DRAM

CIRCUIT DESIGN
Fiindamcntal and High-Speed Topics

Brent Kceth ・ R. Jacob Baker

Brian Johnson ・ Feng Un
DRAM Circuit Design
 IEEE Press
445 Hoes Lane
Piscataway, NJ 08854

IEEE Press Editorial Board

Mohamed E. El-Hawary, Editor in Chief

R. Abari T. G. Croda S. V Kartalopoulos

S. Basu S. Farshchi M. S. Newman
A. Chatteijee B. M. Hammerli
T. Chen R. J. Herrick

Kenneth Moore, Director ofIEEE Book and Information Services (BIS)

Catherine Faduska, Senior Acquisitions Editor
Jeanne Audino, Project Editor

IEEE Solid-State Circuits Society, Sponsor

IEEE SSCS Liaison to the IEEE Press, Stuart K. Tewksbury

Technical Editing and Illustrations

Mary J. Miller, Micron Technology, Inc.

Technical Reviewers
Jim Frenzel, University of Idaho
Zhao Zhang, Iowa State University
DRAM Circuit Design
Fundamental and High-Speed Topics

Brent Keeth
R. Jacob Baker
Brian Johnson
Feng Lin

IEEE Press Series on Microelectronic Systems

Stuart K. Tewksbury and Joe E. Brewer, Series Editors

IEEE Solid-State Circuits Society, Sponsor

令 IEEE
IEEE PRESS

眄 WILEY：
S 2oo7 i

WILEY-INTERSCIENCE
A John Wiley & Sons, lncM Publication
Copyright © 2008 by the Institute of Electrical and Electronic Engineers, Inc. All rights reserved.

Published by John Wiley & Sons, Inc., Hoboken, New Jersey.

Published simultaneously in Canada.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form
or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as
permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior
written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to
the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax
(978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher fbr permission should
be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ
07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permission.

Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in
preparing this book, they make no representations or warranties with respect to the accuracy or
completeness of the contents of this book and specifically disclaim any implied warranties of
merchantability or fitness for a particular purpose. No warranty may be created or extended by sales
representatives or written sales materials. The advice and strategies contained herein may not be
suitable fbr your situation. You should consult with a professional where appropriate. Neither the
publisher nor author shall be liable fbr any loss of profit or any other commercial damages, including
but not limited to special, incidental, consequential, or other damages.

For general information on our other products and services or fbr technical support, please contact our
Customer Care Department within the United States at (800) 762-2974, outside the United States at
(317) 572-3993 or fax (317) 572-4002.

Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may
not be available in electronic format. For information about Wiley products, visit our web site at
www.wiley.com.

Wiley Bicentennial Logo: Richard J. Pacifico

Library of Congress Cataloging-in-Publication Data is available.

ISBN 978-0-470-18475-2

Printed in the United States of America.

10 987654321
To Susi, John, Katie, Gracie, Dory, and Faith (B.K.)
To Julie, Kyrit and Josh (J.B.)
To Cassandra, Parker, Spencer, and Dolpha
To Hui, Luan, and Conrad (EL.)
To Leah and Nicholas (M.M.)
Contents
Preface.......................................................................................... xi

Chapter 1 An Introduction to DRAM

1.1 DRAM Types and Operation...............................................................1
1.1.1 The Ik DRAM (First Generation)........................................... 1
1.1.2 The 4k-64 Meg DRAM (Second Generation)...................... 7
1.1.3 Synchronous DRAM (Third Generation)..............................15
1.2 DRAM Basics....................................................................................22
1.2.1 Access and Sense Operations................................................. 24
1.2.2 Write Operation.......................................................................28
1.2.3 Opening a Row (Summary)............................... 29
1.2.4 Open/Folded DRAM Array Architectures...........................31

Chapter 2 The DRAM Array

2.1 The Mbit Cell................................................................................ 33
2.2 The Sense Amp................................................................................. 44
2.2.1 Equilibration and Bias Circuits.............................................44
2.2.2 Isolation Devices.................................................................... 46
2.2.3 Input/Output Transistors....................................................... 46
2.2.4 Nsense and Psense Amplifiers.............................................. 47
2.2.5 Rate of Activation.................................................................. 49
2.2.6 Configurations......................................................................... 49
2.2.7 Operation.................................................................................. 52
2.3 Row Decoder Elements..................................................................... 54
2.3.1 Bootstrap Wordline Driver........ ............................................. 55
232 NOR Driver............................................................................... 57
2.3.3 CMOS Driver........................................................................... 58

vii
vin Contents

2.3.4 Address Decode Tree............................................................. 58

2.3.5 Static Tree.................................................................................59
2.3.6 P&E Tree................................................................................... 59
23.7 Predecoding.............................................................................. 60
2.3.8 Pass Transistor Tree............................................................... 61
2.4 Discussion........................................................................................... 61

Chapter 3 Array Architectures

3.1 Array Architectures............................................................................ 65
3.1.1 Open Digitline Array Architecture....................................... 65
3.1.2 Folded Array Architecture...................................................... 75
3.2 Design Examples: Advanced Bilevel DRAM Architecture .... 83
3.2.1 Array Architecture Objectives................................................ 84
3.2.2 Bilevel Digitline Construction............................................... 85
3.2.3 Bilevel Digitline Array Architecture..................................... 88
3.2.4 Architectural Comparison......................................................93

Chapter 4 The Peripheral Circuitry

4.1 Column Decoder Elements............................................................... 99
4.2 Column and Row Redundancy.................................................... 102
4.2.1 Row Redundancy............................................................... 104
4.2.2 Column Redundancy.......................................................... 107

Chapter 5 Global Circuitry and Considerations

5.1 Data Path Elements........................................................................ 111
5.1.1 Data Input Buffer................................................................. 111
5.1.2 Data Write Muxes.............................................................. 115
5.1.3 Write Driver Circuit.......................................................... 116
5.1.4 Data Read Path................................................................... 118
5.1.5 DC Sense Amplifier (DCSA)........................................... 119
5.1.6 Helper Flip-Flop (HFF)..................................................... 121
5.1.7 Data Read Muxes.............................................................. 122
5.1.8 Output Buffer Circuit.......................................................... 124
5.1.9 Test Modes............................. 125
5.2 Address Path Elements................................................................. 126
5.2.1 Row Address Path............................................................... 126
5.2.2 Row Address Buffer.......................................................... 127
5.2.3 CBR Counter....................................................................... 127
5.2.4 Predecode Logic................................................................. 128
5.2.5 Refresh Rate.......................................................................... 128
5.2.6 Array Buffers....................................................................... 130
5.2.7 Phase Drivers....................................................................... 131
Contents ix

5.2.8 Column Address Path............................................................ 131

5.2.9 Address Transition Detection............................................... 132
5.3 Synchronization in DRAMs.......................................................... 135
5.3.1 The Phase Detector............................................................... 137
5.3.2 The Basic Delay Element......................................................137
533 Control of the Shift Register................................................ 138
5.3.4 Phase Detector Operation..................................................... 139
5.3.5 Experimental Results.............. ............................................. 140
5.3.6 Discussion.............................................................................. 142

Chapter 6 Voltage Converters

6.1 Internal Voltage Regulators............................................................ 147
6.1.1 Voltage Converters................................................................. 147
6.1.2 Voltage References.............................................................. 148
6.1.3 Bandgap Reference.............................................................. 153
6.1.4 The Power Stage................................................................... 154
6.2 Pumps and Generators..................................................................... 158
6.2.1 Pumps....................................................................................... 158
6.2.2 DVC2 Generator...................................................................165
6.3 Discussion......................................................................................... 165

Chapter 7 An Introduction to High-Speed DRAM

7.1 The Performance Paradigm.............................................................167
7.2 Performance for DRAM Memory Devices.................................. 169
7.3 Underlying Technology Improvements........................................ 171

Chapter 8 High-Speed Die Architectures

8.1 Introduction: Optimizing DRAM Architecture for High Perfor­
mance ........................................................................................ 173
8.2 Architectural Features: Bandwidth, Latency, and Cycle Time . 175
8.2.1 Architectural Limiters: The Array Data Path..................... 176
8.2.2 Architectural Limiters: The Read Data Path.....................183
8.2.3 Architectural Limiters: Latency...........................................186
8.3 Conclusion: Designing fbr High Performance........................... 190

Chapter 9 Input Circuit Paths

9.1 Introduction................................................................................. 193
9.2 Input Receivers................................................................................ 196
9.3 Matched Routing............................................................................. 200
9.4 Capture Latches................................................................................ 203
9.5 Input Timing Adjustments.............................................................. 206
9.6 Current Mode Logic (CML).......................................................... 212
X Contents

Chapter 10 Output Circuit Paths

10.1 Transmission Line, Impedance, and Termination......................220
10.2 Impedance Control....................................................................... 224
10.3 Simultaneous Switching Noise (SSN).......................................230
10.4 Signal Return Path Shift (SRPS)................................................. 238
10.5 Electrostatic Discharge (ESD)....................................................242
10.6 Parallel-to-Serial Conversion...................................................... 244
10.7 Emerging Memory I/O Features........................................ 248

Chapter 11 Timing Circuits

11.1 Introduction..................................................................................... 251
11.2 All-Digital Clock Synchronization Design................................ 254
11.2.1 Timing Analysis................................................................. 255
11.2.2 Digital Delay Line.............................................................. 259
11.2.3 Phase Detector (PD).......................................................... 267
11.2.4 Test and Debug................................................................... 273
11.2.5 Dual-Loop Architecture....................................................274
11.3 Mixed-Mode Clock Synchronization Design............................ 275
11.3.1 Analog Delay Line.............................................................276
11.3.2 Charge-Pump Phase Detector (CPPD)........................... 283
11.3.3 Dual-Loop Analog DLL.................................................... 286
11.3.4 Mixed-Mode DLL and Its Applications......................... 289
11.4 Whafs Next fbr Timing..................................................................... 291

Chapter 12 Control Logic Design

12.1 Introduction..................................................................................... 295
12.2 DRAM Logic Styles..................................................................... 298
12.2.1 Process Limitations............................................................. 298
12.2.2 Array Operation................................................................... 301
12.2.3 Performance Requirements................................................304
12.2.4 Delay-Chain Logic Style................................................... 306
12.2.5 Domino Logic..................................................................... 309
12.2.6 Testability........................................................................... 314
12.3 Command and Address Control................................................. 317
12.3.1 Command Decoder.............................................................318
12.3.2 Read and Write Data Address Registers......................... 324
1^.33 Column Access Control....................................................... 328
12.4 Write Data Latency Timing and Data Demultiplexing.......... 330
12.4.1 Write Latency Timing........................................................ 332
12.4.2 Write Data Demultiplexing ............................................... 338
12.5 Read Data Latency Timing and Data Multiplexing.............. 346
12.5.1 Read Data FIFO................................................................... 350
Contents xi

12.5.2 Read Latency (CL) Tracking............................................ 359

12.6 Comments on Future Direction for DRAM Logic Design ... 367

Chapter 13 Power Delivery

13.1 Power Delivery Network Design................................................. 373
13.2 Device/Package Co-design......................................................... 376
13.3 Full-Chip Simulations.................................................................. 380

Chapter 14 Future Work in High-Performance Memory .... 385

Appendix................................... 391

Glossary................ ................... 407

Index .. ............................................................................... . ... 413

Preface
From the core memory that rocketed into space during the Apollo moon
missions to the solid-state memories used in today9s commonplace com­
puter, memory technology has played an important, albeit quiet, role during
the prior and current centuries. It has been quiet in the sense that memory,
although necessary, is not glamorous and sexy, and is instead being rele­
gated to the role of a commodity. Yet, it is important because memory tech­
nology, specifically, CMOS DRAM technology, has been one of the
greatest driving forces in the advancement of solid-state technology. It
remains a driving force today, despite the segmentation in its market space.
The very nature of the commodity memory market, with high product
volumes and low pricing, is what ultimately drives the technology, lb sur­
vive, let alone remain viable over the long term, memory manufacturers
work aggressively to drive down their manufacturing costs while maintain­
ing, if not increasing, their share of the market. One of the best tools to
achieve this goal remains the ability of manufacturers to shrink their tech­
nology, essentially in getting more memory chips per wafer through process
scaling. Unfortunately, with all memory manufacturers pursuing the same
goals, it is literally a race to see who can get there first. As a result, there is
tremendous pressure to advance the state of the art-more so than in other
related technologies due to the commodity status of memory.
While the memory industry continues to drive forward, most people can
relax and enjoy the benefits-except for those of you who need to join in
the fray. For you, the only way out is straight ahead, and it is for you that we
have written the second edition of this book.
The goal of DRAM Circuit Design: Fundamental and High-Speed Topics
is to bridge the gap between the introduction to memory design available in
most CMOS circuit texts and the advanced articles on DRAM design that

Xlll
XIV Preface

are available in technical journals and symposium digests. The book intro­
duces the reader to DRAM theory, history, and circuits in a systematic, tuto­
rial fashion. The level of detail varies, depending on the topic. In most
cases, however, our aim is merely to introduce the reader to a functional ele­
ment and illustrate it with one or more circuits. After gaining familiarity
with the purpose and basic operation of a given circuit, the reader should be
able to tackle more detailed papers on the subject.
The second half of the book is completely devoted to advanced con­
cepts pertaining to state-of-the-art high-speed and high-performance
DRAM memory. The two halves of the book are worlds apart in content.
This is intentional and serves to rapidly advance the reader from novice to
expert through the turning of a page.
The book begins in Ch叩ter 1 with a brief history of DRAM device evo­
lution from the first 1Kbit device to the 64Mbit synchronous devices. This
chapter introduces the reader to basic DRAM operation in order to lay a
foundation for more detailed discussion later. Chapter 2 investigates the
DRAM memory array in detail, including fundamental array circuits needed
to access the array. The discussion moves into array architecture issues in
Chapter 3, including a design example comparing known architecture types
to a novel, stacked digitline architecture. This design example should prove
useful, fbr it delves into important architectural trade-oflfc and exposes
underlying issues in memory design. Chapter 4 then explores peripheral cir­
cuits that support the memory array, including column decoders and redun­
dancy. The reader should find Chapter 5 very interesting due to the breadth
of circuit types discussed. This includes data path elements, address path
elements, and synchronization circuits. Chapter 6 follows with a discussion
of voltage converters commonly found on DRAM designs. The list of con­
verters includes voltage regulators, voltage references, Fpp/2 generators,
and voltage pumps.
Chapter 7 introduces the concept of high-performance memory and
underlying market forces. Chapter 8 examines high-speed DRAM memory
architectures including a discussion of performance-cost trade-offs. Chapter
9 takes a look at the input circuit path of high-performance DRAM while
Chapter 10 takes a complementary look at the output circuit path. Chapter 11
dives into the complicated world of high-performance timing circuits
including delay-lock-loops and phase-lock-loops. Chapter 12 ties it all
together by tackling the difficult subject of control logic design. This is an
especially important topic in high-performance memory chips. Chapter 13
looks at power delivery and examines methods to improve performance
through careful analysis and design of the power delivery network. Finally,
Chapter 14 discusses future work in high-perfbrmance memory. We wrap
Preface xv

up the book with the Appendix, which directs the reader to a detailed list of
papers from major conferences and journals.

Brent Keeth
R. Jacob Baker
Brian Johnson
Feng Lin
 Chapter
1

An Introduction to DRAM
Dynamic random access memory (DRAM) integrated circuits (ICs) have
existed for more than thirty years. DRAMs evolved from the earliest kilobit
(Kb) generation to the gigabit (Gb) generation through advances in both
semiconductor process and circuit design technology. Tremendous
advances in process technology have dramatically reduced feature size, per­
mitting ever higher levels of integration. These increases in integration have
been accompanied by major improvements in component yield to ensure
that overall process solutions remain cost-effective and competitive. Tech­
nology improvements, however, are not limited to semiconductor process­
ing. Many of the advances in process technology have been accompanied or
enabled by advances in circuit design technology. In most cases, advances
in one have enabled advances in the other. In this chapter, we introduce
some fundamentals of the DRAM IC, assuming that the reader has a basic
background in complementary metal-oxide semiconductor (CMOS) circuit
design, layout, and simulation [1].

1.1 DRAM TYPES AND OPERATION

To gain insight into how modern DRAM chips are designed, it is useful to
look into the evolution of DRAM. In this section, we offer an overview of
DRAM types and modes of operation.

1.1.1 The 1k DRAM (First Generation)

We begin our discussion by looking at the 1,024-bit DRAM (1,024 x
1 bit). Functional diagrams and pin connections appear in Figure 1.1 and
Figure 1.2, respectively. Note that there are 10 address inputs with pin
labels 7?(-7?5 and C]-C5. Each address input is connected to an on-chip
2 Chap. 1 An Introduction to DRAM

address input buffer. The input buffers that drive the row (R) and column
(C) decoders in the block diagram have two purposes: to provide a known
input capacitance (CQ on the address input pins and to detect the input
address signal at a known level so as to reduce timing errors. The level
Vtrip, an idealized trip point around which the input buffers slice the input
signals, is important due to the finite transition times on the chip inputs
(Figure 1.3). Ideally, to avoid distorting the duration of the logic zeros and
ones, VTR1P should be positioned at a known level relative to the maximum
and minimum input signal amplitudes. In other words, the reference level
should change with changes in temperature, process conditions, input maxi­
mum amplitude 化浦，and input minimum amplitude (VJL), Having said
this, we note that the input buffers used in first-generation DRAMs were
simply inverters.
Continuing our discussion of the block diagram shown in Figure LI,
we see that five address inputs are connected through a decoder to the
1,024-bit memory array in both the row and column directions. The total
number of addresses in each direction, resulting from decoding the 5-bit
word, is 32. The single memory array is made up of 1,024 memory ele­
ments laid out in a square of 32 rows and 32 columns. Figure 1.4 illustrates
the conceptual layout of this memory anay. A memory element is located at
the intersection of a row and a column.

% %
? ?

CE* 6

D/N 。 Buffer

R1 -
a
R2 p
o
o Cell matrix
R3 o
p 1,024 bits
-
R4 o:
R5

Figure 1.1 1,024-bit DRAM functional diagram.

Sec. 1.1 DRAM Types and Operation 3

C2 1c ?6 c3

G 2匚 ]15 G

俨 31 Ji4 q

E 4匚 J13 C&

J 5匚 ]l2 D011T

6 6匚 口 11 D,

J 7[ 口10 %

E 8[ 09 %。
Figure 1.2 1,024-bit DRAM pin connections.

Pad Input buffer - Output to

・ decoders

Figure 1.3 Ideal address input buffer.

By applying an address of all zeros to the 10 address input pins, the

memory data located at the intersection of rowO, RA 0, and column 0, CA 0,
is accessed. (It is either written to or read out, depending on the state of the
R/W* input and assuming that the CE* pin is LOW so that the chip is
enabled.)
It is important to realize that a single bit of memory is accessed by using
both a row and a column address. Modem DRAM chips reduce the number
of external pins required for the memory address by using the same pins
fbr both the row and column address inputs (address multiplexing). A clock
signal row address strobe (RAS*) strobes in a row address and then, on the
same set of address pins, a clock signal column address strobe (CAS^)
strobes in a column address at a different time.
4 Chap. 1 An Introduction to DRAM

Column address decoder outputs

号
din 。-ap 。。 p CAO CA1 CA2 CA30 CA31

①
SS8PP8 Moa

Figure 1.4 Layout of a 1,024-bit memory array.

Also note how a first-generation memory array is organized as a logical

square of memory elements. (At this point, we don't know what or how the
memory elements are made. We just know that there is a circuit at the inter­
section of a row and column that stores a single bit of data.) In a modem
DRAM chip, many smaller memory arrays are organized to achieve a larger
memory size. For example, 1,024 smaller memory arrays, each composed
of 256 kbits, may constitute a 256-Meg (256 million bits) DRAM.
LLL1 Reading Data Out of the Ik DRAM. Data can be read out of
the DRAM by first putting the chip in the Read mode by pulling the R/W*
pin HIGH and then placing the chip enable pin C£* in the LOW state. Fig­
ure 1.5 illustrates the timing relationships between changes in the address
inputs and data appearing on the DOUT pin. Important timing specifications
present in this figure are Read cycle time (tRC) and Access time (tAc). The
term tRC specifies how fast the memory can be read. If tRC is 500 ns, then
the DRAM can supply 1-bit words at a rate of 2 MHz. The term tAC spec­
ifies the maximum length of time after the input address is changed before
the output data (DOUT) is valid.
Sec. LI DRAM Types and Operation 5

R/VV* = 1 CE* =10

Figure 1.5 Ik DRAM Read cycle.

1.LL2 Writing to the Ik DRAM. Writing data to the DRAM is

accomplished by bringing the R/W* input LOW with valid data present on
the Din pin. Figure 1.6 shows the timing diagram for a Write cycle. The
term Write cycle time (twc) is related to the maximum frequency at which
we can write data into the DRAM. The term Address to Write delay time
(tAW) specifies the time between the address changing and the R/印* input
going LOW. Finally, Write pulse width (tWp) specifies how long the input
data must be present before the RW* input can go back HIGH in prepara­
tion for another Read or Write to the DRAM. When writing to the DRAM,
we can think of the R/W^ input as a clock signal.
1.LL3 Refreshing the Ik DRAM. The dynamic nature of DRAM
requires that the memory be refreshed periodically so as not to lose the con­
tents of the memory cells. Later we will discuss the mechanisms that lead to
the dynamic operation of the memory cell. At this point, we discuss how
memory Refresh is accomplished fbr the Ik DRAM.
I
q 50 ns

o
A
l

Figure 1.6 Ik DRAM Write cycle.

6 Chap. 1 An Introduction to DRAM

Refreshing a DRAM is accomplished internally: external data to the

DRAM need not be applied. To refresh the DRAM, we periodically access
the memory with every possible row address combination. A timing dia­
gram for a Refresh cycle is shown in Figure 1.7. With the CE* input pulled
HIGH, the address is changed, while the R/W* input is used as a strobe or
clock signal. Internally, the data is read out and then written back into the
same location at full voltage; thus, logic levels are restored (or refreshed).

I - AW _ । ]

耽 / \ !/~
ce* =1 / 1\ /
1 ।

Figure 1.7 Ik DRAM Refresh cycle.

LLL4 A Note on the Power Supplies. The voltage levels used in the
Ik DRAM are unusual by modern-day standards. In reviewing Figure 1.2,
we see that the Ik DRAM chip uses two power supplies: VDD and %s, To
begin,嗅 is a greater voltage than VDD: Vss is nominally 5 Y while Vdd is
-12 V. The value of Vss was set by the need to interface to logic circuits that
were implemented using transistor-transistor logic (TTL) logic. The 17 V
difference between VDD and was necessary to maintain a large signal-
to-noise ratio in the DRAM array. We discuss these topics in greater detail
later in the book. The Vss power supply used in modem DRAM designs, at
the time of this writing, is generally zero; the VDD is in the neighborhood of
L5 V.
LL1.5 The 3-Transistor DRAM Cell. One of the interesting circuits
used in the Ik DRAM (and a few of the 4k and 16k DRAMs) is the 3-tran-
sistor DRAM memory cell shown in Figure 18 The column- and rowlines
shown in the block diagram of Figure 1.1 are split into Write and Read line
pairs. When the Write rowline is HIGH, Ml turns ON. At this point, the
data present on the Write columnline is passed to the gate of M2, and the
information voltage charges or discharges the input capacitance of M2. The
next, and final, step in writing to the mbit cell is to turn OFF the Write row­
line by driving it LOW. At this point, we should be able to see why the
memory is called dynamic. The charge stored on the input capacitance of
M2 will leak off over time.
Sec. 1.1 DRAM Types and Operation 7

Figure 1.8 3-transistor DRAM cell.

If we want to read out the contents of the cell, we begin by first pre­
charging the Read columnline to a known voltage and then driving the Read
rowline HIGH. Driving the Read rowline HIGH turns M3 ON and allows
M2 either to pull the Read columnline LOW or to not change the pre­
charged voltage of the Read columnline. (If M25s gate is a logic LOW, then
M2 will be OFF, having no effect on the state of the Read columnline.) The
main drawback of using the 3-transistor DRAM cell, and the reason it is no
longer used, is that it requires two pairs of column and rowlines and a large
layout area. Modem 1 -transistor, 1 -capacitor DRAM cells use a single row­
line, a single columnline, and considerably less area.

1 .1.2 The 4k-64 Meg DRAM (Second Generation)

We distinguish second-generation DRAMs from first-generation
DRAMs by the introduction of multiplexed address inputs, multiple mem­
ory arrays, and the 1 -transistor/1 -capacitor memory cell. Furthermore, sec­
ond-generation DRAMs offer more modes of operation fbr greater
flexibility or higher speed operation. Examples are page mode, nibble
mode, static column mode, fast page mode (FPM), and extended data out
(EDO). Second-generation DRAMs range in size from 4k (4,096 x 1 bit,
i.e., 4,096 address locations with 1-bit input/output word size) up to 64 Meg
(67,108,864 bits) in memory sizes of 16 Meg x 4 organized as 16,777,216
address locations with 4-bit input/output word size, 8 Meg x 8, or 4 Meg
x 16.
Two other major changes occurred in second-generation DRAMs:
(1) the power supply transitioned to a single 5 V and (2) the technology
advanced from NMOS to CMOS. The change to a single 5 V supply
occurred at the 64kbit density. It simplified system design to a single power
supply fbr the memory, processor, and any TTL (transistor-transistor logic)
used in the system. As a result, rowlines had to be driven to a voltage
8 Chap. 1 An Introduction to DRAM

greater than 5 V to turn the NMOS access devices fully ON (more on this
later), and the substrate held at a potential less than zero. For voltages out­
side the supply range, charge pumps are used (see Chapter 6). The move
from NMOS to CMOS, at the 1Mb density level, occurred because of con­
cerns over speed, power, and layout size. At the cost of process complexity,
complementary devices improved the design.
LI .2.1 Multiplexed Addressing. Figure 1.9 shows a 4k DRAM block
diagram, while Figure 1.10 shows the pin connections for a 4k chip. Note
that compared to the block diagram of the Ik DRAM shown in Figure 1.1,
the number of address input pins has decreased from 10 to 6, even though
the memory size has quadrupled. This is the result of using multiplexed
addressing in which the same address input pins are used for both the row
and column addresses. The row address strobe (RAS*) input clocks the
address present on the DRAM address pins Ao to A5 into the row address
latches on the falling edge. The column address strobe (CAS*) input clocks
the input address into the column address latches on its falling edge.
Figure 1.11 shows the timing relationships between RAS*, CAS*, and
the address inputs. Note that tRC is still (as indicated in the last section) the
random cycle time fbr the DRAM, indicating the maximum rate we can
write to or read from a DRAM. Note too how the row (or column) address
must be present on the address inputs when RAS* (or CAS*) goes LOW.
The parameters and fCAS indicate how long RAS* or CAS* must remain
LOW after clocking in a column or row address. The parameters tASR, tRAH,
Osc, and tCAH indicate the setup and hold times fbr the row and column
addresses, respectively.
2

Column address latches

CAS\

Column decoder

32/
RAS"

Data —a Qx/r
s』

A0— ArrayO (2k)

wq

A1 ― u。一外
-
A2 一
ndf ssappq

A3 —
% 3T
A4
A5 — Arrayl (2k)

Figure 1.9 Block diagram of a 4k DRAM.

Sec. 1.1 DRAM Types and Operation 9

喂1 c •

D/n 21 J15 CAS*

RAV*3Q 1
114 Dour

RAS*a[_ 313 cs*

Ao 5[ J12 A3

Ai 6[ . ]11 A4

A2 71 Jo A5

% 81 J9 %

Figure LIO 4,096-bit DRAM pin connections.

Figure Lil Address timing.

1.1.2.2 Multiple Memory Arrays. As mentioned earlier, second-gen­

eration DRAMs began to use multiple or segmented memory arrays. The
main reason fbr splitting up the memory into more than one array at the cost
of a larger layout area can be understood by considering the parasitics
present in the dynamic memory circuit element. To understand the origins
of these parasitics, consider the modem DRAM memory cell comprising
one MOSFET and one capacitor, as shown in Figure 1.12.
In the next section, we cover the operation of this cell in detail. Here we
introduce the operation of the cell. Data is written to the cell by driving the
rowline (a.k,a., wordline) HIGH, turning ON the MOSFET, and allowing
the columnline (a.k.a., digitline or bitline) to charge or discharge the storage
capacitor. After looking at this circuit for a moment, we can make the fol­
lowing observation.
10 Chap. 1 An Introduction to DRAM

Digitline or
columnline or
bitline

Wordhne .
or rowhne g 1

Storage / JT
capacitor V

Figure 1.12 1 -transistor,

1-capacitor (ITIC) memory cell.

1. The wordline (rowline) may be fabricated using polysilicon (poly).

This allows the MOSFET to be formed by crossing the poly word­
line over an n+ active area.
2. To write a full Vcc logic voltage (where Vcc is the maximum posi­
tive power supply voltage) to the storage capacitor, the rowline must
be driven to a voltage greater than VCc + the n-channel MOSFET
threshold voltage (with body effect). This voltage, > Vcc + VTHj is
often labeled Vcc pumped (VCCP).
3. The bitline (columnline) may be made using metal or polysilicon.
The main concern, as we5ll show in a moment, is to reduce the para­
sitic capacitance associated with the bitline.

Consider the row of N dynamic memory elements shown in

Figure 1.13. Typically, in a modem DRAM, N is 512, which is also the
number of bitlines. When a row address is strobed into the DRAM, via the
address input pins using the falling edge of RAS^, the address is decoded to
drive a wordline (rowline) to Vccp- This turns ON an entire row in a DRAM
memory array. Turning ON an entire row in a DRAM memory array allows
the information stored on the capacitors to be sensed (fbr a Read) via the
bitlines or allows the charging or discharging, via the bitlines, of the storage
capacitors (fbr a Write). Opening a row of data by driving a wordline HIGH
is a very important concept for understanding the modes of DRAM opera­
tion. For Refresh, we only need to supply row addresses during a Refresh
operation. For page Reads-when a row is open-a large amount of data,
which is set by the number of columns in the DRAM array, can be accessed
by simply changing the column address.
We're now in a position to answer the question: "Why are we limited to
increasing the number of columnlines (or bitlines) used in a memory
array?" or "Why do we need to break up the memory into smaller memory
arrays?^^ The answer to these questions comes from the realization that the
Sec. 1.1 DRAM Types and Operation 11

more bitlines we use in an array, the longer the delay through the wordline
(Figure 1.13).

BitlineO Bitlinel BitlineN-2 BitlineN-1

From row
Wordline ■o
address
driver

£
decoder
Wordtine |q
E

v V V V

Figure L13 Row ofN dynamic memory elements.

If we drive the wordline on the left side of Figure 1.13 HIGH, the signal
will take a finite time to reach the end of the wordline (the wordline on the
right side of Figure 1.13). This is due to the distributed resistance/capaci-
tance structure formed by the resistance of the polysilicon wordline and the
capacitance of the MOSFET gates. The delay limits the speed of DRAM
operation. To be precise, it limits how quickly a row can be opened and
closed. To reduce this RC time, a polycide wordline is formed by adding a
silicide, for example, a mixture of a refractory metal such as tungsten
with polysilicon, on top of poly silicon. Using a polycide wordline will have
the effect of reducing the wordline resistance. Also, additional drivers can
be placed at different locations along the wordline, or the wordline can be
stitched at various locations with metal.
The limitations on the additional number of wordlines can be under­
stood by realizing that by adding more wordlines to the array, more para­
sitic capacitance is added to the bitlines. This parasitic capacitance becomes
important when sensing the value of data charge stored in the memory ele­
ment. We'll discuss this in more detail in the next section.
LL2.3 Memory Array Size. A comment is in order about memory
array size and how addressing can be used fbr setting word and page size.
(We'll explain what this means in a moment.) If we review the block dia­
gram of the 4k DRAM shown in Figure 1.9, we see that two 2k-DRAM
memory arrays are used. Each 2k memory is composed of 64 wordlines and
32 bitlines fbr 2,048 memory elements/address locations per array. In the
block diagram, notice that a single bit, coming from the column decoder,
can be used to select data, via the bitlines, from ArrayO or Array 1.
From our discussion earlier, we can open a row in ArrayO while at the
same time opening a row in Array 1 by simply applying a row address to the
input address pins and driving RAS* LOW. Once the rows are open, it is a
simple matter of changing the column address to select different data asso­
12 Chap. 1 An Introduction to DRAM

ciated with the same open row from either array. If our word size is 1 bit, we
could define a page as being 64 bits in length (32 bits from each array). We
could also define our page size as 32 bits with a 2-bit word for input/output.
We would then say that the DRAM is a 4k DRAM organized as 2k x 2. Of
course, in the 4k DRAM, in which the number of bits is small, the concepts
of page reads or size aren't too useful. We present them here simply to illus­
trate the concepts. Lefs consider a more practical and modem configura­
tion.
Suppose we have a 64-Meg DRAM organized as 16 Meg x 4 (4 bits
input/output) using 4k row address locations and 4k column address loca­
tions (12 bits or pins are needed fbr each 4k of addressing). If our (sub)
memory array size is 256kbits, then we have a total of 256 memory arrays
on our DRAM chip. We'll assume that there are 512 wordlines and 512 bit­
lines (digitlines), so that the memory array is logically square. (However,
physically, as we shall see, the array is not square.) Internal to the chip, in
the address decoders, we can divide the row and column addresses into two
parts: the lower 9 bits fbr addressing the wordlines/bitlines in a 256k mem­
ory array and the upper 3 bits for addressing one of the 64 "group・o&fdur”
memory arrays (6 bits total coming from the upper 3 bits of the row and col­
umn addresses).
Our 4-bit word comes from the group-of-fbur memory arrays (one bit
from each memory array). We can define a page of data in the DRAM by
realizing that when we open a row in each of the four memory arrays, we
are accessing 2k of data (512 bits/array x 4 arrays). By simply changing the
column address without changing the row address and thus opening another
group-of-fbur wordlines, we can access the 2k “page" of data. With a little
imagination, we can see different possibilities fbr the addressing. For exam­
ple, we could open 8 group-of-fbur memory arrays with a row address and
thus increase the page size to 16k, or we could use more than one bit at a
time from an array to increase word size.
LL2.4 Refreshing the DRAM. Refreshing the DRAM is accom­
plished by sequentially opening each row in the DRAM. (We'll discuss how
the DRAM cell is refreshed in greater detail later in the book.) If we use the
64-Meg example in the last section, we need to supply 4k row addresses to
the DRAM by changing the external address inputs from 000000000000 to
111111111111 while clocking the addresses into the DRAM using the fall­
ing edge of RAS*. In some DRAMs, an internal row address counter is
present to make the DRAM easier to refresh. The general specification fbr
64-Meg DRAM Refresh is that all rows must be refreshed at least every
64 ms, which is an average of 15.7 jxs per row. This means, that if the Read
cycle time tRC is 100 ns (see Figure 1.11), it will take 4,096 • 100 ns or
Sec. 1.1 DRAM Types and Operation 13

410)is to refresh a DRAM with 4k of row addresses. The percentage of

time the DRAM is unavailable due to Refresh can be calculated as
410 ps/64 ms or 0.64% of the time. Note that the Refresh can be a burst,
taking 410 jis as just described, or distributed, where a row is refreshed
every 15.7 jis.
1.L2.5 Modes of Operation. From the last section, we know that we
can open a row in one or more DRAM arrays concurrently, allowing a page
of data to be written to or read from the DRAM. In this section, we look at
the different modes of operation possible for accessing this data via the col­
umn address decoder. Our goal in this section is not to present all possible
modes of DRAM operation but rather to discuss the modes that have been
used in second-generation DRAMs. These modes are page mode, nibble
mode, static column mode, fest page mode, and extended data out.
Figure 1.14 shows the timing diagram for a age mode Read, Write, and
Read-Modify-Write. We can understand this timing diagram by first notic­
ing that when RAS* goes LOW, we clock in a row address, decode the row
address, and then drive a wordline in one or more memory arrays to VCcp-
The result is an open row(s) of data sitting on the digitlines columnlines).
Only one row can be opened in any single array at a time. Prior to opening
a row, the bitlines are precharged to a known voltage. (Precharging to VCCH
is typically performed using internal circuitry.) Also notice at this time that
data out, DOut, is in a Hi-Z state; that is, the DRAM is not driving the bus
line connected to the DOUT pin.
The next significant timing event occurs when CAS* goes LOW and the
column address is clocked into the DRAM (Figure 1.14). At this time, the
column address is decoded, and, assuming that the data from the open row
is sitting on the digitlines, it is steered using the column address decoder to
Dout- We may have an open row of 512 bits, but we are steering only one
bit to Dout. Notice that when G4S* goes HIGH, DOUT goes back to the Hi-
Z state.
By strobing in another column address with the same open row, we can
select another bit of data (again via the column address decoder) to steer to
the Dout pin. In this case, however, we have changed the DRAM to the
Write mode (Figure 1.14). This allows us to write, with the same row open
via the DIN pin in Figure 1.10, to any column address on the open row.
Later, second-generation DRAMs used the same pins for both data input
and output to reduce pin count. These bidirectional pins are labeled DQ.
The final set of timing signals in Figure 1.14 (the right side) read data
out of the DRAM with R/印* HIGH, change R/W^ to a LOW, and then write
to the same location. Again, when CAS* goes HIGH, Dour goes back to the
Hi-Z state.
14 Chap. 1 An Introduction to DRAM

The remaining modes of operation are simple modifications of page

mode. As seen in Figure 1J5, FPM allows the column address to propagate
into the column circuits while CAS* is HIGH. The speed of the DRAM thus
improves by reducing the delay between CAS* going LOW and valid data
present, or accessed, on DOUT (tCAC)- EDO is simply an FPM DRAM that
doesn't force DOUT to a Hi-Z state immediately when CAS* goes HIGH.
The data out of the DRAM is thus available for a longer period of time,
allowing fbr faster system operation. In general, opening the row is the
operation that takes the longest amount of time. Once a row is open, the
data sitting on the columnlines can be steered to DOUT at a fast rate. Interest­
ingly, using column access modes has been the primary method of boosting
DRAM performance over the years, especially in double-data rate (DDR)
DRAMs (see Chapter 8).
The other popular modes of operation in second-generation DRAMs
were the static column and nibble modes. Static column mode DRAMs
used flow-through latches in the column address path. When a column
address was changed externally, with CAS* LOW, the column address fed
directly to the column address decoder. (The address wasn't clocked on the
falling edge of G45* ) This increased the speed of the DRAM by prevent­
ing the outputs from going into the Hi-Z state with changes in the column
address.

Figure L14 Page mode.

Nibble mode DRAMs used an internal presettable address counter so

that by strobing CAS*, the column address would change internally. Figure
1.16 illustrates the timing operation fbr a nibble mode DRAM. The first
time CAS* transitions LOW (first being defined as the first transition after
RAS* goes LOW), the column address is loaded into the counter. If E4S* is
Sec. 1.1 DRAM Types and Operation 15

held LOW and CAS* is toggled, the internal address counter is incremented,
and the sequential data appears on the output of the DRAM. The term nib­
ble mode comes from limiting the number of CAS* cycles to four (a nibble).
g
SSRPP4

s*
UT

////////

Figure 1.15 Fast page mode.

lRAS

RAS* -1
i NCP ■

CAS* f-Y /~t

’A SC

I 二^ • 一： L 'CAH
cah : : :

«
i। :，*:,
: t
lRAC
Dqut

^rite //〃///[ \7777.

Figure 1.16 Nibble mode.

1.1.3 Synchronous DRAM (Third Generation)

Synchronous DRAMs (SDRAMs) are made by adding a synchronous
interface between the basic core DRAM operation/circuitry of second-gen­
eration DRAMs and the control coming from oflf-chip to make the DRAM
operation fester. All commands and operations to and from the DRAM are
executed on the rising edge of a master or command clock signal that is
common to all SDRAMs and labeled CLK. See Figure 1.17 for the pin con­
nections of a 64Mb SDRAM with 16-bit input/output (I/O).
16 Chap. 1 An Introduction to DRAM

SDRAMs operate with a maximum CLK frequency in the range of

100-167 MHz, clocking data on one edge of the clock. This means that if
a 64Mb SDRAM is organized as a xl6 part (that is, the input/output word
size is 16 bits), the maximum rate at which the words can be written to the
part is 200-334 MB/s.
7
2 Vss
o
3 53 DQ15

RARRRRRRARRFIFIRmRRRRRRmmFIFIQ
V
0 4 52 /sSQ
D DDQ7 51
D 0DQ2 5 50 DQ14
6 49 DQ13
V Q 48
Q %D0
D 5sQ3 7 47
D Q4 8 46 DQ12
45 DQ11
V 9 44
Q “SSQ
DQ 5 43
D 6 10 42 DQ10
11 41 DQ9
VQ 12 40
07M 39 %Q
D - 13
Q 14 38 DQ8
V
15 37
D 16 36 Vss
Q
W s 17 35 NC
C4 18 34 DQMH
19 33 CLK
R4 20 32
Cs s 21 31 CKE
BA ★ 22 30 NC
23 29
B4 o 24 28 A11
A 7
名

25 A9
O 26
A A8
A1 A7
A2 A6
A A5
V3 A4
%
7

Figure 1.17 Pin connections of a 64Mb SDRAM with 16-bit I/O.

A variation of the SDRAM is the double-data-rate SDRAM (DDR

SDRAM, or simply DDR DRAM), The DDR parts register commands and
operations on the rising edge of the clock signal while allowing data to be
transferred on both the rising and falling edges. A differential input clock
signal is used in the DDR DRAM with the labeling of, not surprisingly,
CLK and CLK*. In addition, the DDR DRAM provides an output data
strobe, labeled DQS, synchronized with the output data and the input CLK.
DQS is used at the controller to strobe in data from a DRAM. The big bene­
fit of using a DDR part is that the data transfer rate can be twice the clock
frequency because data can be transferred on both the rising and falling
edges of CLK. This means that when using a 133 MHz clock, the data writ­
ten to and read from the DRAM can be transferred at 266M words/s. Using
Sec. 1.1 DRAM Types and Operation 17

the numbers from the previous paragraph, this means that a 64Mb DDR
SDRAM with an input/output word size of 16 bits will transfer data to and
from the memory controller at 400-572 MB/s.
Figure 1.18 shows the block diagram of a 64Mb SDRAM with 16-bit
I/O. Note that although CLK is now used for transferring data, we still have
the second-generation control signals CS*, WE*, CAS*, and, RAS* present
on the part. (CKE is a clock enable signal which, unless otherwise indi­
cated, is assumed HIGH.) Lefs discuss how these control signals are used
in an SDRAM by recalling that in a second-generation DRAM, a Write was
executed by first driving WE* and CS* LOW. Next a row was opened by
applying a row address to the part and then driving RAS* LOW. (The row
address is latched on the falling edge of RAS^.) Finally, a column address
was ipplied and latched on the falling edge of CAS*. A short time later, the
data applied to the part would be written to the accessed memory location.
For the SDRAM Write, we change the syntax of the descriptions of
whafs happening in the part. However, the fundamental operation of the
DRAM circuitry is the same as that of the second-generation DRAMs. We
can list these syntax changes as follows:

1. The memory is segmented into banks. For the 64Mb memory of Fig­
ure 1.17 and Figure 1.18, each bank has a size of 16Mbs (organized
as 4,096 row addresses [12 bits] x 256 column addresses [8 bits] x 16
bits [16 Z)。I/O pins]), As discussed earlier, this is nothing more than
a simple logic design of the address decoder (the banks can be laid
out so that they are physically in the same area). The bank selected is
determined by the addresses BAO and BA 1.
2. In second-generation DRAMs, we said, “We open a row,“ as dis­
cussed earlier. In SDRAM, we now say, “We activate a row in a
bank." We do this by issuing an active command to the part. Issuing
an active command is accomplished on the rising edge of CLK with a
row/bank address applied to the part with CS* and RAS* LOW, while
CAS^ and WE* are held HIGH.
3. In second-generation DRAMs, we said, “We write to a location given
by a column address,“ by driving CAS* LOW with the column
address applied to the part and then applying data to the part. In an
SDRAM, we write to the part by issuing the Write command to the
part. Issuing a Write command is accomplished on the rising edge of
CLK with a column/bank address applied to the part: CS*, CAS*, and
WE* are held LOW, and RAS* is held HIGH.
18 Chap. 1 An Introduction to DRAM

P
C
0
E
spews

E Cwlrol
C4S*K O logic
D

Mode regisler Row* BankO Banko

DOML,
row- memory
address OQMH
ddress array
latch & (4,096 x
iecodm 256 x16)
Sense amps 16 J Data
output DQO-
register DO15
/Q gating DQM
mask logic
Read data latch
contror Write drivers

266 x
ColurnTT
address
counter/
latch

Figure 1.18 Block diagram of a 64Mb SDRAM with 16-bit I/O.

Table 1.1 shows the commands used in an SDRAM. In addition, this

table shows how inputs/outputs (DQs) can be masked using the DQ mask
(DQM) inputs. This feature is useful when the DRAM is used in graphics
applications.
SDRAMs often employ pipelining in the address and data paths to
increase operating speed. Pipelining is an effective tool in SDRAM design
because it helps disconnect operating frequency and access latency. Without
pipelining, a DRAM can only process one access instruction at a time.
Essentially, the address is held valid internally until data is fetched from the
array and presented to the output buffers. This single instruction mode of
operation ties operating frequency and access time (or latency) together.
However, with pipelining, additional access instructions can be fed into the
SDRAM before prior access instructions have completed, which permits
access instructions to be entered at a higher rate than would otherwise be
allowed. Hence, pipelining increases operating speed.
Sec. 1.1 DRAM Types and Operation 19

Table 1.1 SDRAM commands. (Notes: 1)

Name C5* A45* CAS* 腔* DQM ADDR DQs Notes

Command H X X X X X X —
inhibit (NOP)

No operation L H H H X X X —
(NOP)

Active (select L L H H X Bank/ X 3

bank and row
activate row)

Read (select L H L H L/H8 Bank/ X 4

bank and col­ col
umn, and start
Read burst)

Write (select L H L L L/H8 Bank/ 必lid 4

bank and col­ col
umn, and start
Write burst)

Burst terminate L H H L X X Activ —

e

PRECHARGE L L H L X Code X 5
(deactive row in
bank or banks)

Auto-Refresh L L L H X X X 6,7
ors self-refresh
(enter self­
refresh mode)

Load mode reg­ L L L L X Op­ X 2

ister code

Write ENABLE/ — 一 — — L — Activ 8

output ENABLE e

Write inhibit/ 一 — — —
— H — Hi-Z 8
output Hi-Z

Notes
1. CKE is HIGH fbr all commands shown except for self-refresh.
2. A0-A 11 define the op-code written to the mode register.
3. AO-A11 provide row address, and BAO, BAX determine which bank is made active.
4. /0-49 (x4),力(K48 (x8), or 40-47 (x16) provide column address; J10 HIGH
enables the auto PRECHARGE feature (nonpersistent), while 410 LOW disables the
auto PRECHARGE feature; 840, BA\ determine which bank is being read from or
written to.
5. N10 LOW: BAO, BAX determine the bank being precharged. >410 HIGH: all banks
precharged and BAO, BA 1 are “don't care.”
6. This command is Auto-Refresh if CKE is HIGH and Self-Refresh if CKE is LOW.
20 Chap. 1 An Introduction to DRAM

7. Internal Refresh counter controls row addressing; all inputs and I/Os are "don't
care” except fbr CKE.
8. Activates or deactivates the DQs during Writes (zero-clock delay) and Reads (two-
clock delay).

Pipeline stages in the data path can also be helpful when synchronizing
output data to the system clock. CAS latency refers to a parameter used by
the SDRAM to synchronize the output data from a Read request with a par­
ticular edge of the system clock. A typical Read fbr an SDRAM with CAS
latency set to three is shown in Figure 1.19. SDRAMs must be capable of
reliably functioning over a range of operating frequencies while maintain­
ing a specified CAS latency. This is often accomplished by configuring the
pipeline stage to register the output data to a specific clock edge, as deter­
mined by the CAS latency parameter.

CLK _

Addresses A ； 乂 X X
Output, Q

Figure 1.19 SDRAM with a latency of three.

At this point, we should understand the basics of SDRAM operation,

but we may be asking, 46Why are SDRAMs potentially faster than second-
generation DRAMs such as EDO or FPM?" The answer to this question
comes from the realization that it's possible to activate a row in one bank
and then, while the row is opening, perform an operation in some other
bank (such as reading or writing). In addition, one of the banks can be in a
PRECHARGE mode (the bitlines are driven to Kcc/2) while accessing one
of the other banks and, thus, in effect hiding PRECHARGE and allowing
data to be continuously written to or read from the SDRAM. (Of course,
this depends on which application and memory address locations are used.)
We use a mode register, as shown in Figure 1.20, to put the SDRAM into
specific modes of operation fbr programmable operation, including pipelin­
ing and burst Reads/Writes of data [2].
Sec. 1.1 DRAM Types and Operation 21

An Aw A9 A8 A7 A6 A5 A4 A3 A2 Ao

111111111111Addressbus
/ 11/10/9/ 8/7/6/5/4/3/2/1/0/I mhHa m3
I Reserved* WB )p mode CAS* latency BT Burst length
J
1
Burst length

M2 M1 MO M3 = 0 M3 = 1
0 0 0 1 1
0 0 1 2 2
0 1 0 4 4
0 1 1 8 8
1 0 0 Reserved Reserved
1 0 1 Reserved Reserved
1 1 0 Reserved Reserved
1 1 1 Full page Reserved

M3 Burst type
0 Sequential
1 Interleaved

M6 M5 M4 CAS* latency
0 0 0 Reserved
0 0 1 Reserved
0 1 0 2
0 1 1 3
1 0 0 Reserved
1 0 1 Reserved
1 1 0 Reserved
1 1 1 Reserved

M3 M7 M6-M0 Operating mode

0 0 Defined Standard operation
— — — All other states reserved

M9 Write burst mode •Should program

M11,M10 = 0,0to
0 Programmed burst length
ensure compatibility
1 Single location access with future devices.

Figure 1.20 Mode register

22 Chap. 1 An Introduction to DRAM

1.2 DRAM BASICS

A modem DRAM memory cell or memory bit (mbit), as shown in
Figure 1.21, is formed with one transistor and one capacitor, accordingly
referred to as a 1T1C cell. The mbit is capable of holding binary informa­
tion in the form of stored charge on the capacitor. The mbit transistor oper­
ates as a switch interposed between the mbit capacitor and the digitline.
Assume that the capacitor's common node is biased at which we will
later show as a reasonable assumption. Storing a logic one in the cell
requires a capacitor with a voltage of +匕割/2 across it. Therefore, the
charge stored in the mbit capacitor is

“cc
Q =长C

(ID
where C is the capacitance value in farads. Conversely, storing a logic zero
in the cell requires a capacitor with a voltage of --VCCI2 across it. Note that
the stored charge on the mbit capacitor for a logic zero is

―"cc
Q =受C

(1.2)
The charge is negative with respect to the VCCI2 common node voltage
in this state. Various leakage paths cause the stored capacitor charge to
slowly deplete. To return the stored charge and thereby maintain the stored
data state, the cell must be refreshed. The required refreshing operation is
what makes DRAM memory dynamic rather than static.

Wordline

T or rowline

M1

Potential = Vcc for

logic one and ground Digitline or
for logic zero. columnline

Figure L21 1T1C DRAM memory cell.

(Note the rotation of the rowline and columnline.)

The digitline referred to earlier consists of a conductive line connected

to a multitude of mbit transistors. The conductive line is generally con­
Sec. L2 DRAM Basics 23

structed from either metal or silicide/polycide polysilicon. Because of the

quantity of mbits connected to the digitline and its physical length and
proximity to other features, the digitline is highly capacitive. For instance, a
typical value for digitline capacitance on a 50 nm process might be 120 fF.
Digitline capacitance is an important parameter because it dictates many
other aspects of the design. We discuss this further in Section 2.L For now,
we continue describing basic DRAM operation.
The mbit transistor gate terminal is connected to a wordline rowline).
The wordline, which is connected to a multitude of mbits, is actually
formed of the same polysilicon as that of the transistor gate. The wordline is
physically orthogonal to the digitline. A memory array is formed by tiling a
selected quantity of mbits together such that mbits along a given digitline
do not share a common wordline and mbits along a common wordline do
not share a common digitline. Examples of this are shown in Figures 1.22
and 1.23, In these layouts, mbits are paired to share a common contact to
the digitline, which reduces the array size by eliminating duplication.

D3 D2 Do

Figure L22 Open digitline memory array schematic.

24 Chap. 1 An Introduction to DRAM

Figure 1.23 Open digitline memory array layout.

1.2.1 Access and Sense Operations

Next, we examine the access and sense operations. We begin by assum­
ing that the cells connected to DI, in Figure 1.24, have logic one levels
(+%c/2) stored on them and that the cells connected to DO have logic zero
levels (――2) stored on them. Next, we form a digitline pair by consider­
ing two digitlines from adjacent arrays. The digitline pairs, labeled DO/DO*
and Dl/Dl *, are initially equilibrated to Kcc/2 V. All wordlines are initially
at 0 V, ensuring that the mbit transistors are OFF. Prior to a wordline firing,
the digitlines are electrically disconnected from the VCCH bias voltage and
allowed to float. They remain at the Vcc!2 PRECHARGE voltage due to
their capacitance*
To read mbitl, wordline WLO changes to a voltage that is at least one
transistor VTM above Vcc. This voltage level is referred to as VCcp or Vpp.
To ensure that a full logic one value can be written back into the mbit capac­
itor, VCcp must remain greater than one VTH above The mbit capacitor
begins to discharge onto the digitline at two different voltage levels depend­
ing on the logic level stored in the cell. For a logic one, the capacitor begins
to discharge when the wordline voltage exceeds the digitline PRECHARGE
voltage by For a logic zero, the capacitor begins to discharge when the
wordline voltage exceeds VTH. Because of the finite rise time of the word­
Sec. L2 DRAM Basics 25

line voltage, this difference in turn-on voltage translates into a significant

delay when reading ones, as seen in Figure 1.25.

<
<

VV<
<

Arra

Figure 1.24 Simple array schematic (an open DRAM array).

Accessing a DRAM cell results in charge sharing between the mbit

capacitor and the digitline capacitance. This charge sharing causes the digit­
line voltage either to increase for a stored logic one or to decrease for a
stored logic zero. Ideally, only the digitline connected to the accessed mbit
will change. In reality, the other digitline voltage also changes slightly, due
to parasitic coupling between digitlines and between the firing wordline and
the other digitline. (This is especially true for the folded bitline architecture
discussed later.) Nevertheless, a differential voltage develops between the
two digitlines. The magnitude of this voltage difference, or signal, is a func­
tion of the mbit capacitance (以尻),digitline capacitance 仁厕),and volt­
age stored on the cell prior to access (Vceu). See Figure 1.26. Accordingly,

y = y 'mbit
signal Cell'cdigit^cmbit

(13)
A VSignai of200 mV is yielded from a design in which VceU = 1.00,。或〃 =20
fF, and 孰城=200 fF.
26 Chap. 1 An Introduction to DRAM

Figure 1.25 Cell access waveforms.

After the cell has been accessed, sensing occurs. Sensing is essentially
the amplification of the digitline signal or the differential voltage between
the digitlines. Sensing is necessary to properly read the cell data and refresh
the mbit cells. (The reason fbr forming a digitline pair now becomes appar­
ent) presents a schematic diagram fbr a simplified sense amplifier circuit:
a cross-coupled NMOS pair and a cross-coupled PMOS pair. The sense
amplifiers also appear like a pair of cross-coupled inverters in which ACT
and NLAT* provide power and ground. The NMOS pair or Nsense-amp has
a common node labeled NLAT* (for Nsense-amp latch).
Similarly, the Psense-amp has a common node labeled ACT (fbr Active
pull-up). Initially, NLAT* is biased to VCCH. and ACT is biased to Vss. or
signal ground. Because the digitline pair DI and DI* are both initially at
Vee。, the Nsense-amp transistors are both OFF. Similarly, both Psense-
amp transistors are OFF. Again, when the mbit is accessed, a signal devel­
ops across the digitline pair. While one digitline contains charge from the
cell access, the other digitline does not but serves as a reference for the
Sensing operation. The sense amplifiers are generally fired sequentially: the
Nsense-amp first, then the Psense-amp. Although designs vary at this point,
Sec. L2 DRAM Basics 27

the higher drive of NMOS transistors and better VTH matching offer better
sensing characteristics by Nsense-amps and thus lower error probability
compared to Psense-amps.
Waveforms fbr the Sensing operation are shown in Figure 1.28. The
Nsense-amp is fired by bringing NLAT* (Nsense-amp latch) toward ground.
As the voltage difference between NLAT* and the digitlines (DI and DI* in
)approaches the NMOS transistor, whose gate is connected to the
higher voltage digitline, begins to conduct. This conduction occurs first in
the subthreshold and then in the saturation region as the gate-to-source volt­
age exceeds VTH and causes the low-voltage digitline to discharge toward
the NLAT* voltage. Ultimately, NLAT* will reach ground and the digitline
will be brought to ground potential. Note that the other NMOS transistor
will not conduct: its gate voltage is derived from the low-voltage digitline,
which is being discharged toward ground. In reality, parasitic coupling
between digitlines and limited subthreshold conduction by the second tran­
sistor result in a temporary voltage drop on the high digitline, as seen in
Figure 1.28.

Figure 1.26 DRAM charge sharing..

Psense-amp Nsense-amp

Figure L27 Sense amplifier schematic.

Sometime after the Nsense-amp fires, ACT will be brought toward Vcc
to activate the Psense-amp, which operates in a complementary fashion to
the Nsense-amp. With the low-voltage digitline approaching ground, there
is a strong signal to drive the appropriate PMOS transistor into conduction.
This conduction, again moving from subthreshold to saturation, charges the
28 Chap. 1 An Introduction to DRAM

high-voltage digitline toward ACT, ultimately reaching Vcc. Because the

mbit transistor remains ON, the mbit capacitor is refreshed during the Sens­
ing operation. The voltage, and hence charge, which the mbit capacitor held
prior to accessing, is restored to a full level: Vcc for a logic one and ground
for a logic zero. It should be apparent now why the minimum wordline volt­
age is a Vth above Vcc, If VCCP were anything less, a full Vcc level could
not be written back into the mbit capacitor The mbit transistor source volt­
age Vsource cannot be greater than Vgate- Vth because this would turn OFF
the transistor.

Figure 1.28 Sensing operation waveforms.

1.2.2 Write Operation

A Write operation is similar to a Sensing and Restore operation except
that a separate Write driver circuit determines the data that is placed into the
celL The Write driver circuit is generally a tristate inverter connected to the
digitlines through a second pair of pass transistors, as shown in Figure 1.29.
These pass transistors are referred to as I/O transistors. The gate terminals
of the I/O transistors are connected to a common column select (CSEL) sig­
nal. The CSEL signal is decoded from the column address to select which
pair (or multiple pairs) of digitlines is routed to the output pad or, in this
case, the Write driver.
In most current DRAM designs, the Write driver simply overdrives the
sense amplifiers, which remain ON during the Write operation. After the
new data is written into the sense amplifiers, the amplifiers finish the Write
cycle by restoring the digitlines to full rail-to-rail voltages. An example is
shown in Figure 1.30 in which DI is initially HIGH after the Sensing oper­
ation and LOW after the writing operation. A Write operation usually
involves only 2-4 mbits within an array of mbits because a single CSEL
line is generally connected to only four pairs of I/O transistors. The remain-
Sec. L2 DRAM Basics 29

ing digitlines are accessed through additional CSEL lines that correspond to
different column address locations.

Figure 1.29 Sense amplifier schematic with I/O devices.

Write Restore

Figure 1.30 Write operation waveforms.

1.2.3 Opening a Row (Summary)

Opening a row of mbits in a DRAM array is a fundamental operation
for both reading and writing to the DRAM array. Sometimes the chain of
events from a circuit designer's point of view, which lead to an open row, is
called the RAS* timing chain. We summarize the RAS* timing chain of
events below, assuming that for a second-generation DRAM both RAS* and
CAS* are HIGH. (Ifs easy to extend our discussion to third-generation
DRAMs where RAS* and CAS* are effectively generated from the control
logic.)
30 Chap. 1 An Introduction to DRAM

1. Initially, both RAS* and CAS* are HIGH. All bitlines in the DRAM
are driven to 展/2, while ail wordlines are at 0 V This ensures that
all of the mbit's access transistors in the DRAM are OFF.
2. A valid row address is applied to the DRAM and RAS* goes LOW.
While the row address is being latched, on the falling edge of RAS*,
and decoded, the bitlines are disconnected from the Vcc/2 bias and
allowed to float. The bitlines at this point are charged to VCCH^ and
they can be thought of as capacitors.
3. The row address is decoded and applied to the wordline drivers. This
forces only one rowline in at least one memory array to Vccp- Driv­
ing the wordline to VCCP turns ON the mbits attached to this rowline
and causes charge sharing between the mbit capacitance and the
capacitance of the corresponding bitline. The result is a small pertur­
bation (upwards for a logic one and downwards for a logic zero) in
the bitline voltages.
4. The next operation is Sensing, which has two purposes: a) to deter­
mine if a logic one or zero was written to the cell and b) to refresh
the contents of the cell by restoring a full logic zero (0 V) or one
(Vcc) to the capacitor. Following the wordlines going HIGH, the
Nsense-amp is fired by driving, via an n-channel MOSFET, NLAT*
to ground. The inputs to the sense amplifier are two bitlines: the bit­
line we are sensing and the bitline that is not active (a bitline that is
still charged to Vcc2—an inactive bitline). Pulling NLAT* to ground
results in one of the bitlines going to ground. Next, the ACT signal is
pulled up to Vcc, driving the other bitline to Vcc. Some important
notes:

(a) It doesn't matter if a logic one or logic zero was sensed because
the inactive and active bitlines are pulled in opposite directions.
(b) The contents of the active cell, after opening a row, are restored
to full voltage levels (either 0 V or VCc)- The entire DRAM can
be refreshed by opening each row.

Now that the row is open, we can write to or read from the DRAM. In
either case, it is a simple matter of steering data to or from the active
array(s) using the column decoder. When writing to the array, buffers set the
new logic voltage levels on the bitlines. The row is still open because the
wordline remains HIGH. (The row stays open as long as RAS^ is LOW.)
When reading data out of the DRAM, the values sitting on the bitlines
are transmitted to the output buffers via the I/O MOSFETs. To increase the
speed of the reading operation, this data, in most situations, is transmitted to
Sec. 1.2 DRAM Basics 31

the output buffer (sometimes called a DQ buffer) either through a helper

flip-flop or another sense amplifier.
A note is in order here regarding the word size stored in or read out of
the memory array. We may have 512 active bitlines when a single rowline
in an array goes HIGH (keeping in mind once again that only one wordline
in an array can go HIGH at any given time), This literally means that we
could have a word size of 512 bits from the active anay. By adjusting the
word size, or the number of these 512 bits we send to the chip's output, we
can change the speed on power performance of the chips. For example, if
our chip's I/O is x4 (by 4, that is, 4-bit input/output), then sending eight bits
at a time (out of these 512) provides data fbr two clock edges. Sending 32
bits provides data for eight clock edges. The cost of transmitting larger
word sizes is larger bus widths (more layout area). However, the benefit is
that more time is available to send the data once it is in the pipeline. The
output word size is the basic difference between the DDR, DDR2, and
DDR3 DRAM topologies.

1.2.4 Open/Folded DRAM Array Architectures

Throughout the book, we make a distinction between the open array
architecture as shown in Figures 1.22 and 1.24 and the folded DRAM array
used in many modem DRAMs and seen in Figure 1.31. At the cost of
increased layout area, folded arrays increase noise immunity by moving
sense amp inputs next to each other. These sense amp inputs come directly
from the DRAM array. The term folded comes from taking the DRAM
arrays seen in Figure 1.24 and folding them together to form the topology
seen in Figure 1.31.

Figure 1.31 A folded DRAM array.

32 Chap. 1 An Introduction to DRAM

REFERENCES

[1] R. J. Baker, CMOS: Circuit Design, Layout, and Simulation, 2nd ed.
Hoboken, NJ: John Wiley & Sons, Inc., and IEEE Press, 2005.
[2] Micron Technology, Inc., Synchronous DRAM datasheet, 1999.
 Chapter
2

The DRAM Array

This chapter begins a more detailed examination of standard DRAM array
elements. This examination is necessary fbr a clear understanding of funda­
mental DRAM elements and how they are used in memory block construc­
tion. A common point of reference is required before considering the
analysis of competing array architectures. Included in this chapter is a
detailed discussion of mbits, array configurations, sense amplifier elements,
and row decoder elements.

2.1 THE MBIT CELL

The primary advantage of DRAM over other types of memory technology
is low cost. This advantage arises from the simplicity and scaling character­
istics of its 1T1C memory cell [1]. Although the DRAM mbit is simple con­
ceptually, its actual design and implementation are highly complex.
Therefore, successful, cost-effective DRAM designs require a tremendous
amount of process technology.
Figure 2.1 presents the layout of a modem buried capacitor DRAM
mbit pair. (Buried means that the capacitor is below the digitline.) This type
of mbit is also referred to as a bitline over capacitor (BOC) cell. Because
sharing a contact significantly reduces overall cell size, DRAM mbits are
constructed in pairs. In this way, a digitline contact can be shared. The mbits
comprise an active area rectangle (in this case, an n+ active area), a pair of
polysilicon wordlines, a single digitline contact, a metal or polysilicon dig­
itline, and a pair of cell capacitors formed with an oxide-nitride-oxide
dielectric between two layers of polysilicon. For most processes, the word­
line polysilicon is silicided to reduce sheet resistance, permitting longer
wordline segments without reducing speed. The mbit layout, as shown in

33
34 Chap. 2 The DRAM Array

Figure 2.1, is essentially under the control of process engineers, for every
aspect of the mbit must meet stringent performance and yield criteria.

storage node ^tline contact

Metal 1
digitline

Poly3
cellplate
N+ active area Wordlines
(polyl)

Figure 2.1 Mbit pair layout.

A small array of mbits appears in Figure 2.2. This figure is useful to

illustrate several features of the mbit. First, note that the digitline pitch
(width plus space) dictates the active area pitch and the capacitor pitch. Pro­
cess engineers adjust the active area width and the field oxide width to max­
imize transistor drive and minimize transistor-to-transistor leakage. Field
oxide technology greatly impacts this balance. A thicker field oxide or a
shallower junction depth affords a wider transistor active area. Second, the
wordline pitch (width plus space) dictates the space available for the digit­
line contact, transistor length, active area, field poly width, and capacitor
length. Optimization of each of these features by process engineers is neces­
sary to maximize capacitance, minimize leakage, and maximize yield. Con­
tact technology, subthreshold transistor characteristics, photolithography,
and etch and film technology dictate the overall design.

Figure 2.2 Layout to show anay pitch.

At this point in the discussion, it is appropriate to introduce the concept

of feature size and how it relates to cell size. The mbit shown in Figures 2.1
Sec. 2.1 The Mbit Cell 35

and 2.2 is by definition an eight-square feature cell (8F2) [2][3]. The

intended definition of feature (F) in this case is minimum realizable process
dimension but in fact equates to a dimension that is one-half the wordline
(row) or digitline (column) pitch. A 45 nm process having wordline and
digitline pitches of 100 nm yields an mbit size that is
7
(8) • SQnm = 0.02|im

(2・D
It is easier to explain the 8F2 designation with the aid of Figure 2.3. An
imaginary box drawn around the mbit defines the cell's outer boundary.
Along the x-axis, this box includes one-half digitline contact feature, one
wordline feature, one capacitor feature, one field poly feature, and one-half
poly space feature fbr a total of four features. Along the y-axis, this box
contains two one-half field oxide features and one active area feature fbr a
total of two features. The area of the mbit is therefore

4F-2F = 8F2

(2.2)

Figure 2.3 Layout to show 8F2 derivation.

36 Chap. 2 The DRAM Array

The folded array architecture, as shown in Figure 2.2, always produces

an 8F2 mbit. This results from the fact that each wordline connects or forms
a crosspoint with an mbit transistor on every other digitline and must pass
around mbit transistors as field poly on the remaining digitlines. The field
poly in each mbit cell adds two square features to what would have been a
6F2 cell. Although the folded array yields a cell that is 25% larger than other
array architectures, it also produces superior signal-to-noise performance,
especially when combined with some form of digitline twisting [4], Supe­
rior low-noise performance made folded array architecture the architecture
of choice since the 64kbit generation [5].
A folded array is schematically depicted in Figure 24 Sense amplifier
circuits placed at the edge of each array connect to both true and comple­
ment digitlines (D and D*) coming from a single anay. Optional digitline
pair twisting in one or more positions reduces and balances the coupling to
adjacent digitline pairs and improves overall signal-to-noise characteristics
[4]. Figure 2.5 shows the variety of twisting schemes used throughout the
DRAM industry [6].

Figure 2.4 Folded digitline array schematic.

Sec. 2.1 The Mbit Cell 37

Ideally, a twisting scheme equalizes the coupling terms from each digit­
line to all other digitlines, both true and complement. If implemented prop­
erly, the noise terms cancel or at least produce only common-mode noise to
which the differential sense amplifier is more immune.
Each digitline twist region consumes valuable silicon area. Thus, design
engineers resort to the simplest and most efficient twisting scheme to get the
job done. Because the coupling between adjacent metal lines is inversely
proportional to the line spacing, the signal-to-noise problem gets increas­
ingly worse as DRAMs scale to smaller and smaller dimensions. Hence, the
industry trend is toward use of more complex twisting schemes on succeed­
ing generations [6][7].

Figure 2.5 Digitline twisting scheme.

38 Chap. 2 The DRAM Array

An alternative to the folded array architecture, popular prior to the

64kbit generation [1], was the open digitline architecture. Seen schemati­
cally in Figure 2.6, this architecture also features the sense amplifier circuits
between two sets of arrays [8]. Unlike the folded array, however, true and
complement digitlines (D and D*) connected to each sense amplifier pair
come from separate arrays [9]. This arrangement precludes using digitline
twisting to improve signal-to-noise performance, which is the prevalent rea­
son why the industry switched to folded arrays. Note that unlike the folded
array architecture, each wordline in an open digitline architecture connects
to mbit transistors on every digitline, creating crosspoint-style arrays.
This feature permits a 25% reduction in mbit size to only 6F2 because
the wordlines do not have to pass alternate mbits as field poly. The layout
for an anay of standard 6F2 mbit pairs is shown in Figure 2.7 [2]. A box is
drawn around one of the mbits to show the 6F2 cell boundary. Again, two
mbits share a common digitline contact to improve layout efficiency. While
folded array architectures were used extensively in second-generation
DRAMs [8] and the open array architectures, as just mentioned were used
extensively in first-generation DRAMs, the third-generation designs are
seeing a return to the open architecture. As the device/array size continue to
shrink with scaling, the effects of coupled noise on the arrays is reduced,
which makes the open array architecture both viable and preferable (due to
the reduced size).

Figure 2.6 Open digitline array schematic.

Sec. 2.1 The Mbit Cell 39

Figure 2.7 Open digitline array layout.

(Feature size, F, is equal to one-half digitline pitch.)

Digitline capacitive components, contributed by each mbit, include

junction capacitance, digitline-to-cellplate (poly3), digitline-to-wordline,
digitline-to-digitline, digitline-to-substrate, and, in some cases, digitline-to-
storage cell (poly2) capacitance. Therefore, each mbit connected to the dig­
itline adds a specific amount of capacitance to the digitline. Most modem
DRAM designs have no more than 512 mbits connected to a digitline seg­
ment.
Two factors dictate this quantity. First, for a given cell size, as deter­
mined by row and column pitches, a maximum storage capacitance can be
achieved without resorting to exotic processes or excessive cell height. For
processes in which the digitline is above the storage capacitor (buried
capacitor), contact technology determines the maximum allowable cell
height. This fixes the volume available (cell area multiplied by cell height)
in which to build the storage capacitor. Second, as the digitline capacitance
increases, the power associated with charging and discharging this capaci­
tance during Read and Write operations increases. Any given wordline
essentially accesses (crosses) all of the columns within a DRAM. For a 256-
Meg DRAM, each wordline crosses 16,384 columns. With a multiplier such
as that, it is easy to appreciate why limits to digitline capacitance are neces­
sary to keep power dissipation low.
40 Chap. 2 The DRAM Array

Figure 2.8 presents a process cross section fbr the buried capacitor mbit
depicted in Figure 2.2, and Figure 2.9 shows a SEM image of the buried
capacitor mbit. This type of mbit, employing a buried capacitor structure,
places the digitline physically above the storage capacitor [10]. The digit­
line is constructed from either metal or polycide, while the digitline contact
is formed using a metal or polysilicon plug technology. The mbit capacitor
is formed with polysilicon (poly2) as the bottom plate, an oxide-nitride-
oxide (ONO) dielectric, and a sheet of polysilicon (poly3). This top sheet of
polysilicon becomes a common node shared by all mbit capacitors. The
capacitor shape can be simple, such as a rectangle, or complex, such as con­
centric cylinders or stacked discs. The most complex capacitor structures
are the topic of many DRAM process papers [11][12][13].

Digitline

ONO dielectric
tse Poly3
c ■\\ cellplate
o
o
①
U
S6e o s

Z = Field poly
A = Wordline
o 6
」

d Q ri~~h
Field oxide
VN+ acHve area)

P-substrate

Figure 2.8 Buried capacitor cell process cross section.

Figure 2.9 Buried capacitor cell process SEM image.

The ONO dielectric undergoes optimization to achieve maximum

capacitance with minimum leakage. It must also tolerate the maximum
DRAM operating voltage without breaking down. For this reason, the cell­
plate (poly3) is normally biased at + Kcc/2 V, ensuring that the dielectric
Sec 2.1 The Mbit Cell 41

has no more than VCCH V across it for either stored logic state, a logic one
at + VCCH V ora logic zero at -2/2 V.
Two other basic mbit configurations are used in the DRAM industry.
The first, shown in Figures 2.10, 2.11, and 2.12, is referred to as a buried
digitline or capacitor over bitline (COB) cell [14][15]. The digitline in this
cell is almost always formed of polysilicon rather than of metal.

Figure 240 Buried digitline Mbit cell layout.

P-substrate Field poly N+ active area

Figure 2.11 Buried digitline mbit process cross section.

42 Chap. 2 The DRAM Array

Figure 2.12 Buried digitline Mbit process SEM image.

As viewed from the top, the active area is normally bent or angled to
accommodate the storage capacitor contact that must drop between digit­
lines. An advantage of the buried digitline cell over the buried capacitor cell
of Figure 2.8 is that its digitline is physically very close to the silicon sur­
face, making digitline contacts much easier to produce. The angled active
area, however, reduces the effective active area pitch, constraining the isola­
tion process even further In buried digitline cells, it is also very difficult to
form the capacitor contact. Because the digitline is at or near minimum
pitch fbr the process, insertion of a contact between digitiines can be diffi­
cult.
Figures 2.13 and 2.14 present a process cross section of the third type of
mbit used in the construction of DRAMs. Using trench storage capacitors,
this cell is accordingly called a trench cell [12][13]. Trench capacitors are
formed in the silicon substrate, rather than above the substrate, after etching
deep holes into the wafer The storage node is a doped polysilicon plug,
which is deposited in the hole following growth or deposition of the capaci­
Sec. 2.1 The Mbit Cell 43

tor dielectric. Contact between the storage node plug and the transistor drain
is usually made through a poly strap.

Figure 2.14 Trench capacitor Mbit process SEM image.

With most trench capacitor designs, the substrate serves as the com­
mon-node connection to the capacitors, preventing the use of +%。/2 bias
and thinner dielectrics. The substrate is heavily doped around the capacitor
to reduce resistance and improve the capacitor's CV characteristics. A real
44 Chap. 2 The DRAM Array

advantage of the trench cell is that the capacitance can be increased by

merely etching a deeper hole into the substrate [16].
Furthermore, the capacitor does not add stack height to the design,
greatly simplifying contact technology. The disadvantage of trench capaci­
tor technology is the difficulty associated with reliably building capacitors
in deep silicon holes and connecting the trench capacitor to the transistor
drain terminal.

2.2 THE SENSE AMP

The term sense amplifier actually refers to a collection of circuit elements
that pitch up to the digitlines of a DRAM array. This collection most gener­
ally includes isolation transistors, devices for digitline equilibration and
bias, one or more Nsense amplifiers, one or more Psense amplifiers, and
devices connecting selected digitlines to I/O signal lines. All of the circuits
along with the wordline driver circuits, as discussed in Section 2.3, are
called pitch cells. This designation comes from the requirement that the
physical layout fbr these circuits is constrained by the digitline and wordline
pitches of an array of mbits. For example, the sense amplifiers for a specific
digitline pair (column) are generally laid out within the space of four digit­
lines. With one sense amplifier fbr every four digitlines, this is commonly
referred to as quarter pitch or four pitch.

2.2.1 Equilibration and Bias Circuits

The first elements analyzed are the equilibration and bias circuits. As
we can recall from the discussions of DRAM operation in Section 1.1, the
digitlines start at 匕0/2 V prior to cell Access and Sensing [17]. It is vitally
important to the Sensing operation that both digitlines, which form a col­
umn pair, are at the same voltage before the wordline is driven HIGH. Any
offset voltage appearing between the pair directly reduces the effective sig­
nal produced during the Access operation [5], Equilibration of the digitlines
is accomplished with one or more NMOS transistors connected between the
digitline conductors. NMOS are used because of their higher drive capabil­
ity and the resulting faster equilibration.
An equilibration transistor, together with bias transistors, is shown in
Figure 2.15. The gate terminal is connected to a signal called equilibrate
(EQ). EQ is held to Vcc whenever the external row address strobe signal
RAS^ is HIGH. This indicates an inactive or PRECHARGE state for the
DRAM. After RAS* has fallen, EQ transitions LOW, turning the equilibra­
tion transistor OFF just prior to a wordline going HIGH. EQ will again tran­
sition HIGH at the end of a RAS* cycle to force equilibration of the
Sec. 2.2 The Sense Amp 45

digitlines. The equilibration transistor is sized large enough to ensure rapid

equilibration of the digitlines to prepare the part fbr a subsequent access.
As shown in Figure 2.15, two more NMOS transistors accompany the
EQ transistor to provide a bias level of VCCI2 V. These devices operate in
conjunction with equilibration to ensure that the digitline pair remains at the
prescribed voltage fbr Sensing. Normally, digitlines that are at Vcc and
ground equilibrate to Vcc。V [5]. The bias devices ensure that this occurs
and also that the digitlines remain at VCCI2. despite leakage paths that
would otherwise discharge them. Again, fbr the same reasons as for the
equilibration transistor, NMOS transistors are used. Most often, the bias and
equilibration transistors are integrated to reduce their overall size. VCCH V
PRECHARGE is used on most modem DRAMs because it reduces power
consumption and Read-Write times and improves Sensing operations.
Power consumption is reduced because a VCCI2 PRECHARGE voltage can
be obtained by equilibrating the digitlines (which are at Vcc and ground,
respectively) at the end of each cycle.

Figure 2.15 Equilibration schematic.

The charge-sharing between the digitlines produces VCCH without

additional ICc current. Some designs, however, equilibrate the digitlines to
Vdd [18]. Because the wordlines and digitlines are both at Vcc when the
part is inactive, row-to-column shorts that exist do not contribute to
increased standby current. Vcc。PRECHARGE DRAMs, on the other hand,
suffer higher standby current with row-to-column shorts because the word­
lines and digitlines are at different potentials when the part is inactive. A
layout for the equilibration and bias circuits is shown in Figure 2.16.
46 Chap. 2 The DRAM Array

Figure 2.16 Equilibration and bias circuit layout.

2.2.2 Isolation Devices

Isolation devices are important to the sense amplifier circuits. These
devices are NMOS transistors placed between the array digitlines and the
sense amplifiers (see Figure 2.19). Isolation transistors are physically
located on both ends of the sense amplifier layout. In quarter-pitch sense
amplifier designs, there is one isolation transistor for every two digitlines.
Although this is twice the active area width and space of an array, it never­
theless sets the limit fbr isolation processing in the pitch cells.
The isolation devices provide two functions. First, if the sense amps are
positioned between and connected to two arrays, they electrically isolate
one of the two arrays. This is necessary whenever a wordline fires in one
array because isolation of the second array reduces the digitline capacitance
driven by the sense amplifiers, thus speeding Read-Write times, reducing
power consumption, and extending Refresh for the isolated array. Second,
the isolation devices provide resistance between the sense amplifier and the
digitlines. This resistance stabilizes the sense amplifiers and speeds up the
Sensing operation by somewhat isolating the highly capacitive digitlines
from the low-capacitance sense nodes [19]. Capacitance of the sense nodes
between isolation transistors is generally less than 15fF, permitting the
sense amplifier to latch much faster than if it were solidly connected to the
digitlines. The isolation transistors slow Write-Back to the mbits, but this is
far less of a problem than initial Sensing.

2.2.3 Input/Output Transistors

The input/output (I/O) transistors allow data to be read from and written
to specific digitline pairs. A single I/O transistor is connected to each sense
node as shown in Figure 2.17. The outputs of each I/O transistor are con­
nected to I/O signal pairs. Commonly, there are two pairs of I/O signal lines,
Sec. 2.2 The Sense Amp 47

which permit four I/O transistors to share a single column select (CSEL)
control signal. DRAM designs employing two or more metal layers run the
column select lines across the arrays in either Metal2 or Metal3. Each col­
umn select can activate four I/O transistors on each side of an array to con­
nect four digitline pairs (columns) to peripheral data path circuits. The I/O
transistors must be sized carefully to ensure that instability is not introduced
into the sense amplifiers by the I/O bias voltage or remnant voltages on the
I/O lines.
Although designs vary significantly as to the numerical ratio, I/O tran­
sistors are designed to be two to eight times smaller than the Nsense ampli­
fier transistors. This is sometimes referred to as beta ratio. A beta ratio
between five and eight is considered standard; however, it can only be veri­
fied with silicon. Simulations may foil to adequately predict sense amplifier
instability, although theory would predict better stability with higher beta
ratio and better Write times with lower beta ratio. During Write, the sense
amplifier remains ON and must be overdriven by the Write driver (see
Section 1.2.2).
I/O b* I/O b I/O a* I/O a

Figure 2.17 I/O transistors.

2.2.4 Nsense and Psense Amplifiers

The fundamental elements of any sense amplifier block are the Nsense
amplifier and the Psense amplifier. These amplifiers, as previously dis­
cussed, work together to detect the access signal voltage and drive the digit­
lines, accordingly to Vcc and ground. The Nsense amplifier depicted in
Figure 2.18 consists of cross-coupled NMOS transistors. The Nsense ampli­
48 Chap. 2 The DRAM Array

fier drives the low-potential digitline to ground. Similarly, the Psense ampli­
fier consists of cross-coupled PMOS transistors and drives the HIGH-
potential digitline to KCc-

Figure 2.18 Basic sense amplifier block.

The sense amplifiers are carefully designed to guarantee correct detec­

tion and amplification of the small signal voltage produced during cell
access (less than 200mV) [5]. Matching of transistor VTH^ transconductance,
and junction capacitance within close tolerances helps ensure reliable sense
amplifier operation. Ultimately, the layout dictates the overall balance and
performance of the sense amplifier block. As a result, a tremendous amount
of time is spent ensuring that the sense amplifier layout is optimum. Sym­
metry and exact duplication of elements are critical to a successful design.
This includes balanced coupling to all sources of noise, such as I/O lines
and latch signals (NLAT* and ACT). Balance is especially critical for layout
residing inside the isolation transistors. And because the sense node capaci­
tance is very low, it is more sensitive to noise and circuit imbalances.
While the majority of DRAM designs latch the digitlines to Vcc and
ground, a certain number of designs choose to reduce these levels. Various
technical papers report improved Refresh times and lower power dissipation
through reductions in latch voltages [20][21]. At first, this appears contra­
dictory: writing a smaller charge into the memory cell would require
refreshing the cell more often. However, by maintaining lower drain-to-
source voltages (Vds) and negative gate-to-source voltages (VGS) across
non«accessed mbit transistors, substantially lower subthreshold leakage and
longer Refresh times can be realized, despite the smaller stored charge. An
important concern that we are not mentioning here is the leakage resulting
from defects in the silicon crystal structure. These defects can result in an
increase in the drain/substrate diode saturation current and can practically
limit the Refresh times.
Most designs that implement reduced latch voltages generally raise the
Nsense amplifier latch voltage without lowering the Psense amplifier latch
voltage. Designated as boosted sense ground designs, they write data into
each mbit using full Vcc for a logic one and boosted ground for a logic zero.
The sense ground level is generally a few hundred millivolts above true
ground. In standard DRAMs, which drive digitlines fully to ground, the VGS
Sec. 2.2 The Sense Amp 49

of nonaccessed mbits becomes zero when the digitlines are latched. This
results in high subthreshold leakage for a stored one level because full VCc
exists across the mbit transistor while the VGS is held to zero. Stored zero
levels do not suffer from prolonged subthreshold leakage: any amount of
cell leakage produces a negative VGS for the transistor. The net effect is that
a stored one level leaks away much faster than a stored zero level. One's
level retention, therefore, establishes the maximum Refresh period for most
DRAM designs. Boosted sense ground extends Refresh by reducing sub­
threshold leakage for stored ones. This is accomplished by guaranteeing
negative gate-to-source bias on nonaccessed mbit transistors. The benefit of
extended Refresh from these designs is somewhat diminished, though, by
the added complexity of generating boosted ground levels and the problem
of digitlines that no longer equilibrate at KCc/2 V.

2.2.5 Rate of Activation

The rate at which the sense amplifiers are activated has been the subject
of some debate. A variety of designs use multistage circuits to control the
rate at which NLAT* fires. Especially prevalent with boosted sense ground
designs are two-stage circuits that initially drive NLAT* quickly toward true
ground to speed sensing and then bring NLAT* to the boosted ground level
to reduce cell leakage. An alternative to this approach, again using two-
stage drivers, drives NLAT* to ground, slowly at first to limit current and
digitline disturbances. This is followed by a second phase in which NLAT*
is driven more strongly toward ground to complete the Sensing operation.
This phase usually occurs in conjunction with ACT activation. Although
these two designs have contrary operation, each meets specific performance
objectives: trading ofFnoise and speed.

2.2.6 Configurations
Figure 2.19 shows a sense amplifier block commonly used in double- or
triple-metal designs. It features two Psense amplifiers outside the isolation
transistors, a pair of £g/bias (EQb) devices, a single Nsense amplifier, and
a single I/O transistor for each digitline. Because only half of the sense
amplifiers fbr each array are on one side, this design is quarter pitch, as are
the designs in Figures 2.20 and 2.21. Placement of the Psense amplifiers
outside the isolation devices is necessary because a full one level (Vcc) can­
not pass through unless the gate terminal of the ISO transistors is driven
above Vcc. £0/bias transistors are placed outside of the ISO devices to per­
mit continued equilibration of digitlines in arrays that are isolated. The I/O
transistor gate terminals are connected to a common CSEL signal fbr four
adjacent digitlines. Each of the four I/O transistors is tied to a separate I/O
50 Chap. 2 The DRAM Array

bus. This sense amplifier, though simple to implement, is somewhat larger

than other designs due to the presence of two Psense amplifiers.

Figure 2.20 Complex sense amplifier block.

Figure 2.21 Reduced sense amplifier block.

Figure 2.20 shows a second, more complicated style of sense amplifier

block. This design employs a single Psense amplifier and three sets of
Nsense amplifiers. Because the Psense amplifier is within the isolation tran­
sistors, the isolation devices must be the NMOS depletion type, the PMOS
enhancement type, or the NMOS enhancement type with boosted gate drive
Sec. 2.2 The Sense Amp 51

to permit writing a full logic one into the array mbits. The triple Nsense
amplifier is suggestive of PMOS isolation transistors; it prevents fiill zero
levels to be written unless the Nsense amplifiers are placed adjacent to the
arrays. In this more complicated style of sense amplifier block, using three
Nsense amplifiers guarantees faster sensing and higher stability than a simi­
lar design using only two Nsense amplifiers. The inside Nsense amplifier is
fired before the outside Nsense amplifiers. However, this design will not
yield a minimum layout, an objective that must be traded off against perfor­
mance needs.
The sense amplifier block of Figure 2.21 can be considered a reduced
configuration. This design has only one Nsense-amp and one Psense-amp,
both of which are placed within the isolation transistors. To write full logic
levels, either the isolation transistors must be depletion mode devices or the
gate voltage must be boosted above VCc by at least one VTH. This design
still uses a pair of EQ/bias circuits to maintain equilibration on isolated
arrays.
Only a handful of designs operates with a single Eg/bias circuit inside
the isolation devices, as shown in Figure 2.22. Historically, DRAM engi­
neers tended to shy away from designs that permitted digitlines to float for
extended periods of time. However, as of this writing, at least three manu­
facturers in volume production have designs using this scheme.
A sense amplifier design for single-metal DRAMs is shown in
Figure 2.23. Prevalent on 1-Meg and 4-Meg designs, single-metal processes
conceded to multi-metal processes at the 16-Meg generation. Unlike the
sense amplifiers shown in Figures 2.19, 2.20, 2.21, and 2.22, single-metal
sense amps are laid out at half pitch: one amplifier for every two array digit­
lines. This type of layout is extremely difficult and places tight constraints
on process design margins. With the loss of Metal2, the column select sig­
nals are not brought across the memory arrays. Generating column select
signals locally for each set of I/O transistors requires a full column decode
block.

Figure 2.22 Minimum sense amplifier block.

52 Chap. 2 The DRAM Array

As shown in Figure 2.23, the Nsense-amp and Psense-amps are placed

on separate ends of the arrays. Sharing sense amps between arrays is espe­
cially beneficial for single-metal designs. As illustrated in Figure 2.23, two
Psense-amps share a single Nsense-amp. In this case, with the I/O devices
on only one end, the right Psense-amp is activated only when the right array
is accessed. Conversely, the left Psense-amp is always activated, regardless
of the accessed array, because all Read and Write data must pass through the
left Psense-amp to get to the I/O devices.

Figure 2.23 Single-metal sense amplifier block.

2.2.7 Operation
A set of signal waveforms is illustrated in Figure 2.24 for the sense
amplifier of Figure 2.19. These waveforms depict a Read-Modify-Write
cycle (Late Write) in which the cell data is first read out and then new data
is written back. In this example, a one level is read out of the cell, as indi­
cated by DO rising above DO during cell access. A one level is always
+Vcc/2 in the mbit cell, regardless of whether it is connected to a true or
complement digitline. The correlation between mbit cell data and the data
appearing at the DRAM's data terminal (DQ) is a function of the data topol­
ogy and the presence of data scrambling. Data or topo scrambling is imple­
mented at the circuit level: it ensures that the mbit data state and the DQ
logic level are in agreement. An mbit one level (+@72) corresponds to a
logic one at the DQ, and an mbit zero level (一喂⑵ conesponds to a logic
zero at the DQ terminal.
Writing specific data patterns into the memory arrays is important to
DRAM testing. Each type of data pattern identifies the weaknesses or sensi­
tivities of each cell to the data in surrounding cells. These patterns include
solids, row stripes, column stripes, diagonals, checkerboards, and a variety
of moving patterns. Test equipment must be programmed with the data
topology of each type of DRAM to correctly write each pattern. Often the
tester itself guarantees that the pattern is correctly written into the arrays,
unscrambling the complicated data and address topology as necessary to
write a specific pattern. On some newer DRAM designs, part of this task is
Sec. 2.2 The Sense Amp 53

implemented on the DRAM itself, in the form of a topo scrambler, such that
the mbit data state matches the DQ logic level. This implementation some­
what simplifies tester programming.

vCCP

%/2

Access Sense Write Restore

Figure 2.24 Waveforms for the Read-Modify-Write cycle.

Returning to Figure 2.24, we see that a wordline is fired in Array 1. Prior

to this, ISOa* will go LOW to isolate ArrayO, and EQb will go LOW to dis­
able the Eg/bias transistors connected to Array 1. The wordline then fires
HIGH, accessing an mbit, which dumps charge onto DO*. NLAT*, which is
initially at VCCI2. drives LOW to begin the Sensing operation and pulls £>0
toward ground. Then, ACT fires, moving from ground to VCCi activating the
Psense amplifier and driving DO* toward VCc- After separation has com­
menced, CSELO rises to Vcc. turning ON the I/O transistors so that the cell
data can be read by peripheral circuits. The I/O lines are biased at a voltage
close to Vcc, which causes DI to rise while the column is active. After the
Read is complete, Write drivers in the periphery turn ON and drive the I/O
lines to opposite data states (in our example).
This new data propagates through the I/O devices and writes over the
existing data stored in the sense amplifiers. Once the sense amplifiers latch
the new data, the Write drivers and the I/O devices can be shut down, allow­
ing the sense amplifiers to finish restoring the digitlines to full levels. Fol­
lowing this restoration, the wordline transitions LOW to shut OFF the mbit
transistor Finally, EQb and ISOa* fire HIGH to equilibrate the digitlines
back to Fcc/2 and reconnect AnayO to the sense amplifiers. The timing for
each event of Figure 2.24 depends on circuit design, transistor sizes, layout,
device performance, parasitics, and temperature. While timing for each
event must be minimized to achieve optimum DRAM performance, it can­
not be pushed so far as to eliminate all timing margins. Margins are neces­
sary to ensure proper device operation over the expected range of process
variations and the wide range of operating conditions.
54 Chap. 2 The DRAM Array

Again, there is not one set of timing waveforms that covers all design
options. The sense amps of Figures 2.19=2.23 all require slightly different
signals and timings. Various designs actually fire the Psense amplifier prior
to or coincident with the Nsense amplifier. This obviously places greater
constraints on the Psense amplifier design and layout, but these constraints
are balanced by potential performance benefits. Similarly, the sequence of
events as well as the voltages for each signal can vary. There are almost as
many designs for sense amplifier blocks as there are DRAM design engi­
neers. Each design reflects various influences, preconceptions, technolo­
gies, and levels of understanding. The bottom line is to maximize yield and
performance and minimize everything else.

2.3 ROW DECODER ELEMENTS

Row decode circuits are similar to sense amplifier circuits in that they pitch
up to mbit arrays and have a variety of implementations. A row decode
block is comprised of two basic elements: a wordline driver and an address
decoder tree. There are three basic configurations for wordline driver cir­
cuits: the NOR driver, the inverter (CMOS) driver, and the bootstrap driver.
In addition, the drivers and associated decode trees can be configured either
as local row decodes for each array section or as global row decodes that
drive a multitude of array sections.
Global row decodes connect to multiple arrays through metal wordline
straps. The straps are stitched to the polysilicon wordlines at specific inter­
vals dictated by the polysilicon resistance and the desired RC wordline time
constant. Some processes that strap wordlines with metal do not silicide the
polysilicon, although doing so would reduce the number of stitch regions
required. Strapping wordlines and using global row decoders obviously
reduces die size [22], very dramatically in some cases. The disadvantage of
strapping is that it requires an additional metal layer at minimum array
pitch. This puts a tremendous burden on process technologists in which
three conductors are at minimum pitch: wordlines, digitlines, and wordline
straps.
Local row decoders, on the other hand, require additional die size rather
than metal straps. It is highly advantageous to reduce the polysilicon resis­
tance in order to stretch the wordline length and lower the number of row
decodes needed. This is commonly achieved with silicided polysilicon pro­
cesses. On large DRAMs, such as the 1Gb, the area penalty can be prohibi­
tive, making low-resistance wordlines all the more necessary.
Sec. 2.3 Row Decoder Elements 55

2.3.1 Bootstrap Wordline Driver

The bootstrap wordline driver shown in Figure 2.25 is built exclusively
from NMOS transistors [22], resulting in the smallest layout of the three
types of driver circuits. The absence of PMOS transistors eliminates large
Nwell regions from the layout. As the name denotes, this driver relies on
bootstrapping principles to bias the output transistor's gate terminal. This
bias voltage must be high enough to allow the NMOS transistor to drive the
wordline to the boosted wordline voltage VCcp-

Figure 2.25 Bootstrap wordline driver.

Operation of the bootstrap driver is described with help from Figure

2.26. Initially, the driver is OFF with the wordline and PHASE terminals at
ground. The wordline is held at ground by transistor M2 because the
decoder output signal DEC* is at Vcc, The gate of M3, the pass transistor, is
fixed at Kcc. The signals DEC and DEC* are fed from a decode circuit that
will be discussed later. As a complement pair, DEC and DEC* represent the
first of two terms necessary to decode the correct wordline. PHASE0, which
is also fed from a decode circuit, represents the second term. If DEC rises to
Vcc and DEC* drops to ground, as determined by the decoder, the boot
node labeled 51 will rise to Vcc~ VTN V, and M2 will turn OFF. The word­
line continues to be held to ground by Ml because PHASE0 is still
grounded. After Bl rises to Vcc- VTN. the PHASE。signal fires to the
boosted wordline voltage VCCP. Because of the gate-to-drain and gate-to-
source capacitance of Ml, the gate of M1 boots to an elevated voltage,
Vboot. This voltage is determined by the parasitic capacitance of node Bl,
Cgsi, CGD], Vccp，and the initial voltage at Bl,喂——. Accordingly,
56 Chap. 2 The DRAM Array

一忌喏%—
(2.3)
In conjunction with the wordline voltage rising from ground to VCCP, the
gate-to-source capacitance of Ml provides a secondary boost to the boot
node. The secondary boost helps to ensure that the boot voltage is adequate
to drive the wordline to a full Vccp level.

Figure 2.26 Bootstrap operation waveforms.

The layout of the boot node is very important to the bootstrap wordline
driver. First, the parasitic capacitance of node which includes routing,
junction, and overlap components, must be minimized to achieve maximum
boot efficiency. Second, charge leakage from the boot node must be mini­
mized to ensure adequate VGS fbr transistor Ml such that the wordline
remains at VCcp for the maximum RAS* LOW period. Low leakage is often
achieved by minimizing the source area for M3 or by using donut gate
structures that surround the source area, as illustrated in Figure 2.27.

Figure 2.27 Donut gate structure layout.

Sec. 23 Row Decoder Elements 57

The bootstrap driver is turned OFF by first driving the PHASE0 signal
to ground. Ml remains ON because node Bl cannot drop below VCc- %h；
Ml substantially discharges the wordline toward ground. This is followed
by the address decoder turning OFF, bringing DEC to ground and DEC* to
Vcc. With DEC* at Vcc. transistor M2 turns ON and fully clamps the word­
line to ground. A voltage level translator is required for the PHASE signal
because it operates between ground and the boosted voltage VCCP. For a glo­
bal row decode configuration, this requirement is not much of a burden. For
a local row decode configuration, however, the requirement for level trans­
lators can be very troublesome. Generally, these translators are placed either
in the array gap cells at the intersection of the sense amplifier blocks and
row decode blocks or distributed throughout the row decode block itself.
The translators require both PMOS and NMOS transistors and must be
capable of driving large capacitive loads. Layout of the translators is
exceedingly difficult, especially because the overall layout needs to be as
small as possible.

2.3.2 NOR Driver

The second type of wordline driver is similar to the bootstrap driver in
that two decode terms drive the output transistor from separate terminals.
The NOR driver, as shown in Figure 2.28, uses a PMOS transistor for Ml
and does not rely on bootstrapping to derive the gate bias. Rather, the gate is
driven by a voltage translator that converts DEC from VCc to VCCP voltage
levels. This conversion is necessary to ensure that Ml remains OFF for
unselected wordlines as the PHASE signal, which is shared by multiple
drivers, is driven to VCCP.

V
V CCP V
v CCP

PHASE

Figure 2.28 NOR driver.

58 Chap. 2 The DRAM Array

To fire a specific wordline, DEC must be HIGH and the appropriate

PHASE must fire HIGH. Generally, there are four to eight PHASE signals
per row decoder block. The NOR driver requires one level translator for
each PHASE and DEC signal. By comparison, the bootstrap driver only
requires level translators for the PHASE signal.

2.3.3 CMOS Driver

The final wordline driver configuration is shown in Figure 2.29. In gen­
eral, it lacks a specific name: it is sometimes referred to as a CMOS inverter
driver or a CMOS driver. Unlike the first two drivers, the output transistor
Ml has its source terminal permanently connected to Vy This driver,
therefore, features a voltage translator for each and every wordline. Both
decode terms DEC and PHASE* combine to drive the output stage through
the translator. The advantage of this driver, other than simple operation, is
low power consumption. Power is conserved because the translators drive
only the small capacitance associated with a single driver. The PHASE
translators of both the bootstrap and NOR drivers must charge considerable
junction capacitance. The disadvantages of the CMOS driver are layout
complexity and standby leakage current. Standby leakage current is a prod­
uct of Vccp voltage applied to Ml and its junction and subthreshold leakage
currents. For a large DRAM with high numbers of wordline drivers, this
leakage current can easily exceed the entire standby current budget unless
great care is exercised in designing output transistor M1.

2.3.4 Address Decode Tree

With the wordline driver circuits behind us, we can turn our attention to
the address decoder tree. There is no big secret to address decoding in the
row decoder network. Just about any type of logic suffices: static, dynamic,
Sec. 2.3 Row Decoder Elements 59

pass gate, or a combination thereof With any type of logic, however, the
primary objectives in decoder design are to maximize speed and minimize
die area. Because a great variety of methods have been used to implement
row address decoder trees, it is next to impossible to cover them all. Instead,
we will give an insight into the possibilities by discussing a few of them.
Regardless of the type of logic with which a row decoder is imple­
mented, the layout must completely reside beneath the row address signal
lines to constitute an efficient, minimized design. In other words, the metal
address tracks dictate the die area available for the decoder. Any additional
tracks necessary to complete the design constitute wasted silicon. For
DRAM designs requiring global row decode schemes, the penalty for ineffi­
cient design may be insignificant; however, for distributed local row decode
schemes, the die area penalty may be significant. As with mbits and sense
amplifiers, time spent optimizing row decode circuits is time well spent.

2.3.5 Static Tree

The most obvious form of address decode tree uses static CMOS logic.
As shown in Figure 2.30, a simple tree can be designed using two-input
NAND gates. While easy to design schematically, static logic address trees
are not popular. They waste silicon and are very difficult to lay out effi­
ciently. Static logic requires two transistors for each address term, one
NMOS and one PMOS, which can be significant for many address terms.
Furthermore, static gates must be cascaded to accumulate address terms,
adding gate delays at each level. For these and other reasons, static logic
gates are not used in row decode address trees in today's state-of-the-art
DRAMs.

Figure 2.30 Static decode tree.

236 P&E Tree

The second type of address tree uses dynamic logic, the most prevalent
being precharge and evaluate (P&E) logic. Used by the majority of DRAM
manufacturers, P&E address trees come in a variety of configurations,
although the differences between them can be subtle. Figure 2.31 shows a
simplified schematic for one version of a P&E address tree designed fbr use
with bootstrapped wordline drivers. P&E address tree circuits feature one or
60 Chap. 2 The DRAM Array

more PMOS PRECHARGE transistors and a cascade of NMOS ENABLE

transistors M2-M4. This P&E design uses half of the transistors required by
the static address tree of Figure 2.30. As a result, the layout of the P&E tree
is much smaller than that of the static tree and fits more easily under the
address lines. The PRE transistor is usually driven by a PRECHARGE* sig­
nal under the control of the RAS* chain logic. PRECHARGE^ and transistor
Ml ensure that DEC* is precharged HIGH, disabling the wordline driver
and preparing the tree for row address activation.
M7 is a weak PMOS transistor driven by the DEC inverter (M5 and
M6). Together, M7 and the inverter form a latch to ensure that DEC*
remains HIGH for all decoders that are not selected by the row addresses.
At the beginning of any RAS^ cycle, PRECHARGE* is LOW and the row
addresses are all disabled (LOW). After RAS^ falls, PRECHARGE^ transi­
tions HIGH to turn OFF Ml; then the row addresses are enabled. If
RAX-RA3 all go HIGH, then M2-M4 turn ON, overpowering M7 and driv­
ing DEC* to ground and subsequently DEC to Kcc. The output of this tree
segment normally drives four bootstrapped wordline drivers, each con­
nected to a separate PHASE signal. Therefore, for an array with 256 word­
lines, there will be 64 such decode trees.

2.3.7 Predecoding
The row address lines shown as RAX-RA2)can be either true and com­
plement or predecoded. Predecoded address lines are formed by logically
combining (AND) addresses as shown in Table 2.1.
Sec. 2.4 Discussion 61

The advantages of using predecoded addresses include lower power

(fewer signals make transitions during address changes) and higher effi­
ciency (only three transistors are necessary to decode six addresses for the
circuit of Figure 2.31). Predecoding is especially beneficial in redundancy
circuits. In fact, predecoded addresses are used throughout most DRAM
designs today.

Table 2.1 Predecoded address truth table.

RAO RAI PR01<n> PR01<0> PRO1<1> PR01<2> PR01<3>

0 0 0 1 0 0 0

1 0 1 0 1 0 0

0 1 2 0 0 1 0

1 1 3 0 0 0 1

2.3.8 Pass Transistor Tree

The final type of address tree to be examined, shown in Figure 2.32,
uses pass transistor logic. Pass transistor address trees are similar to P&E
trees in numerous ways. Both designs use PMOS PRECHARGE transistors
and NMOS address ENABLE transistors. Unlike P&E logic, however, the
NMOS cascade does not terminate at ground. Rather, the cascade of
M2-M4 goes to a PHASE* signal, which is HIGH during PRECHARGE
and LOW during EVALUATE. The address signals operate the same as in
the P&E tree: HIGH to select and LOW to deselect. The pass transistor tree
is shown integrated into a CMOS wordline drive匚 This is necessary because
the pass transistor tree and the CMOS wordline driver are generally used
together and their operation is complementary. The cross-coupled PMOS
transistors of the CMOS level translator provide a latch necessary to keep
the final interstage node biased at Vcc. Again the latch has a weak pull-up,
which is easily overpowered by the cascaded NMOS ENABLE transistors.
A pass transistor address tree is not used with bootstrapped wordline drivers
because the PHASE signal feeds into the address tree logic rather than into
the driver, as required by the bootstrap driver.

2.4 DISCUSSION
We have briefly examined the basic elements required in DRAM row
decoder blocks. Numerous variations are possible. No single design is best
fbr all applications. As with sense amplifiers, design depends on technology
and performance and cost trade-offs.
62 Chap. 2 The DRAM Array

Figure 2.32 Pass transistor decode tree.

REFERENCES

[1] K, Itoh, “Trends in megabit DRAM circuit design,“ IEEE Journal of Solid­
State Circuits, vol. 25, pp. 778-791, June 1990.
[2] D. Takashima, S. Watanabe, H. Nakano, Y. Oowaki, and K. Ohuchi,
“Open/fblded bit-line arrangement for ultra-high-density DRAMs,“ IEEE
Journal of Solid-State Circuits, vol. 29, pp. 539-542, April 1994.
[3] Hideto Hidaka, Yoshio Matsuda, and Kazuyasu Fujishima,44A divided/shared
bit-line sensing scheme for ULSI DRAM cores/' IEEE Journal of Solid-State
Circuits, vol. 26, pp. 473-477, April 1991.
[4] M. Aoki, Y. Nakagome, M. Horiguchi, H. Tanaka, S. Ikenaga, J. Etoh,
Y. Kawamoto, S. Kimura, E. Takeda, H. Sunami, and K. Itoh, “A 60-ns 16-
Mbit CMOS DRAM with a transposed data-line stnicture,11 IEEE Journal of
Solid-State Circuits, vol. 23, pp. 1113-1119, October 1988.
[5] R. Kraus and K. Hoffmann, ''Optimized sensing scheme of DRAMs,“ IEEE
Journal of Solid-State Circuits, vol. 24, pp. 895-899, August 1989.
[6] T. Yoshihara, H. Hidaka, Y. Matsuda, and K. Fujishima, “A twisted bitline
technique for multi-Mb DRAMs/' in 1988 IEEE ISSCC Digest of Technical
Papers, pp. 238-239.
[7] Yukihito Oowaki, Kenji Tsuchida, Yohji Watanabe, Daisaburo Takashima,
Masako Ohta, Hiroaki Nakano, Shigeyoshi Watanabe, Akihiro Nitayama,
Fumio Horiguchi, Kazunori Ohuchi, and Fujio Masuoka, “A 33-ns 64Mb
DRAM," IEEE Journal of Solid-State Circuits, vol. 26, pp. 1498-1505,
November 1991.
Sec. 2.4 Discussion 63

[8] M. Inoue, H. Kotani, T. Yamada, H. Yamauchi, A. Fujiwara, J. Matsushima,

H. Akamatsu, M. Fukumoto, M. Kubota, I. Nakao, N. Aoi, G. Fuse, S. Ogawa,
S. Odanaka. A. Ueno, and Y. Yamamoto, UA 16Mb DRAM with an open bit­
line architecture,“ in 1988 IEEE ISSCC Digest of Technical Papers,
pp. 246-247.
[9] Y. Kubota, Y. Iwase, K. Iguchi, J. Takagi, T. Watanabe, and K. Sakiyama,
uAlternatively-activated open bitline technique for high density DRAM's,"
IEICE Trans. Electron,, vol. E75-C,pp. 1259-1266, October 1992.
[10] T. Hamada, N. Tanabe, H. Watanabe, K. Takeuchi, N. Kasai, H. Hada,
K. Shibahara, K. Tokashiki, K. Nakajima, S. Hirasawa, E. Ikawa, T. Saeki,
E. Kakehashi, S. Ohya, and T. Kunio, UA split-level diagonal bit-line (SLDB)
stacked capacitor cell for 256Mb DRAMs,1992 IEDM Technical Digest,
pp. 799-802.
[11] Toshinori Morihara, Yoshikazu Ohno, Takahisa Eimori, Toshiharu Katayama,
Shinichi Satoh, Tadashi Nishimura, and Hirokazu Miyoshi, “Disk-shaped
stacked capacitor cell for 256Mb dynamic random-access memory,Japan
Journal ofApplied Physics, vol. 33, Part 1, pp. 4570-4575, August 1994.
[12] J. H. Ahn, Y. W. Park, J. H. Shin, S. T. Kim, S. P. Shim, S. W. Nam,
W. M. Park, H. B. Shin, C. S. Choi, K. T. Kim, D. Chin, O. H. Kwon, and
C. G Hwang, “Micro villus patterning (MVP) technology for 256Mb DRAM
stack cell," in 1992 Symposium on VLSI Tech. Digest of Technical Papers,
pp. 12-13.
[13] Kazuhiko Sagara, Tokuo Kure, Shoji Shukuri, Jiro Yiigami, Norio Hasegawa,
Hidekazu Goto, and Hisaomi Yamashita, c<Recessed memory array technology
for a double cylindrical stacked capacitor cell of256M DRAM,“ IEICE Trans.
Electron., vol. E75-C, pp. 1313-1322, November 1992.
[14] S. Ohya. 1994. Semiconductor memory device having stacked-type capacitor
of large capacitance. US Patent 5,298,775, March 29, 1994.
[15] M. Sakao, N. Kasai, T. Ishijima, E. Ikawa, H. Watanabe, K. Terada, and
T. Kikkawa, UA capacitor-over-bit-line (COB) cell with hemispherical-grain
storage node for 64Mb DRAMs,“ in 1990 IEDM Technical Digest,
pp. 655-658.
[16] G Bronner, H. Aochi, M. Gall, J. Gambino, S. Gernhardt, E. Hammerl, H. Ho,
J. Iba, H. Ishiuchi, M. Jaso, R. Kleinhenz, T. Mii, M. Narita, L. Nesbit,
W. Neumueller, A. Nitayama, T. Ohiwa, S. Parke, J. Ryan, T. Sato, H. Takato,
and S. Yoshikawa, “A fully planarized 0.25|im CMOS technology for
256Mbit DRAM and beyond,M in 1995 Symposium on VLSI Tech. Digest of
Technical Papers, pp. 15-16.
[17] N. C.-C. Lu and H. H. Chao, "Half Fpp / bit-line sensing scheme in CMOS
DRAMs,“ in IEEE Journal of Solid-State Circuits, vol. SC19, p. 451, August
1984.
[18] E. Adler; J. K. DeBrosse; S. F. Geissler; S. J. Holmes; M. D. Jaffe;
J. B. Johnson; C. W. Koburger, III; J. B. Lasky; B. Lloyd; G L. Miles;
64 Chap. 2 The DRAM Array

J. S. Nakos; W. P. Noble, Jr.; S. H. Vbldman; M. Armacost; and R. Ferguson;

“The evolution of IBM CMOS DRAM technology,n IBM Journal ofResearch
and Development, vol. 39, pp. 167-188, March 1995.
[19] R. Kraus, “Analysis and reduction of sense-amplifier offset,n in IEEE Journal
ofSolid-State Circuits, vol. 24, pp. 1028-1033, August 1989.
[20] M. Asakura, T. Ohishi, M. Tsukude, S. Tomishima, H. Hidaka, K. Arimoto,
K. Fujishima, T. Eimori, Y. Ohno, T. Nishimura, M. Yasunaga, T. Kondoh,
S. I. Satoh, T. Yoshihara, and K. Demizu, “A 34ns 256Mb DRAM with
boosted sense-ground scheme/' in 1994 IEEE ISSCC Digest of Technical
Papers, pp. 140-141.
[21] T. Ooishi, K. Hamade, M. Asakura, K. Yasuda, H. Hidaka, H. Miyamoto, and
H. Ozaki, uAn automatic temperature compensation of internal sense ground
fbr sub-quarter micron DRAMs J in 1994 Symposium on VLSI Circuits Digest
of Technical Papers, pp. 77-78.
[22] K. Noda, T. Saeki, A. Tsujimoto, T. Murotani, and K. Koyama, “A boosted
dual word-line decoding scheme for 256 Mb DRAMs,” in 1992 Symposium on
VLSI Circuits Digest of Technical Papers, pp. 112-113.
 Chapter
3

Array Architectures
This chapter presents a detailed description of the two most prevalent array
architectures under consideration for future large-scale DRAMs: the afore­
mentioned open architectures and olded digitline architectures.

3.1 TYPES OF ARRAY ARCHITECTURES

To provide a viable point for comparison, each architecture is employed in
the theoretical construction of 32-Mbit memory blocks for use in a 256-
Mbit DRAM. Design parameters and layout rules from a typical 0.25 gm
DRAM process provide the necessary dimensions and constraints for the
analysis. Some of these parameters are shown in Table 3.1. By examining
DRAM architectures in the light of a real-world design problem, an objec­
tive and unbiased comparison can be made. In addition, using this approach,
we readily detect the strengths and weaknesses of either architecture.

3.1.1 Open Digitline Array Architecture

The open digitline array architecture was the prevalent architecture
prior to the 64kbit DRAM. A modem embodiment of this architecture, as
shown in Figure 3.1 [1] [2], is constructed with multiple crosspoint array
cores separated by strips of sense amplifier blocks in one axis and either
row decode blocks or wordline stitching regions in the other axis. Each
128kbit array core is built using 6F2 Mbit cell pairs. There are 131,072 (2刃
functionally addressable Mbits arranged in 264 rows and 524 digitlines. In
the 264 rows, there are 256 actual wordlines, 4 redundant wordlines, and 4
dummy wordlines. In the 524 digitlines, there are 512 actual digitlines, 8
redundant digitlines, and 4 dummy digitlines. Photolithography problems
usually occur at the edge of large repetitive structures, such as Mbit arrays.

65
66 Chap. 3 Array Architectures

These problems produce malformed or nonunifbrm structures, rendering the

edge cells useless. Therefore, including dummy Mbits, wordlines, and digit­
lines on each array edge ensures that these problems occur only on dummy
cells, leaving live cells unaffected. Although dummy structures enlarge
each array core, they also significantly improve device yield. Thus, they are
necessary on all DRAM designs.

Table 3.1 0.25 pm design parameters.

Parameter Value

Digitline width WDL 0.3 pm

Digitline pitch PDL 0.6 pm

Wordline width Wwl 0.3 gm

Wordline pitch fbr 8F2 0.6 gm

Mbit PWL%

Wordline pitch for 6F2 0.9 gm

Mbit Pwls

Cell capacitance Cc 30fF

Digitline capacitance 0.8fF

per Mbit Cdm

Wordline capacitance 0.6fF

per 8F2 Mbit

Wordline capacitance 0.5fF

per 6F2 Mbit Cwh

Wordline sheet 6D/sq

resistance

Array core size, as measured in the number of Mbits, is restricted by

two factors: a desire to keep the quantity of Mbits a power of two and the
practical limits on wordline and digitline length. The need for a binary
quantity of Mbits in each array core derives from the binary nature of
DRAM addressing. Given N row addresses and M column addresses for a
given part, there are 2n+m addressable Mbits. Address decoding is greatly
simplified within a DRAM if array address boundaries are derived directly
from address bits.
Sec. 3.1 Types of Array Architectures 67

Figure 3.1 Open digitline architecture schematic.

Because addressing is binary, the boundaries naturally become binary.

Therefore, the size of each array core must necessarily have 2X addressable
rows and 2丫 addressable digitlines. The resulting anay core size is 2x+y
Mbits, which is, of course, a binary number.
The second set of factors limiting array core size involves practical lim­
its on digitline and wordline length. From earlier discussions in Section 2.1,
the digitline capacitance is limited by two factors. First, the ratio of cell
capacitance to digitline capacitance must fall within a specified range to
ensure reliable sensing. Second, operating current and power for the DRAM
is in large part determined by the cunent required to charge and discharge
the digitlines during each active cycle. Power considerations restrict digit­
line length for the 256-Mbit generation to approximately 128 Mbit pairs
(256 rows), with each Mbit connection adding capacitance to the digitline.
The power dissipated during a Read or Refresh operation is proportional to
the digitline capacitance (Cd), the supply voltage (VCc)f the number of
active columns (N), and the Refresh period (P). Accordingly, the power dis­
sipated is given as

N(Cd+C0
pd = vc& ---- 2P--- watts
(3.1)
On a 256-Mbit DRAM in 8k (rows) Refresh, there are 32,768 (215) active
columns during each Read, Write, or Refresh operation. The active array
current and power dissipation for a 256-Mbit DRAM appear in Table 3.2
for a 90 ns Refresh period (-5 timing) at various digitline lengths. The bud­
68 Chap. 3 Array Architectures

get for the active array current is limited to 200mA for this 256-Mbit
design. To meet this budget, the digitline cannot exceed a length of
256 Mbits.
Wordline length, as described in Section 2.1, is limited by the maximum
allowable RC time constant of the wordline. To ensure acceptable access
time fbr the 256-Mbit DRAM, the wordline time constant should be kept
below 4 nanoseconds. For a wordline connected to N Mbits, the total resis­
tance and capacitance follow:

L ”

(3.2)

CWL =。印g - Nfarads

(3.3)
where PDL is the digitline pitch, WLW is the wordline width, and Cw^ is the
wordline capacitance in an 8F2 Mbit cell.

Table 3.2 Active current and power versus digitline length.

Digitline Power
Digitline Active Current
Capacitance Dissipation
Length (Mbits) (mA)
g (mW)

128 102 60 199

256 205 121 398

512 410 241 795

Table 3.3 contains the effective wordline time constants fbr various
wordline lengths. As shown in the table, the wordline length cannot exceed
512 Mbits (512 digitlines) if the wordline time constant is to remain under 4
nanoseconds.
The open digitline architecture does not support digitline twisting
because the true and complement digitlines, which constitute a column, are
in separate array cores. Therefore, no silicon area is consumed for twist
regions. The 32-Mbit array block requires a total of two hundred fifty-six
128kbit array cores in its construction. Each 32-Mbit block represents an
address space comprising a total of 4,096 rows and 8,192 columns. A prac­
tical configuration for the 32-Mbit block is depicted in Figure 3.2.
Sec. 3.1 Types of Array Architectures

Table 3.3 Wordline time constant versus wordline length.

Wordline Time Constant

Rwl (ohms) 。牝（伍）
Length (Mbits) (ns)

128 1,536 64 0.098

256 3,072 128 0.39

512 6,144 256 1.57

1024 12,288 512 6.29

In Figure 3.2, the 256 array cores appear in a 16 x 16 arrangement. The

xl6 arrangement produces 2-Mbit sections of 256 wordlines and 8,192 dig­
itlines (4,096 columns), Sixteen 2-Mbit sections are required to form the
complete 32-Mbit block: sense amplifier strips are positioned vertically
between each 2-Mbit section, and row decode strips or wordline stitching
strips are positioned horizontally between each array core.
4.224 Wordlines

IMI
w

s
x.!gg
・ ・ ・ ・

:!ig
ig
1

:!
ig
:!
ig
ig

Figure 3.2 Open digitline 32-Mbit anay block.

Layout can be generated fbr the various 32-Mbit elements depicted in

Figure 3.2. This layout is necessary to obtain reasonable estimates for pitch
cell size. With these size estimates, overall dimensions of the 32-Mbit mem­
ory block can be calculated. The results of these estimates appear in
70 Chap. 3 Array Architectures

Table 3.4. Essentially, the overall height of the 32-Mbit block can be found
by summing the height of the row decode blocks (or stitch regions) together
with the product of the digitline pitch and the total number of digitlines.
Accordingly,

HeightM = (TR - HLDEC) PDl) microns

(3.4)

where TR is the number of local row decoders, HLDEC is the height of each
decoder, TDL is the number of digitlines including redundant and dummy
lines, and PDL is the digitline pitch. Similarly, the width of the 32-Mbit
block is found by summing the total width of the sense amplifier blocks
with the product of the wordline pitch and the number of wordlines. This bit
of math yields

Width32 = (Tsa- Wamp) + (Twl- P%6)microns

(3.5)
where TSA is the number of sense amplifier strips, WAMP is the width of the
sense amplifiers, TWL is the total number of wordlines including redundant
and dummy lines, and PWl^ is the wordline pitch for the 6F2 Mbit.
Table 3.4 contains calculation results fbr the 32-Mbit block shown in
Figure 3.2. Although overall size is the best measure of architectural effi­
ciency, a second popular metric is array efficiency. Array efficiency is deter­
mined by dividing the area consumed by functionally addressable Mbits by
the total die area. To simplify the analysis in this text, peripheral circuits are
ignored in the array efficiency calculation. Rather, the calculation considers
only the 32-Mbit memory block, ignoring all other factors. With this simpli­
fication, the array efficiency fbr a 32-Mbit block is given as

(100 -2 255
Efficiency = ---------------------------- percent
Areas!

(3.6)
where 225 is the number of addressable Mbits in each 32-Mbit block. The
open digitline architecture yields a calculated array efficiency of 51.7%.
Unfortunately, the ideal open digitline architecture presented in
Figure 3.2 is difficult to realize in practice. The difficulty stems from an
interdependency between the memory array and sense amplifier layouts in
which each array digitline must connect to one sense amplifier and each
sense amplifier must connect to two array digitlines.
Sec. 3.1 Types of Array Architectures 71

Table 3.4 Open digitline (local row decode): 32-Mbit size calculations.

Description Parameter Size

Number of sense tsa 17

amplifier strips

Width of sense WAMp 88 pm

amplifiers

Number of local Tldec 17

decode strips

Height of local decode Hldec 93 pm

strips

Number of digitlines Tdl 8,400

Number of wordlines Twl 4,224

Height of 32-Mbit Height! 6,621 gm

block

Width of 32-Mbit Width32 5,298 gm

block

Area of 32-Mbit block AreaYL 35,078,058 jim2

This interdependency, which exists for all array architectures, becomes

problematic for the open digitline architecture. The two digitlines, which
connect to a sense amplifier, must come from two separate memory arrays.
As a result, sense amplifier blocks must always be placed between memory
arrays for open digitline array architectures [3], unlike the depiction in
Figure 3.2.
Two layout approaches may be used to achieve this goal. First, design
the sense amplifiers so that the sense amplifier block contains a set of sense
amplifiers fbr each digitline in the array. This single-pitch solution, shown
in Figure 3.3, eliminates the need for sense amplifiers on both sides of an
array core because all of the digitlines connect to a single sense amplifier
block. Not only does this solution eliminate the edge problem, but it also
reduces the 32-Mbit block size. There are now only eight sense amplifier
strips instead of the seventeen of Figure 3.2. Unfortunately, it is nearly
impossible to lay out sense amplifiers in this fashion [4]. Even a single­
metal sense amplifier layout, considered the tightest layout in the industry,
achieves only one sense amplifier for every two digitlines (double-pitch).
72 Chap. 3 Array Architectures

Figure 3.3 Single-pitch open digitline architecture.

A second approach to solving the interdependency problem in open dig­

itline architectures is to maintain the configuration shown in Figure 3.2 but
include some form of reference digitline fbr the edge sense amplifiers. The
reference digitline can assume any form as long as it accurately models the
capacitance and behavior of a true digitline. Obviously, the best type of ref­
erence digitline is a true digitline. Therefore, with this approach, more
dummy array cores are added to both edges of the 32-Mbit memory block,
as shown in Figure 3.4. The dummy array cores need only half as many
wordlines as the true array core because only half of the digitlines are con­
nected to any single sense amplifier strip. The unconnected digitlines dou­
ble the effective length of the reference digitlines.
This approach solves the array edge problem. However, by producing a
larger 32-Mbit memory block, array efficiency is reduced. Dummy arrays
solve the array edge problem inherent in open digitline architecture but
require sense amplifier layouts that are on the edge of impossible. The prob­
lem of sense amplifier layout is all the more difficult because global column
select lines must be routed through. For all intents and purposes, therefore,
the sense amplifier layout cannot be completed without the presence of an
additional conductor, such as a third metal, or without time multiplexed
sensing. Thus, fbr the open digitline architecture to be successful, an addi­
tional conductor must be added to the DRAM process.
Sec. 3.1 Types of Array Architectures 73

Figure 3.4 Open digitline architecture with dummy arrays.

With the presence of the additional conductor, such as MetaI3, the sense
amplifier layout and either a full or hierarchical global row decoding
scheme is made possible. A full global row decoding scheme using word­
line stitching places great demands on metal and contact/via technologies;
however, it represents the most efficient use of the additional metal. Hierar­
chical row decoding using bootstrap wordline drivers is slightly less effi­
cient. Wordlines no longer need to be strapped with metal on pitch, and,
thus, process requirements are relaxed significantly [5].
For a balanced perspective, both global and hierarchical approaches are
analyzed. The results of this analysis for the open digitline architecture are
summarized in Tables 3.5 and 3.6. Array efficiency for global and hierarchi­
cal row decoding calculate to 60.5% and 55.9%, respectively, for the 32
Mbit memory blocks based on data from these tables.
74 Chap. 3 Array Architectures

Table 3.5 Open digitline (dummy arrays and global row decode): 32-Mbit size
calculations.

Description Parameter Size

Number of sense Tsa 17

amplifier strips

Width of sense wAMP 88 pm

amplifiers

Number of global Tgdec 1

decode strips

Height of global Hgdec 200呷

decode strips

Number of stitch Nst 17

regions

Height of stitch regions Hst 10 |im

Number of digitlines Tdl 8,400

Number of wordlines Twl 4,488

Height of 32-Mbit Height3! 5,410 pm

block

Width of 32-Mbit WidMl 5,535 pm

block

Area of 32-Mbit block Area32 29,944,350 呷2

Sec. 3.1 Types of Array Architectures 75

Table 3.6 Open digitline (dummy arrays and hierarchical row decode): 32-Mbit
size calculations.

Description Parameter Size

Number of sense Tsa 17

amplifier strips

Width of sense Mmp 88 gm

amplifiers

Number of global T'gdec 1

decode strips

Height of global Hgdec 190 gm

decode strips

Number ofhier decode Thdec 17

strips

Height ofhier decode Hhdec 37 gm

strips

Number of digitlines T,dl 8,400

Number of wordlines Twl 4,488

Height of 32-Mbit Height32 5,859 岬

block

Width of 32-Mbit Widths 5,535 gm

block

Area of 32-Mbit block Area32 32,429,565 |im2

3.1.2 Folded Array Architecture

The folded array architecture depicted in Figure 3.5 is the standard
architecture of many modem DRAM designs. The folded architecture is
constructed with multiple array cores separated by strips of sense amplifiers
and either row decode blocks or wordline stitching regions. Unlike the open
digitline architecture, which uses 6F2 Mbit cell pairs, the folded array core
uses 8F2 Mbit cell pairs [6]. Modem array cores include 262,144 (218) fiinc-
tionally addressable Mbits arranged in 532 rows and 1,044 digitlines. In the
532 rows, there are 512 actual wordlines, 4 redundant wordlines, and 16
dummy wordlines. Each row (wordline) connects to Mbit transistors on
alternating digitlines. In the 1,044 digitlines, there are 1,024 actual digit­
lines (512 columns), 16 redundant digitlines (8 columns), and 4 dummy dig­
itlines. As discussed in Section 3.1.1, because of the additive or subtractive
photolithography effects, dummy wordlines and digitlines are necessary to
76 Chap. 3 Array Architectures

guardband live digitlines. These photo effects are pronounced at the edges
of large repetitive structures such as the array cores.
Sense amplifier blocks are placed on both sides of each array core. The
sense amplifiers within each block are laid out at quarter pitch: one sense
amplifier for every four digitlines. Each sense amplifier connects through
isolation devices to columns (digitline pairs) from both adjacent array cores.
Odd columns connect on one side of the core, and even columns connect on
the opposite side. Each sense amplifier block is therefore connected to only
odd or even columns and is never connected to both odd and even columns
within the same block. Connecting to both odd and even columns requires a
half^pitch sense amplifier layout: one sense amplifier for every two digit­
lines. While half-pitch layout is possible with certain DRAM processes, the
bulk of production DRAM designs remain quarter pitch due to the ease of
laying them out. The analysis presented in this chapter is accordingly based
on quarter-pitch design practices.

Row decode Row decode

Figure 3.5 Folded digitline array architecture schematic.

The location of row decode blocks fbr the array core depends on the
number of available metal layers. For one- and two-metal processes, local
row decode blocks are located at the top and bottom edges of the core. For
three- and four-metal processes, global row decodes are used. Global row
decodes require only stitch regions or local wordline drivers at the top and
bottom edges of the core [7]. Stitch regions consume much less silicon area
than local row decodes, substantially increasing array efficiency fbr the
DRAM. The anay core also includes digitline twist regions running parallel
to the wordlines. These regions provide the die area required for digitline
twisting. Depending on the particular twisting scheme selected fbr a design
Sec. 3.1 Types of Array Architectures 77

(see Section 2.1), the array core needs between one and three twist regions.
For the sake of analysis, a triple twist is assumed, as it offers the best overall
noise performance and has been chosen by some DRAM manufacturers fbr
select advanced large-scale applications [8]. Because each twist region con­
stitutes a break in the array structure, it is necessary to use dummy word­
lines. For this reason, there are 16 dummy wordlines (2 fbr each array edge)
in the folded array core rather than 4 dummy wordlines as in the open digit­
line architecture.
There are more Mbits in the array core fbr folded digitline architectures
than there are fbr open digitline architectures. Larger core size is an inherent
feature of folded architectures, arising from the very nature of the architec­
ture. The term folded architecture comes from the fact that folding two open
digitline array cores one on top of the other produces a folded array core.
The digitlines and wordlines from each folded core are spread apart (double
pitch) to allow room fbr the other folded core. After folding, each constitu­
ent core remains intact and independent, except for the Mbit changes (8F2
conversion) necessary in the folded architecture. The array core size dou­
bles because the total number of digitlines and wordlines doubles in the
folding process. It does not quadruple as one might suspect because the two
constituent folded cores remain independent: the wordlines from one folded
core do not connect to Mbits in the other folded core.
Digitline pairing (column formation) is a natural outgrowth of the fold­
ing process; each wordline only connects to Mbits on alternating digitlines.
The existence of digitline pairs (columns) is the one characteristic of folded
digitline architectures that produces superior signal-to-noise performance.
Furthermore, the digitlines that form a column are physically adjacent to
one another. This feature permits various digitline twisting schemes to be
used, as discussed in Section 2.1, further improving signal-to-noise perfor­
mance.
Similar to the open digitline architecture, digitline length fbr the folded
digitline architecture is again limited by power dissipation and minimum
cell-to-digitline capacitance ratio. For the 256-Mbit generation, digitlines
were restricted from connecting to more than 256 cells (128 Mbit pairs).
The analysis used to arrive at this quantity is similar to that fbr the open dig­
itline architecture. (Refer to Table 3.2 to view the calculated results of
power dissipation versus digitline length fbr a 256-Mbit DRAM in 8k
Refresh.) Wordline length is again limited by the maximum allowable RC
time constant of the wordline.
Contrary to an open digitline architecture in which each wordline con­
nects to Mbits on each digitline, the wordlines in a folded digitline architec­
ture connect to Mbits only on alternating digitlines. Therefore, a wordline
can cross 1,024 digitlines while connecting to only 512 Mbit transistors.
78 Chap. 3 Array Architectures

The wordlines have twice the overall resistance, but only slightly more
capacitance because they run over field oxide on alternating digitlines.
Table 3.7 presents the effective wordline time constants for various word­
line lengths for a folded array core. For a wordline connected to N Mbits,
the total resistance and capacitance follow:

Rwl = (2 - Rs - N - PD[) ohms

(3.7)

CWL = C/8 - Nfarads

(3.8)
where Pdl is the digitline pitch and is the wordline capacitance in an
8F2 Mbit cell. As shown in Table 3.7, the wordline length cannot exceed
512 Mbits (1,024 digitlines) fbr the wordline time constant to remain under
4 nanoseconds. Although the wordline connects to only 512 Mbits, it is two
times longer (1,024 digitlines) than wordlines in open digitline array cores.
The folded digitline architecture therefore requires half as many row decode
blocks or wordline stitching regions as the open digitline architecture.

Table 3.7 Wordline time constant versus wordline length (folded).

Wordline Length Time Constant

Rwl (ohms) CW(fF)
(Mbits) (ns)

128 3,072 77 0.24

256 6,144 154 0.95

512 12,288 307 3.77

1,024 24,576 614 15.09

A diagram of a 32-Mbit array block using folded digitline architecture

is shown in Figure 3.6. This block requires a total of one hundred twenty­
eight 256kbit array cores. In this figure, the array cores are arranged in an 8-
row and 16-column configuration. The x8 row arrangement produces 2-
Mbit sections of 256 wordlines and 8,192 digitlines (4,096 columns). A
total of sixteen 2-Mbit sections form the complete 32-Mbit array block.
Sense amplifier strips are positioned vertically between each 2-Mbit sec­
tion, as in the open digitline architecture. Again, row decode blocks or
wordline stitching regions are positioned horizontally between the array
cores.
Sec. 3.1 Types of Array Architectures 79

The 32-Mbit array block shown in Figure 3.6 includes size estimates for
the various pitch cells. The layout was generated where necessary to arrive
at the size estimates. The overall size for the folded digitline 32-Mbit block
can be found by again summing the dimensions for each component.
Accordingly,

Heights = (Jr - HRDEC) + (TDL - PD[) microns

(3.9)
where TR is the number of row decoders, HRDEC is the height of each
decoder, TDL is the number of digitlines including redundant and dummy,
and PDL is the digitline pitch. Similarly,

Width32 = (Tsa -町”尸)+ (丁醇,尸郎z,8)+ TWIST' 印T叼ST)microns

(3.10)
where TSA is the number of sense amplifier strips, WAMP is the width of the
sense amplifiers, TWL is the total number of wordlines including redundant
and dummy, PWL^ is the wordline pitch for the 8F2 Mbit, is the total
number of twist regions, and WTWlST is the width of the twist regions.
Table 3.8 shows the calculated results fbr the 32-Mbit block shown in
Figure 3.6. In this table, a double-metal process is used, which requires
local row decoder blocks. Note that Table 3.8 for the folded digitline archi­
tecture contains approximately twice as many wordlines as does Table 3.5
fbr the open digitline architecture. The reason for this is that each wordline
in the folded array only connects to Mbits on alternating digitlines, whereas
each wordline in the open array connects to Mbits on every digitline. A
folded digitline design therefore needs twice as many wordlines as a compa­
rable open digitline design.
Array efficiency for the 32-Mbit memory block from Figure 3.6 is again
found by dividing the area consumed by functionally addressable Mbits by
the total die area. For the simplified analysis presented in this text, the
peripheral circuits are ignored. Array efficiency fbr the 32-Mbit block is
therefore given as which yields 59.5% fbr the folded array design example.

(100 . 225.PD£.2P^8)
Efficiency - ------------ ------------------- percent

(3.11)
80 Chap. 3 Array Architectures

8,512 Wordlines

S
OJW
三

6
b
z
Lne

od

一
Row
7 256Kbit Sense I!
decode core amplifier 45 pm

Figure 3.6 Folded digitline architecture 32-Mbit array block.

With the addition of Metal3 to the DRAM process, either a global or

hierarchical row decoding scheme, similar to the open digitline analysis,
can be used. While global row decoding and stitched wordlines achieve the
smallest die size, they also place greater demands on the fabrication pro­
cess. For a balanced perspective, both approaches were analyzed for the
folded digitline architecture. The results of this analysis are presented in
Tables 3.9 and 3.10. Array efficiency for the 32-Mbit memory blocks using
global and hierarchical row decoding calculate to 74.0% and 70.9%, respec­
tively.
Sec. 3.1 Types of Array Architectures 81

Table 3.8 Folded digitline (local row decode): 32-Mbit size calculations.

Description Parameter Size

Number of sense Tsaf 17

amplifier strips

Width of sense W4Mp 45 pm

amplifiers

Number of local Tldec 9

decode strips

Height of local decode Hldec 93

strips

Number of digitlines Tdl 8,352

Number of wordlines Twl\ 8,512

Number of twist Ttwist 48

regions

Width of twist regions 犷TWIST 6 gm

Height of 32-Mbit Height32 6,592 gm

block

Width of 32-Mbit Width32 6,160 pm

block

Area of 32-Mbit block Areail 40,606,720 gm2

82 Chap. 3 Array Architectures

Table 3.9 Folded digitline (global decode): 32-Mbit size calculations.

Description Parameter Size

Number of sense Tsa 17

amplifier strips

Width of sense 叫MP 45 pm

amplifiers

Number of global T,gdec 1

decode strips

Height of global Hgdec 200 gm

decode strips

Number of stitch Nst 9

regions

Height of stitch regions Hst 10岬

Number of digitlines Tdl 8,352

Number of wordlines Twl 8,512

Number of twist TWIST 48

regions

Width of twist regions TWIST 6 gm

Height of 32-Mbit Height32 5,301 |xm

block

Width of 32-Mbit Widths 6,160 gm

block

Area of 32-Mbit block Area32 32,654,160 Hm2

Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 83

Table 3.10 Folded digitline (hierarchical row decode): 32-Mbit size

calculations.

Description Parameter Size

Number of sense Tsa 17

amplifier strips

Width of sense ^AMP 45 p.m

amplifiers

Number of global Tgdec 1

decode strips

Height of global Hgdec 190 pm

decode strips

Number of hier Nhdec 9

decode strips

Height of hier Hhec 37 pm

decode strips

Number of Tdl 8,352

digitlines

Number of Twl 8,512

wordlines

Number of twist Ttwist 48

regions

Width of twist TWIST 6 gm

regions

Height of 32-Mbit Height32 5,534 pm

block

Width of 32-Mbit Width32 6,160 pm

block

Area of 32-Mbit AredSI 34,089,440

block

3.2 DESIGN EXAMPLES: ADVANCED BILEVEL DRAM

ARCHITECTURE
This section introduces a novel, advanced architecture possible fbr use on
future large-scale DRAMs. First, we discuss technical objectives for the
proposed architecture. Second, we develop and describe the concept for an
advanced array architecture capable of meeting these objectives. Third, we
conceptually construct a 32-Mbit memory block with this new architecture
84 Chap. 3 Array Architectures

fbr a 256-Mbit DRAM. Finally, we compare the results achieved with the
new architecture to those obtained fbr the open digitline and folded digitline
architectures from Section 3.1.

3.2.1 Array Architecture Objectives

Both the open digitline and folded digitline architectures have distinct
advantages and disadvantages. While open digitline architectures achieve
smaller array layouts by virtue of using smaller 6F2 Mbit cells, they suffer
from poor signal-to-noise performance. A relaxed wordline pitch, which
stems from the 6F2 Mbit, simplifies the task of wordline driver layout.
Sense amplifier layout, however, is difficult because the array configuration
is inherently half pitch: one sense amplifier for every two digitlines. The
superior signal-to-noise performance [9] of folded digitline architectures
comes at the expense of larger, less efficient array layouts. Good signal-to-
noise performance stems from the adjacency of true and complement digit­
lines and the capability to twist these digitline pairs. Sense amplifier layout
is simplified because the array configuration is quarter pitch— —that is, one
sense amplifier for every four digitlines. Wordline driver layout is more dif­
ficult because the wordline pitch is effectively reduced in folded architec­
tures.
The main objective of the new array architecture is to combine the
advantages, while avoiding the disadvantages, of both folded and open dig­
itline architectures. To meet this objective, the architecture needs to include
the following features and characteristics:
• Open digitline Mbit configuration
• Small 6F2 Mbit
• Small, efficient array layout
• Folded digitline sense amplifier configuration
• Adjacent true and complement digitlines
• Twisted digitline pairs
• Relaxed wordline pitch
• High signal-to-noise ratio
An underlying goal of the new architecture is to reduce overall die size
beyond that obtainable from either the folded or open digitline architec­
tures. A second yet equally important goal is to achieve signal-to-noise per­
formance that meets or approaches that of the folded digitline architecture.
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 85

3.2.2 Bilevel Digitline Construction

A bilevel digitline architecture resulted from 256-Mbit DRAM research
and design activities carried out at Micron Technology, Inc., in Boise,
Idaho. This bilevel digitline architecture is an innovation that evolved from
a comparative analysis of open and folded digitline architectures. The anal­
ysis served as a design catalyst, ultimately leading to the creation of a new
DRAM array configuration—one that allows the use of 6F2 Mbits in an oth­
erwise folded digitline array configuration. These memory cells are a by­
product of crosspoint-style (open digitline) array blocks. Crosspoint-style
array blocks require that every wordline connect to Mbit transistors on
every digitline, precluding the formation of digitline pairs. Yet, digitline
pairs (columns) remain an essential element in folded digitline-type opera­
tion. Digitline pairs and digitline twisting are important features that pro­
vide fbr good signal-to-noise performance.
The bilevel digitline architecture solves the crosspoint and digitline pair
dilemma through vertical integration. In vertical integration, essentially,
two open digitline crosspoint array sections are placed side by side, as seen
in Figure 3.7. Digitlines in one array section are designated as true digit­
lines, while digitlines from the second array section are designated as com­
plement digitlines. An additional conductor is added to the DRAM process
to complete the formation of the digitline pairs. The added conductor allows
digitlines from each array section to route across the other array section
with both true and complement digitiines vertically aligned. At the juncture
between each section, the true and complement signals are vertically
twisted. With this twisting, the true digitline connects to Mbits in one array
section, and the complement digitline connects to Mbits in the other array
section. This twisting concept is illustrated in Figure 3.8.
lb improve the signal-to-noise characteristics of this design, the single
twist region is replaced by three twist regions, as illustrated in Figure 3.9. A
benefit of adding multiple twist regions is that only half of the digitline
pairs actually twist within each region, making room in each region fbr the
twists to occur. The twist regions are equally spaced at the 25%, 50%, and
75% marks in the overall array. Assuming that even digitline pairs twist at
the 50% mark, odd digitlines twist at the 25% and 75% marks. Each compo­
nent of a digitline pair, true and complement, spends half of its overall
length on the bottom conductor connecting to Mbits and half of its length on
the top conductor. This characteristic balances the capacitance and the num­
ber of Mbits associated with each digitline. Furthermore, the triple twisting
scheme guarantees that the noise terms are balanced fbr each digitline, pro­
ducing excellent signal-to-noise performance.
86 Chap. 3 Array Architectures

Figure 3.7 Development of bilevel digitline architecture.

Figure 3.8 Digitline vertical twisting concept.

A variety of vertical twisting schemes is possible with the bilevel digit­

line architecture. As shown in Figure 3.10, each scheme uses conductive
layers already present in the DRAM process to complete the twist. Vertical
twisting is simplified because only half of the digitlines are involved in a
given twist region. The final selection of a twisting scheme is based on yield
factors, die size, and available process technology.
To further advance the bilevel digitline architecture concept, its 6F2
Mbit was modified to improve yield. Shown in an arrayed form in
Figure 3.11, the plaid Mbit is constructed using long parallel strips of active
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 87

area vertically separated by traditional field oxide isolation. Wordlines run

perpendicular to the active area in straight strips of polysilicon. Plaid Mbits
are again constructed in pairs that share a common contact to the digitline.
Isolation gates (transistors) formed with additional polysilicon strips pro­
vide horizontal isolation between Mbits. Isolation is obtained from these
gates by permanently connecting the isolation gate polysilicon to either a
ground or a negative potential. Using isolation gates in this Mbit design
eliminates one- and two-dimensional encroachment problems associated
with normal isolation processes. Furthermore, many photolithography prob­
lems are eliminated from the DRAM process as a result of the straight, sim­
ple design of both the active area and the polysilicon in the Mbit. The plaid
designation for this Mbit is derived from the similarity between an array of
Mbits and tartan fabric that is apparent in a color array plot.
In the bilevel and folded digitline architectures, both true and comple­
ment digitlines exist in the same array core. Accordingly, the sense ampli­
fier block needs only one sense amplifier fbr every two digitline pairs. For
the folded digitline architecture, this yields one sense amplifier for every
four Metal 1 digitlines-quarter pitch. The bilevel digitline architecture that
uses vertical digitline stacking needs one sense amplifier fbr every two
Metal 1 digitlines—
—half pitch. Sense amplifier layout is therefore more diffi­
cult for bilevel than for folded designs. The three-metal DRAM process
needed for bilevel architectures concurrently enables and simplifies sense
amplifier layout. Metal 1 is used for lower level digitlines and local routing
within the sense amplifiers and row decodes. Metal2 is available for upper
level digitlines and column select signal routing through the sense amplifi­
ers. Metal3 can therefore be used for column select routing across the arrays
and fbr control and power routing through the sense amplifiers. The func­
tion of Metal2 and Metal3 can easily be swapped in the sense amplifier
block depending on layout preferences and design objectives.

Figure 39 Bilevel digitline architecture schematic.

88 Chap. 3 Array Architectures

Figure 3.10 Vertical twisting schemes.

Wordline pitch is effectively relaxed for the plaid 6F2 Mbit of the
bilevel digitline architecture. The Mbit is still built using the minimum pro­
cess feature size of 0.3 [xm. The relaxed wordline pitch stems from struc­
tural differences between a folded digitline Mbit and an open digitline or
plaid Mbit. There are essentially four wordlines running across each folded
digitline Mbit pair compared to two wordlines running across each open
digitline or plaid Mbit pair. Although the plaid Mbit is 25% shorter than a
folded Mbit (three versus four features), it also has half as many wordlines,
effectively reducing the wordline pitch. This relaxed wordline pitch makes
layout of the wordline drivers and the address decode tree much easier. In
fact, both odd and even wordlines can be driven from the same row decoder
block, thus eliminating half of the row decoder strips in a given array block.
This is an important distinction, as the tight wordline pitch for folded digit­
line designs necessitates separate odd and even row decode strips.

3.2.3 Bilevel Digitline Array Architecture

The bilevel digitline array architecture depicted in Figure 3.12 is a
potential architecture for tomorrow's large-scale DRAM designs. The
bilevel architecture is constructed with multiple array cores separated by
strips of sense amplifiers and either row decode blocks or wordline stitching
regions. Wordline stitching requires a fbur-metal process, while row decode
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 89

blocks can be implemented in a three-metal process. The array cores

include 262,144 (225) functionally addressable plaid 6F2 Mbits arranged in
532 rows and 524 bilevel digitline pairs. The 532 rows comprise 512 actual
wordhnes, 4 redundant wordlines, and 16 dummy wordlines. There are also
267 isolation gates in each array due to the use of plaid Mbits. Because they
are accounted for in the wordline pitch, however, they can be ignored. The
524 bilevel digitline pairs comprise 512 actual digitline pairs, 8 redundant
digitline pairs, and 4 dummy digitline pairs. The term digitline pair
describes the array core structure because pairing is a natural product of the
bilevel architecture. Each digitline pair consists of one digitline on Metal 1
and a vertically aligned complementary digitline on Metal2.

Figure 3.11 Plaid 6F2 Mbit array.

(The isolation gates are tied to ground or a negative voltage.
90 Chap. 3 Array Architectures

Figure 3.12 Bilevel digitline array schematic.

Sense amplifier blocks are placed on both sides of each array core. The
sense amplifiers within each block are laid out at half pitch: one sense
amplifier fbr every two Metal 1 digitlines. Each sense amplifier connects
through isolation devices to columns (digitline pairs) from two adjacent
array cores. Similar to the folded digitline architecture, odd columns con­
nect on one side of the array core, and even columns connect on the other
side. Each sense amplifier block is then exclusively connected to either odd
or even columns, never to both.
Unlike a folded digitline architecture that uses a local row decode block
connected to both sides of an array core, the bilevel digitline architecture
uses a local row decode block connected to only one side of each core. As
stated earlier, both odd and even rows can be driven from the same local
row decoder block with the relaxed wordline pitch. Because of this feature,
the bilevel digitline architecture is more efficient than alternative architec­
tures. A four-metal DRAM process allows local row decodes to be replaced
by either stitch regions or local wordline drivers. Either approach could sub­
stantially reduce die size. The array core also includes the three twist
regions necessary fbr the bilevel digitline architecture. The twist region is
larger than that used in the folded digitline architecture, owing to the com­
plexity of twisting digitlines vertically. The twist regions again constitute a
break in the array structure, making it necessary to include dummy word­
lines.
As with the open digitline and folded digitline architectures, the bilevel
digitline length is limited by power dissipation and a minimum cell-to-digit-
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 91

line capacitance ratio. In the 256-Mbit generation, the digitlines are again
restricted from connecting to more than 256 Mbits (128 Mbit pairs). The
analysis to arrive at this quantity is the same as that fbr the open digitline
architecture, except that the overall digitline capacitance is higher. The
bilevel digitline runs over twice as many cells as the open digitline with the
digitlirie running in equal lengths in both Metal2 and Metall. The capaci­
tance added by the Metal2 component is small compared to the already
present Metall component because Metal2 does not connect to Mbit transis­
tors. Overall, the digitline capacitance increases by about 25% compared to
an open digitline. The power dissipated during a Read or Refresh operation
is proportional to the digitline capacitance (Cd), the supply (internal) volt­
age (Vcc), the external voltage (Vccx), the number of active columns (N)t
and the Refresh period (P). It is given as

p %(分 + 4))

(3.12)
On a 256-Mbit DRAM in 8k Refresh, there are 32,768 (215) active col­
umns during each Read, Write, or Refresh operation. Active array current
and power dissipation fbr a 256-Mbit DRAM are given in Table 3.11 for a
90 ns Refresh period (-5 timing) at various digitline lengths. The budget fbr
active array current is limited to 200mA for this 256-Mbit design. Tb meet
this budget, the digitline cannot exceed a length of 256 Mbits.
Wordline length is again limited by the maximum allowable (RC) time
constant of the wordline. The calculation for bilevel digitline is identical to
that performed fbr open digitline due to the similarity of array core design.
These results are given in Table 3.3. Accordingly, if the wordline time con­
stant is to remain under the required 4-nanosecond limit, the wordline
length cannot exceed 512 Mbits (512 bilevel digitline pairs).

Table 3.11 Active current and power versus bilevel digitline length.

Digitline Length Digitline Power

Active Current
(Mbits) Capacitance (fF) Dissipation (mW)

128 128 75 249

256 256 151 498

512 513 301 994

A layout of various bilevel elements was generated to obtain reason­

able estimates of pitch cell size. With these size estimates, overall dimen­
92 Chap. 3 Array Architectures

sions for a 32・Mbit array block could be calculated. The diagram for a 32-
Mbit array block using the bilevel digitline architecture is shown in Figure
3.13. This block requires a total of one hundred twenty-eight 256kbit array
cores. The 128 array cores are ananged in 16 rows and 8 columns. Each 4-
Mbit vertical section consists of 512 wordlines and 8,192 bilevel digitline
pairs (8,192 columns). Eight 4-Mbit strips are required to form the complete
32-Mbit block. Sense amplifier blocks are positioned vertically between
each 4-Mbit section.
Row decode strips are positioned horizontally between every anay
core. Only eight row decode strips are needed for the sixteen array cores, for
each row decode contains wordline drivers for both odd and even rows. The
32-Mbit array block shown in Figure 3.13 includes pitch cell layout esti­
mates. Overall size for the 32-Mbit block is found by summing the dimen­
sions fbr each component
As before,

Heights = (%• HrdeQ +( A . PDl) microns

(3.13)
S 8 u一e

6
5
z
lnE

cd

Figure 3J3 Bilevel digitline architecture 32-Mbit array block.

Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 93

where TR is the number of bilevel row decoders, H^ec is the height of each
decoder, TDL is the number of bilevel digitline pairs including redundant and
dummy, and PDL is the digitline pitch. Also,

Height^! = (TSA + twist, microns

(3.14)
where TSA is the number of sense amplifier strips, WAMP is the width of the
sense amplifiers, TWL is the total number of wordlines including redundant
and dummy, PWL6 is the wordline pitch for the plaid 6F2 Mbit, Ttwist is the
total number of twist regions, and is the width of the twist regions.
Table 3.12 shows the calculated results for the bilevel 32-Mbit block
shown in Figure 3.13. A three-metal process is assumed in these calcula­
tions because it requires the local row decoders. Array efficiency for the
bilevel digitline 32-Mbit array block, which yields 63.1% for this design
example, is given as
. (100-2 755 PDr2-P^6)
Efficiency = ----------- - ----- -------------percent

(3.15)
With Metal4 added to the bilevel DRAM process, the local row decoder
scheme can be replaced by a global or hierarchical row decoder scheme.
The addition of a fourth metal to the DRAM process places even greater
demands on process engineers. Regardless, an analysis of 32-Mbit array
block size was performed assuming the availability of Metal4. The results
of the analysis are shown in Tables 3.13 and 3.14 for the global and hierar­
chical row decode schemes. Array efficiency for the 32-Mbit memory block
using global and hierarchical row decoding is 74.5% and 72.5%, respec­
tively.

3.2.4 Architectural Comparison

Although a straight comparison of DRAM architectures might appear
simple, in fact it is very complicated. Profit remains the critical test of archi­
tectural efficiency and is the true basis fbr comparison. This in turn requires
accurate yield and cost estimates for each alternative. Without these esti­
mates and a thorough understanding of process capabilities, conclusions are
elusive and the exercise is academic. The data necessary to perform the
analysis and render a decision also varies from manufacturer to manufac­
turer. Accordingly, a conclusive comparison of the various array architec­
tures is beyond the scope of this text. Rather, the architectures are compared
in light of the available data. To better facilitate a comparison, the 32-Mbit
94 Chap. 3 Array Architectures

array block size data from Sections 3.1 and 3.2 is summarized in Table 3.15
fbr the open digitline, folded digitline, and bilevel digitline architectures.

Table 3.12 Bilevel digitline (local row decode): 32-Mbit size

calculations.

Description Parameter Size

Number of sense Tsa 9

amplifier strips

Width of sense Wamp 65 gm

amplifiers

Number of local T'ldec 8

decode strips

Height of local Hldec 149 gm

decode strips

Number of Tdl 8,352

digitlines

Number of Twl 4,256

wordlines

Number of twist Wrw/sn 9 pm

regions

Width of twist 力TWIST 9 gm

regions

Height of 32-Mbit HeightM 6,203 gm

block

Width of 32-Mbit Width32 4,632 Rin

block

Area of 32-Mbit Area32 28,732,296 um2

block

From Table 3.15, it seems that overall die size (32-Mbit area) is a better
metric for comparison than array efficiency. For instance, the three-metal
folded digitline design using hierarchical row decodes has an area of
34,089,440mm2 and an efficiency of 70.9%. The three-metal bilevel digit­
line design with local row decodes has an efficiency of only 63.1% but an
overall area of 28,732,296 mm2. Array efficiency for the folded digitline is
higher. This is misleading, however, because the folded digitline yields a die
that is 18.6% larger for the same number of conductors.
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 95

Table 3.13 Bilevel digitline (global decode): 32-Mbit size

calculations.

Description Parameter Size

Number of sense Tsa 9

amplifier strips

Width of sense 畋MP 65呷

amplifiers

Number of global Tgdec 1

decode strips

Height of global Hgdec 200 gm

decode strips

Number of stitch Nst 4

regions

Height of stitch Hst 10 gm

regions

Number of T'dl 8,352

digitlines

Number of Twl 4,256

wordlines

Number of twist Ttwist 24

regions

Width of twist TWIST 9 pm

regions

Height of 32-Mbit Height32 5,251 pm

block

Width of 32-Mbit WidMl 4,632 pm

block

Area of 32-Mbit Areail 24,322,632 "m2

block

Table 3.15 also illustrates that the bilevel digitline architecture always
yields the smallest die area, regardless of the configuration. The smallest
folded digitline design at 32,654,160mm2 and the smallest open digitline
design at 29,944,350mm2 are still larger than the largest bilevel digitline
design at 28,732,296mm2. It is also apparent that both the bilevel and open
digitline architectures need at least three conductors in their construction.
96 Chap. 3 Array Architectures

The folded digitline architecture still has a viable design option using only
two conductors. The penalty of using two conductors is a much larger die
size-a full 41% than the three-metal bilevel digitline design.

Table 3.141 Bilevel digitline (hierarchical row decode):

32-Mbit size calculations.

Description Parameter Size

Number of Tsa 9
sense amplifier
strips

Width of sense
amplifiers
* 65 gm

Number of glo­ T'gdec 1

bal decode strips

Height of glo­ Hgdec 190 gm

bal decode strips

Number of hier Nhdec 4

decode strips

Height of hier Hhdec 48 gm

decode strips

Number of Tdl 8,352

digitlines

Number of T\vl 4,256

wordlines

Number of twist Ttwist 24

regions

Width of twist %TWIST 9 pm

regions

Height of 32- Height^ 5,393 pm

Mbit block

Width of 32- Widths 4,632 gm

Mbit block

Area of32-Mbit Area32 24.930,376 gm2

block
Sec. 3.2 Design Examples: Advanced Bilevel DRAM Architecture 97

Table 3.15 32-Mbit size calculations summary.

32-Mbit Area Efficiency

Architecture Row Decode Metals
3m2) (%)

Open digit Global 3 29,944,350 60.5

Open digit Hier 3 32,429,565 55.9

Folded digit Local 2 40,606,720 59.5

Folded digit Global 3 32,654,160 74.0

Folded digit Hier 3 34,089,440 70.9

Bilevel digit Local 3 28,732,296 63.1

Bilevel digit Global 4 24,322,632 74.5

Bilevel digit Hier 4 24,980,376 72.5

REFERENCES

[1] H. Hidaka, Y. Matsuda, and K. Fujishima, t4A divided/shared bit-line sensing

scheme for ULSI DRAM cores/' IEEE Journal of Solid-State Circuits, vol.
26, pp. 473777, April 1991.
[2] T. Hamada, N. Tanabe, H. Watanabe, K. Takeuchi, N. Kasai, H. Hada,
K. Shibahara, K. Tokashiki, K. Nakajima, S. Hirasawa, E. Ikawa, T. Saeki,
E. Kakehashi, S. Ohya, and T. A. Kunio, “A split-level diagonal bit-line
(SLDB) stacked capacitor cell for 256Mb DRAMs J in 1992IEDM Technical
Digest, pp. 799-802.
[3] M. Inoue, H. Kotani, T. Yamada, H. Yamauchi, A. Fujiwara, J. Matsushima,
H. Akamatsu, M. Fukumoto, M. Kubota, I. Nakao, N. Aoi, G. Fuse, S. Ogawa,
S. Odanaka, A. Ueno, and H. Yamamoto, 4tA 16Mb DRAM with an open bit­
line architecture,in 1988 IEEE ISSCC Digest of Technical Papers,
pp. 246-247.
[4] M. Inoue, T. Yamada, H. Kotani, H. Yamauchi, A. Fujiwara, J. Matsushima,
H. Akamatsu, M. Fukumoto, M. Kubota, I. Nakao, N. Aoi, G. Fuse, S. Ogawa,
S. Odanaka, A. Ueno, and H. Yamamoto, t4A 16-Mbit DRAM with a relaxed
sense-amplifier-pitch open-bit-line architecture,n IEEE Journal of Solid-State
Circuits, vol. 23, pp. 1104-1112, October 1988.
[5] K. Noda, T. Saeki, A. Tsujimoto, T. Murotani, and K. Koyama, UA boosted
dual word-line decoding scheme for 256Mb DRAMs," in 1992 Symposium on
VLSI Circuits Digest of Technical Papers, pp. 112-113.
98 Chap. 3 Array Architectures

[6] D. Takashima, S. Watanabe, H. Nakano, Y. Oowaki, and K. Ohuchi,

uOpen/folded bit-line arrangement for ultra-high-density DRAMs,“ IEEE
Journal ofSolid-State Circuits, vol. 29, pp. 539-542, April 1994.
[7] Y. Oowaki, K. Tsuchida, Y. Watanabe, D. Takashima, M. Ohta, H. Nakano,
S. Watanabe, A. Nitayama, F. Horiguchi, K. Ohuchi, and F. Masuoka, 4*A 33­
ns 64Mb DRAM,“ IEEE Journal of Solid-State Circuits, vol. 26,
pp. 1498-1505, November 1991.
[8] Y. Nakagome, M. Aoki, S. Ikenaga, M. Horiguchi, S. Kimura, Y. Kawamoto,
and K. Itoh, “The impact of data-line interference noise on DRAM scaling,M
IEEE Journal ofSolid-State Circuits, vol. 23, pp. 1120-1127, October 1988.
[9] T. Yoshihara, H. Hidaka, Y. Matsuda, and K. Fujishima, ctA twisted bitline
technique for multi-Mb DRAMs J in 1988 IEEE ISSCC Digest of Technical
Papers, pp. 238-239.
 Chapter
4

The Peripheral Circuitry

In this chapter, we briefly discuss the peripheral circuitry. In particular, we
discuss the column decoder and its implementation. We also cover the
implementation of row and column redundancy.

4.1 COLUMN DECODER ELEMENTS

The column decoder circuits represent the final DRAM elements that pitch
up to the array mbits. Historically, column decode circuits were simple and
straightforward: static logic gates were generally used for both the decode
tree elements and the driver output. Static logic was used primarily because
of the nature of column addressing in fast page mode (FPM) and extended
data out (EDO) devices. Unlike row addressing, which occurred once per
RAS* cycle, column addressing could occur multiple times per RAS* cycle,
with each column held open until a subsequent column address appeared.
The transition interval from one column to the next had to be minimized,
allowing just enough time to turn OFF the previous column, turn ON the
new column, and equilibrate the necessary data path circuits. Furthermore,
because column address transitions were unpredictable and asynchronous, a
column decode logic that was somewhat forgiving of random address
changes was required-hence the use of static logic gates.
With the advent of synchronous DRAMs (SDRAMs) and high-speed,
packet-based DRAM technology, the application of column, or row,
addresses became synchronized to a clock. More importantly, column
addressing and column timing became more predictable, allowing design
engineers to use pipelining techniques along with dynamic logic gates in
constructing column decode elements.
Column redundancy adds complexity to the column decoder because
the redundancy operation in FPM and EDO DRAMs requires the decode

99
100 Chap. 4 The Peripheral Circuitry

circuit to terminate column transitions, prior to completion. In this way,

redundant column elements can replace normal column elements. Gener­
ally, the addressed column select is allowed to fire normally. If a redundant
column match occurs for this address, the normal column select is subse­
quently turned OFF; the redundant column select is fired. The redundant
match is timed to disable the addressed column select before enabling the
I/O devices in the sense amplifier.
The fire-and-cancel operation used on the FPM and EDO column
decoders is best achieved with static logic gates. In packet-based and syn­
chronous DRAMs, column select firing can be synchronized to the clock.
Synchronous operation, however, does not favor a fire-and-cancel mode,
preferring instead that the redundant match be determined prior to firing
either the addressed or redundant column select. This match is easily
achieved in a pipeline architecture because the redundancy match analysis
can be performed upstream in the address pipeline before presenting the
address to the column decode logic.
A typical FPM- or EDO-type column decoder realized with static logic
gates is shown schematically in Figure 4.1. The address tree is composed of
combinations ofNAND or NOR gates. In this figure, the address signals are
active HIGH, so the tree begins with two-input NAND gates. Using prede­
coded address lines is again preferred. Predecoded address lines both sim­
plify and reduce the decoder logic because a single input can represent two
or more address terms. In the circuit shown in Figure 4.1, the four input sig­
nals C423, C445, C467, and CAS represent seven address terms, permitting
1 of 128 decoding. Timing of the column selection is controlled by a signal
called column decode enable (CDE), which is usually combined with an
input signal, as shown in Figure 4.2, or as an additional term in the tree.
The final signal shown in the figure is labeled RED. This signal disables
the normal column term and enables redundant column decoders whenever
the column address matches any of the redundant address locations, as
determined by the column-redundant circuitry. A normal column select
begins to turn ON before RED makes a LOW-to-HIGH transition. This fire-
and-cancel operation maximizes the DRAM speed between column
accesses, making column timing all the more critical. An example, which
depicts a page mode column transition, is shown in Figure 4.2. During the
fire-and-cancel transition, I/O equilibration (EQIO) envelops the deselec­
tion of the old column select CSEL<0> and the selection of a new column
select RCSEL<\>. In this example, CSEL<\> initially begins to turn ON
until RED transitions HIGH, disabling CSEL<\> and enabling redundant
column RCSEL<\>.
Sec. 4.1 Column Decoder Elements 101

Figure 4.2 Column selection timing.

The column decode output driver is a simple CMOS inverter because

the column select signal (CSEL) only needs to drive to Vy On the other
hand, the wordline driver, as we have seen, is rather complex; it needed to
drive to a boosted voltage, VCCP. Another feature of column decoders is that
their pitch is very relaxed relative to the pitch of the sense amplifiers and
row decoders. From our discussion in Section 2.2 concerning I/O transistors
and CSEL lines, we learned that a given CSEL is shared by four to eight I/O
transistors- Therefore, the CSEL pitch is one CSEL fbr every eight to sixteen
digitlines. As a result, the column decoder layout is much less difficult to
implement than either the row decoder or the sense amplifiers.
A second type of column decoder, realized with dynamic P&E logic, is
shown in Figure 4.3. This particular design was first implemented in an
800MB/s packet-based SLDRAM device. The packet nature of SLDRAM
and the extensive use of pipelining supported a column decoder built with
P&E logic. The column address pipeline included redundancy match cir­
cuits upstream from the actual column decoder, so that both the column
102 Chap. 4 The Peripheral Circuitry

address and the corresponding match data could be presented at the same
time. There was no need fbr the fire-and-cancel operation: the match data
was already available.
Therefore, the column decoder fires either the addressed column select
or the redundant column select in synchrony with the clock. The decode tree
is similar to that used fbr the CMOS wordline driver; a pass transistor was
added so that a decoder enable term could be included. This term allows the
tree to disconnect from the latching column select driver while new address
terms flow into the decoder. A latching driver was used in this pipeline
implementation because it held the previously addressed column select
active with the decode tree disconnected. Essentially, the tree would discon­
nect after a column select was fired, and the new address would flow into
the tree in anticipation of the next column select. Concurrently, redundant
match information would flow into the phase term driver along with CA 45
address terms to select the correct phase signal. A redundant match would
then override the normal phase term and enable a redundant phase term.
Operation of this column decoder is shown in Figure 4.4. Once again,
deselection of the old column select CSEL<0> and selection of a new col­
umn select RCSEL<\> are enveloped by EQIO. Column transition timing is
under the control of the column latch signal CLATCH七 This signal shuts
OFF the old column select and enables firing of the new column select.
Concurrent with CLATCH* firing, the decoder is enabled with decoder
enable (DECEN) to reconnect the decode tree to the column select driver.
After the new column select fires, DECEN transitions LOW to once again
isolate the decode tree.

4.2 COLUMN AND ROW REDUNDANCY

Redundancy has been used in DRAM designs since the 256k generation to
improve yield and profitability. In redundancy, spare elements such as rows
and columns are used as logical substitutes fbr defective elements. The sub­
stitution is controlled by a physical encoding scheme. As memory density
and size increase, redundancy continues to gain importance. The early
designs might have used just one form of repairable elements, relying
exclusively on row or column redundancy. Yet as processing complexity
increased and feature size shrank, both types of redundancy-row and col­
umn一became mandatory.
Sec. 4.2 Column and Row Redundancy 103

CA1 CA2 DECEN CA3

I , ■ I I

FL CSEL

Normal CSEL driver

LC>A45*<n>

户I RCSEL

Redundant CSEL driver

LCMAT*<n>

CMAT*<0:3> <0> LCMAT*<Q:3>

<1>

<2>

<3>
RED

CA45*〈0:3： <0> LC45*<0:3>

<1>

Y <2>
I
<3>

CLATCH*
0.5 ns

RESET*

Figure 43 Column decode: P&E logic.

104 Chap. 4 The Peripheral Circuitry

Figure 4.4 Column decode waveforms.

Today various DRAM manufacturers are experimenting with additional

forms of repair, including replacing entire subarrays. The most advanced
type of repair, however, involves using movable saw lines as realized on a
prototype 1Gb design [1]. Essentially, any four good adjacent quadrants
from otherwise bad die locations can be combined into a good die by simply
sawing the die along different scribe lines. Although this idea is far from
reaching production, it illustrates the growing importance of redundancy.

4.2.1 Row Redundancy

The concept of row redundancy involves replacing bad wordlines with
good wordlines. There could be any number of problems on the row to be
repaired, including shorted or open wordlines, wordline-to-digitline shorts,
or bad Mbit transistors and storage capacitors. The row is not physically but
logically replaced. In essence, whenever a row address is strobed into a
DRAM by RAS*, the address is compared to the addresses of known bad
rows. If the address comparison produces a match, then a replacement
wordline is fired in place of the normal (bad) wordline.
The replacement wordline can reside anywhere on the DRAM. Its loca­
tion is not restricted to the array containing the normal wordline, although
its range may be restricted by architectural considerations. In general, the
redundancy is considered local if the redundant wordline and the normal
wordline must always be in the same subarray.
If, however, the redundant wordline can exist in a subarray that does not
contain the normal wordline, the redundancy is considered global. Global
repair generally results in higher yield because the number of rows that can
be repaired in a single subarray is not limited to the number of its redundant
rows. Rather, global repair is limited only by the number of fuse banks,
termed repair elements, that are available to any subarray.
Local row repair was prevalent through the 16-Meg generation, produc­
ing adequate yield fbr minimal cost. Global row repair schemes are becom­
Sec. 4.2 Column and Row Redundancy 105

ing more common fbr 64-Meg or greater generations throughout the

industry. Global repair is especially effective for repairing clustered failures
and offers superior repair solutions on large DRAMs.
Dynamic logic is a traditional favorite among DRAM designers for row
redundant match circuits. Dynamic gates are generally much faster than
static gates and well suited to row redundancy because they are used only
once in an entire RAS* cycle operation. The dynamic logic we are referring
to again is called precharge and evaluate (P&E). Match circuits can take
many forms; a typical row match circuit is shown in Figure 4.5. It consists
of a PRECHARGE transistor Ml, match transistors M2-M5, laser fuses
F1-F4, and static gate II. In addition, the node labeled row PRECHARGE
(RPRE*) is driven by static logic gates. The fuses generally consist of nar­
row polysilicon lines that can be blown or opened with either a precision
laser or a fusing current provided by additional circuits (not shown).
For our example using predecoded addresses, three of the four fuses
shown must be blown in order to program a match address. for instance,
F2-F4 were blown, the circuit would match for RA 12<0> but not for
7M12<1:3>. Prior to RAS* falling, the node labeled RED* is precharged to
Vcc by the signal RPRE*, which is LOW. Assuming that the circuit is
enabled by fuse F5, EVALUATE* will be LOW. After RAS^ falls, the row
addresses eventually propagate into the redundant block. If RA 12<0> fires
HIGH, RED* discharges through M2 to ground, If, however, RA 12<0>
does not fire HIGH, RED* remains at Vcc, indicating that a match did not
occur. A weak latch composed of II and M6 ensures that RED* remains at
VCc and does not discharge due to junction leakage.
This latch can easily be overcome by any of the match transistors. The
signal RED* is combined with static logic gates that have similar RED* sig­
nals derived from the remaining predecoded addresses. If all of the RED*
signals for a redundant element go LOW, then a match has occurred, as
indicated by row match (RMAT*) firing LOW. The signal RMAT* stops the
normal row from firing and selects the appropriate replacement wordline.
The fuse block in Figure 4.5 shows additional fuses F5-F6 fbr enabling and
disabling the fuse bank. Disable fuses are important in the event that the
redundant element fails and the redundant wordline must itself be repaired.
The capability to pretest redundant wordlines is an important element in
most DRAM designs today. For the schematic shown in Figure 4.5, the pre­
test is accomplished through the set of static gates driven by the redundant
test (REDTEST) signal. The input signals labeled TRA 12n, 77M34n,
77^4 56n, and TODDE VEN are programmed uniquely fbr each redundant
element through connections to the appropriate predecoded address line.
106 Chap. 4 The Peripheral Circuitry

REDTEST is HIGH whenever the DRAM is in the redundant row pre­

test mode. If the current row address corresponds to the programmed pretest
address, RMAT* will be forced LOW, and the corresponding redundant
Sec. 4.2 Column and Row Redundancy 107

wordline rather than the normal wordline will be fired. This pretest capabil­
ity permits all of the redundant wordlines to be tested prior to any laser pro­
gramming.
Fuse banks or redundant elements, as shown in Figure 4.5, are physi­
cally associated with specific redundant wordlines in the array. Each ele­
ment can fire only one specific wordline, although generally in multiple
subarrays. The number of subarrays that each element controls depends on
the DRAM's architecture, refresh rate, and redundancy scheme. It is not
uncommon in 16-Meg DRAMs for a redundant row to replace physical
rows in eight separate subarrays at the same time. Obviously, the match cir­
cuits must be fast. Generally, firing of the normal row must be held off until
the match circuits have enough time to evaluate the new row address. As a
result, time wasted during this phase shows up directly on the parfs row
access (tRAC) specification.

4.2.2 Column Redundancy

Column redundancy is the second type of repair available on most
DRAM designs. In Section 4.1, it was stated that column accesses can occur
multiple times per cycle. Each column is held open until a subsequent
column appears. Therefore, column redundancy was generally implemented
with circuits that are very different from those seen in row redundancy. As
shown in Figure 4.6, a typical column fuse block is built from static logic
gates rather than from P&E dynamic gates. P&E logic, though extremely
fast, needs to be precharged prior to each evaluation. The transition from
one column to another must occur within the span of a few nanoseconds.
On FPM and EDO devices, which had unpredictable and asynchronous col­
umn address transitions, there was no guarantee that adequate time existed
for this PRECHARGE to occur. Yet the predictable nature of column
addressing on the newer type of synchronous and packet-based DRAMs
affords today's designers an opportunity to use P&E-type column redun­
dancy circuits. In these cases, the column redundant circuits appear very
similar to P&E-style row redundant circuits.
An example of a column fuse block for an FPM or EDO design is
shown in Figure 4.6. This column fuse block has four sets of column fuse
circuits and additional enable/disable logic. Each column fuse circuit, corre­
sponding to a set of predecoded column addresses, contains compare logic
and two sets of laser fuse/latch circuits. The laser fuse/latch reads the laser
fuse whenever column fuse power (CFP) is enabled, generally on POW-
ERUP and during RAS* cycles. The fuse values are held by the simple
inverter latch circuits composed of 10 and II. Both true and complement
data are fed from the fuse/latch circuit into the comparator logic. The com­
108 Chap. 4 The Peripheral Circuitry

parator logic, which appears somewhat complex, is actually quite simple as

shown in the following Boolean expression where F0 without the bar indi­
cates a blown fuse:

CAM =（布•丽•丙）+ （加•广0 •西）+ （姮•丽・尸1） + （前•尸0•尸1）

（4.1）
The column address match （CAM） signals from all of the predecoded
addresses are combined with static logic gates to create a column match
（CMAT^） signal for the column fuse block. The CMAT* signal, when
active, cancels normal CSEL signals and enables redundant RCSEL signals,
as described in Section 4.1. Each column fuse block is active only when its
corresponding enable fuse is blown. The column fuse block usually con­
tains a disable fuse for the same reason as a row redundant block: to repair a
redundant element. Column redundant pretest is implemented somewhat
differently in Figure 4.6 than row redundant pretest here. In Figure 4.6, the
bottom fuse terminal is not connected directly to ground. Rather, all of the
signals for the entire column fuse block are brought out and programmed
either to ground or to a column pretest signal from the test circuitry.
During standard part operation, the pretest signal is biased to ground,
allowing the fuses to be read normally. However, during column redundant
pretest, this signal is brought to %, which makes the laser fuses appear to
be programmed. The fuse/latch circuits latch the apparent fuse states on the
next RAS* cycle. Subsequent column accesses allow the redundant column
elements to be pretested by addressing their pre-programmed match
addresses.
The method of pretesting just described always uses the match circuits
to select a redundant column. It is a superior method to that described fbr
the row redundant pretest because it tests both the redundant element and its
match circuit Furthermore, as the match circuit is essentially unaltered dur­
ing redundant column pretest, the test is a better measure of the obtainable
DRAM performance when the redundant element is active.
Obviously, the row and column redundant circuits that are described in
this section are only one embodiment of what could be considered a wealth
of possibilities. It seems that all DRAM designs use some alternate form of
redundancy. Other types of fuse elements could be used in place of the laser
fuses that are described. A simple transistor could replace the laser fuses in
either Figure 4.5 or Figure 4.6, its gate being connected to an alternative
fuse element. Furthermore, circuit polarity could be reversed and non-pre-
decoded addressing and other types of logic could be used. The options are
nearly limitless. Figure 4.7 shows a SEM image of a typical set of poly
fuses.
Sec. 4.2 Column and Row Redundancy 109

CFP*・

PRO

PR1
C401
<0,3>
CFP Column address 2 & 3 fusebank
E>
CFP*
PR2 ♦—— FO
PR3 < CAM*
CA23\ A<0:3>
<0:3>

CFP Column address 4 & 5 fuse bank

CFP* CAM*
PR4
PR 5
845. A<0:3> CU4T*
<0:3>

CFP Column address 6 & 7 fusebank

CFP*
PR6 CAM*
PR7
C467* A<0:3>
<0:3>

EFDS

REDTSTC ENABLE
DISRED7

Figure 4.6 Column fuse block.

110 Chap. 4 The Peripheral Circuitry

Figure 4.7 8-Meg x 8-sync DRAM poly fuses.

REFERENCE

[1] T. Sugibayashi, L Naritake, S. Utsugi, K. Shibahara, R. Oikawa, H. Mori,

S. Iwao, T. Murotani, K. Koyama, S, Fukuzawa, T. Itani, K. Kasama,
T. Okuda, S. Ohya, and M. Ogawa, UA 1Gbit DRAM for file applications,M in
Digest ofInternational Solid-State Circuits Conferencey 1995, pp. 254-255.
 Chapter
5

Global Circuitry
and Considerations
In this chapter, we discuss the circuitry and design considerations associ­
ated with the circuitry external to the DRAM memory array and memory
array peripheral circuitry. We call this global circuitry.

5.1 DATA PATH ELEMENTS

The typical DRAM data path is bidirectional, allowing data to be both writ­
ten to and read from specific memory locations. Some of the circuits
involved are truly bidirectional, passing data in fbr Write operations and out
for Read operations. Most of the circuits, however, are unidirectional, oper­
ating on data in only a Read or Write operation. To support both operations,
therefore, unidirectional circuits occur in complementary pairs: one fbr
reading and one fbr writing. Operating the same regardless of the data direc­
tion, sense amplifiers, I/O devices, and data muxes are examples of bidirec­
tional circuits. Write drivers and data amplifiers, such as direct current
sense amplifiers (DCSAs) or data input buffers and data output buffers, are
examples of paired, unidirectional circuits. In this chapter, we explain the
operation and design of each of these elements and then show how the ele­
ments combine to form a DRAM data path. In addition, we discuss address
test compression circuits and data test compression circuits and how they
affect the overall data path design.

5.1.1 Data Input Buffer

The first element of any DRAM data path is the data input buffer. An
early example is shown in Figure 5.1, the input buffer consists of both

111
112 Chap. 5 Global Circuitry and Considerations

NMOS and PMOS transistors, basically forming a pair of cascaded invert­

ers. The first inverter stage has ENABLE transistors MI and M2, allowing
the buffer to be powered down during inactive periods. The transistors are
carefully sized to provide high-speed operation and specific input trip
points. In first- and second-generation DRAMs, the high-input trip point
匕h is set to 2.0 V for low-voltage TTL (L VTTLJ-compatible DRAMs, while
the \ow-input trip point VfL is set to 0.8 V.
Designing an input buffer to meet specified input trip points generally
requires a flexible design with a variety of transistors that can be added or
deleted with edits to the metal mask. This is apparent in Figure 5.1 by the
presence of switches in the schematic; each switch represents a particular
metal option available in the design. Because of variations in temperature,
device, and process, the final input buffer design is determined with actual
silicon, not simulations. For a DRAM that is 8 bits wide (x8), there will be
eight input buffers, each driving into one or more Write driver circuits
through a signal labeled DW<v> (Data Write where n corresponds to the
specific data bit 0-7).

Figure 5.1 Data input buffer.

As supply voltage levels dropped with process scaling, the basic

inverterbased input buffer shown in Figure 5.1 found less use in DRAM.
The required noise margins and speed of the interconnecting bus between
the memory controller and the DRAM were difficult to meet. Other topolo­
gies are used. One high-speed bus topology, called stub series terminated
logic (SSTL), is shown in Figure 5.2 [1]. Tightly controlled transmission
line impedances and series resistors transmit high-speed signals with little
distortion. Figure 5.2a shows the bus fbr clocks, command signals, and
addresses. Figure 5.2b shows the bidirectional bus fbr transmitting data to
Sec. 5.1 Data Path Elements 113

and from the DRAM controller. In either circuit, Vpp and VREF are set to
Vcc/2.

Figure 5.2 Stub series terminated logic (SSTL).

From this topology, we can see that a differential input buffer should be
used: an inverter won't work. Some examples of fully differential input
buffers are seen in Figure 5.3, Figure 5.4, and Figure 5,5 [1].
Figure 5.3 is simply a CMOS differential amplifier with an inverter out­
put to generate valid CMOS logic levels. Common-mode noise on the difF­
amp inputs is, ideally, rejected while amplifying the difference between the
input signal and the reference signal. The difFamp input common-mode
range, say a hundred mV, sets the minimum input signal amplitude (cen­
tered around ”尸)required to cause the output to change stages. The speed
of this configuration is limited by the difTamp biasing current. Using a large
current will increase input receiver speed and, at the same time, decrease
amplifier gain and reduce the di住amp's input common-mode range.
The input buffer of Figure 5.3 requires an external biasing circuit. The
circuit of Figure 5.4 is self-biasing. This circuit is constructed by connecting
the gate of the NMOS device to the gates of the PMOS load transistors. This
circuit is simple and, because of the adjustable biasing connection, poten­
tially very fast. The output inverter is needed to ensure that valid output
logic levels are generated.
Both of the circuits in Figures 5.3 and 5.4 suffer from duty-cycle distor­
tion at high speeds. The PULLUP delay doesn't match the PULLDOWN
114 Chap. 5 Global Circuitry and Considerations

delay. Duty-cycle distortion is more of a factor in input buffer design in syn­

chronous DRAMs where clocking is performed on both the rising and fall­
ing edges of the system clock. The differential self-biased input receiver of
Figure 5.5 is used to increase the allowable input signal range for wide-
swing operation.

Figure 5.3 Differential amplifier-based input receiver.

Figure 5.4 An (n-flavor) input buffer fbr high-speed digital design.

Sec. 5.1 Data Path Elements 115

Figure 5.5 A rail-to-rail input buffer.

5.1.2 Data Write Muxes

Data muxes are often used to extend the versatility of a design.
Although some DRAM designs connect the input buffer directly to the
Write driver circuits, most architectures place a block of Data Write muxes
between the input buffers and the Write drivers. The muxes allow a given
DRAM design to support multiple configurations, such as x4, x8, and xl6
I/O. A typical schematic for these muxes is shown in Figure 56 As shown
in this figure, the muxes are programmed according to the bond option con­
trol signals labeled 0PTX4, 0PTX8, and OPTXX6. For xl6 operation, each
input buffer is muxed to only one set of DW lines. For x8 operation, each
input buffer is muxed to two sets of DW lines, essentially doubling the
quantity of Mbits available to each input buffer. For x4 operation, each input
buffer is muxed to four sets of DW lines, again doubling the number of
Mbits available to the remaining four operable input buffers.
Essentially, as the quantity of input buffers is reduced, the amount of
column address space for the remaining buffers is increased. This concept is
easy to understand as it relates to a 16Mb DRAM. As a xl6 part, this
DRAM has 1 Mbit per data pin; as a x8 part, 2 Mbits per data pin; and as a
116 Chap. 5 Global Circuitry and Considerations

x4 part, 4 Mbits per data pin. For each configuration, the number of array
sections available to an input buffer must change. By using, Data Write
maxes that permit a given input buffer to drive as few or as many Write
driver circuits as required, design flexibility is easily accommodated.

5.1.3 Write Driver Circuit

The next element in the data path to be considered is the Write driver
circuit. This circuit, as the name implies, writes data from the input buffers
into specific memory locations. The Write driver, as shown in Figure 5.7,
drives specific I/O lines coming from the Mbit arrays. A given Write driver
is generally connected to only one set of I/O lines, unless multiple sets of
I/O lines are fed by a single Write driver circuit via additional muxes. Using
maxes between the Write driver and the arrays to limit the number of Write
drivers and DCSA circuits is quite common. Regardless, the Write driver
uses a tristate output stage to connect to the I/O lines. Tristate outputs are
necessary because the I/O lines are used for both Read and Write opera­
tions. The Write driver remains in a high-impedance state unless the signal
labeled Write is HIGH and either DW or DW* transitions LOW from the
initial HIGH state, indicating a Write operation. As shown in Figure 5.7, the
Write driver is controlled by specified column addresses, the Write signal,
and DW<n>> The driver transistors are sized large enough to ensure a quick,
efficient Write operation. This is important because the array sense amplifi­
ers usually remain ON during a Write cycle.
The remaining elements of the Write Data path reside in the array and
pitch circuits. As previously discussed in Sections 1.2, 2.1, and 2.2, the
Mbits and sense amplifier block constitute the end of the Write Data path.
The new input data is driven by the Write driver, propagating through the
I/O transistors and into the sense amplifier circuits. After the sense amplifi­
ers are overwritten and the new data is latched, the Write drivers are no
longer needed and can be disabled. Completion of the Write operation into
the Mbits is accomplished by the sense amplifiers, which restore the digit­
lines to full VCc and ground levels. See Sections 1.2 and 2.2 for further dis­
cussion.
Sec. 5.1 Data Path Elements 117

DW<n> DIN*<vn>
x16 x8 _x4
0 0 0
1 1 1
0PTX16 2 2 0
3 3 1
0PTX8 4 4 0
5 5 1
OPTX4 6 6 0
7 7 1
DIN*<0> DW<Q>

DW<1>

DIN*<2> O- B DW<2>

A I
r
DIN*<3> O
B ]
I
A
B 1
1
l r DW<4>
D/A/*<4> O C
L

A
E
D//V*<5> O 1 1
a
D/N*<6> O B
1 ^>o---- ► DW<6>

A
D/A/*<7> O
B
) 1
Figure 5.6 Data Write mux.
118 Chap. 5 Global Circuitry and Considerations

5.1.4 Data Read Path

The Data Read path is similar, yet complementary, to the Data Write
path. It begins, of course, in the anay, as previously discussed in Sections
1.2 and 2.2. After data is read from the Mbit and latched by the sense ampli­
fiers, it propagates through the I/O transistors onto the I/O signal lines and
into a DC sense amplifier (DCS A) or helper flip-flop (HFF). The I/O lines,
prior to the column select (CSEL) firing, are equilibrated and biased to a
voltage approaching Vcc. The actual bias voltage is determined by the I/O
bias circuit, which serves to control the I/O lines through every phase of the
Read and Write cycles. This circuit, as shown in Figure 5.8, consists of a
group of bias and equilibrate transistors that operate in concert with a vari­
ety of control signals. When the DRAM is in an idle state, such as when
RAS* is HIGH, the I/O lines are generally biased to VCc- During a Read
cycle and prior to CSEL, the bias is reduced to approximately one VTH
below VCc-
The actual bias voltage for a Read operation is optimized to ensure
sense amplifier stability and fast sensing by the DCSA or HFF circuits. Bias
is maintained continually throughout a Read cycle to ensure proper DCSA
operation and to speed equilibration between cycles by reducing the range
over which the I/O lines operate. Furthermore, because the DCSA s or HFFs
are very high-gain amplifiers, raikto-rail input signals are not necessary to
drive the outputs to CMOS levels. In fact, it is important that the input lev­
els not exceed the DCSA or HFF common-mode operating range. During a
Write operation, the bias circuits are disabled by the Write signal, permit­
ting the Write drivers to drive rail-to-rail.
Sec. 5.1 Data Path Elements 119

I/O

BIASIO*
BIASIO

Figure 5.8 I/O bias circuit

Operation of the bias circuits is seen in the signal waveforms shown in

Figure 5.9. For the Read-Modify-Write cycle, the I/O lines start at VCc dur­
ing standby; transition to Vcc-VTli at the start of a Read cycle; separate but
remain biased during the Read cycle; drive rail-to-rail during a Write cycle;
recover to Read cycle levels (termed Write Recovery); and equilibrate to
Vcc-Vth in preparation for another Read cycle.

5.1.5 DC Sense Amplifier (DCSA)

The next data path element is the DC sense amplifier (DCSA). This
amplifier, termed Data amplifier or Read amplifier by DRAM manufactur­
ers, is an essential component in modem, high-speed DRAM designs and
takes a variety of forms. In essence, the DCSA is a high-speed, high-gain
differential amplifier for amplifying very small Read signals appearing on
the I/O lines into full CMOS data signals used at the output data buffer. In
most designs, the I/O lines connected to the sense amplifiers are very capac­
itive.
The array sense amplifiers have very limited drive capability and are
unable to drive these lines quickly. Because the DCSA has very high gain, it
amplifies even the slightest separation of the I/O lines into full CMOS lev­
els, essentially gaining back any delay associated with the I/O lines. Good
DCSA designs can output full rail-to-rail signals with input signals as small
as 15mV This level of performance can only be accomplished through very
careful design and layout. Layout must follow good analog design princi­
ples, with each element a direct copy (no minored layouts) of any like ele­
ments.
120 Chap. 5 Global Circuitry and Considerations

Figure 5.9 I/O bias and operation waveforms.

As illustrated in Figure 5.10, a typical DCSA consists of four differen­

tial pair amplifiers and self-biasing CMOS stages. The differential pairs are
configured as two sets of balanced amplifiers. Generally, the amplifiers are
built with an NMOS differential pair using PMOS active loads and NMOS
current mirrors. Because NMOS has higher mobility, providing for smaller
transistors and lower parasitic loads, NMOS amplifiers usually offer faster
operation than PMOS amplifiers. Furthermore, VTH matching is usually bet­
ter for NMOS, offering a more balanced design. The first set of amplifiers is
fed with I/O and /ZO* signals from the array; the second set, with the output
signals from the first pair, labeled DX and DX*. Bias levels into each stage
are carefully controlled to provide optimum performance.
The outputs from the second stage, labeled DY and DY\ feed into self­
biasing CMOS inverter stages for fast operation. The final output stage is
capable of tristate operation to allow multiple sets of DCSA to drive a given
set of Data Read lines (DR<n> and The entire amplifier is equil­
ibrated prior to operation, including the self-biasing CMOS inverter stages,
by all of the devices connected to the signals labeled EQSA, EQSA *, and
EQSA 2. Equilibration is necessary to ensure that the amplifier is electrically
balanced and properly biased before the input signals are applied. The
amplifier is enabled whenever ENSA * is brought LOW, turning ON the out­
put stage and the current mirror bias circuit, which is connected to the dif­
ferential amplifiers via the signal labeled CM. For a DRAM Read cycle,
operation of this amplifier is depicted in the waveforms shown in Figure
5.11. The bias levels are reduced for each amplifier stage, approaching
Vee? ft>r the final stage.
Sec. 5.1 Data Path Elements 121

Figure 5.10 DC sense amp.

5.1.6 Helper Flip-Flop (HFF)

The DCSA of the last section can require a large layout area. To reduce
the area, a helper flip-flop (HFF) as seen in Figure 5.12, can be used. The
HFF is basically a clocked connection of two inverters as a latch [3]. When
CLK is LOW, the I/O lines are connected to the inputs/outputs of the invert­
ers. The inverters don't see a path to ground because Ml is OFF when CLK
is LOW. When CLK transitions HIGH, the outputs of the HFF amplify, in
effect, the inputs into full logic levels.

Figure 5.11 DCSA operation waveforms.

122 Chap. 5 Global Circuitry and Considerations

Figure 5.12 A helper flip-flop.

For example, if I/O = 1.25 V and UO* = 1.23 V then I/O becomes VCCl
and I/O* goes to zero when CLK transitions HIGH. Using positive feedback
makes the HFF sensitive and fast. Note that HFFs can be used at several
locations on the I/O lines due to the small size of the circuit.

5.1.7 Data Read Muxes

The Read Data path proceeds from the DCSA block to the output buff­
ers. The connection between these elements can either be direct or through
Data Read muxes. Similar to Data Write muxes, Data Read muxes are com­
monly used to accommodate multiple-part configurations with a single
design. An example of this is shown in Figure 5.13. This schematic of a
Data Read mux block is similar to that found in Figure 5.6 for the Data
Write mux block. For xl6 operation, each output buffer has access to only
one Data Read line pair (DR<n> and DRt<ri>). For x8 operation, the eight
output buffers each have two pairs of DR<ti> lines available, doubling the
quantity of Mbits accessible by each output. Similarly, for x4 operation, the
four output buffers have four pairs of DR<n> lines available, again dou­
bling the quantity of Mbits available for each output. For those configura­
tions with multiple pairs available, address lines control which D/?<n> pair
is connected to an output buffer.
Sec. 5.1 Data Path Elements 123

o LDQ<n> Read data

DR<6> o
DR<4> o x16 x8 x4
DR<2> o 0 0 0,2 024,6
DR<0>
1 1 1.3 1,3,57
2 2
X4CTR<0:3> O 3 3
4 4
5 5
DR<7> o 6 6 4,6
o 7
DR<5> o 7 5.7
DR<3> o
DR<1>

OP7X16 6

DR<2> Q LDQ<2>

DR<3> O LDQ<3>

DR<4> a LDQ<4>

DR<5> O LDQ<5>

DR<6> a LDQ<6>
DR<4> Q

XBCTRL<0:1>0-
0PTX4 a

DR<7> o LDQ<7>
DR<5> o

Figure 5.13 Data Read mux.

124 Chap. 5 Global Circuitry and Considerations

5.1.8 Output Buffer Circuit

The final element in our Read Data path is the output buffer circuit. It
consists of an output latch and an output driver circuit. A schematic for an
output buffer circuit is shown in . The output driver on the right side of uses
three NMOS transistors to drive the output pad to either VCcx or ground.
Vccx is the external supply voltage to the DRAM, which may or may not be
the same as Vcc. depending on whether or not the part is internally regu­
lated. Output drivers using only NMOS transistors are common because
they offer better latch-up immunity and ESD protection than CMOS output
drivers. Nonetheless, PMOS transistors are still used in CMOS output driv­
ers by several DRAM manufacturers, primarily because they are much eas­
ier to drive than full NMOS stages. CMOS outputs are also more prevalent
on high-speed synchronous and double-data rate (DDR) designs because
they operate at high data rates with less duty-cycle degradation.

Figure 5.14 Output buffer;

In , two NMOS transistors are placed in series with Vccx to reduce sub­
strate injection currents. Substrate injection currents result from impact ion­
ization, occurring most commonly when high drain-to-source and high
gate-to-source voltages exist concurrently. These conditions usually occur
when an output driver is firing to Vccx，especially for high-capacitance
loads, which slow the output transition. Two transistors in series reduce this
effect by lowering the voltages across any single device. The output stage is
tristated whenever both signals PULLUP and PULLDOWN are at
ground.
The signal PULLDOWN is driven by a simple CMOS inverter, whereas
PULLUP is driven by a complex circuit that includes voltage charge pumps.
The pumps generate a voltage to drive PULLUP higher than one VTH above
Vqcx- This is necessary to ensure that the series output transistors drive the
Sec. 5.1 Data Path Elements 125

pad to Vccx. The output driver is enabled by the signal labeled OE. Once
enabled, it remains tristated until either DQ or DQ* fires LOW. If DR fires
LOW, PULLDOWN fires HIGH, driving the pad to ground through M3. If
DR* fires LOW, PULLUP fires HIGH, driving the pad to Vccx through Ml
and M2.
The output latch circuit shown in controls the output driver operation.
As the name implies, it contains a latch to hold the output data state. The
latch frees the DCSA or HFF and other circuits upstream to get subsequent
data fbr the output. It is capable of storing not only one and zero states, but
also a high-impedance state (tristate). It offers transparent operation to
allow data to quickly propagate to the output driver. The input to this latch is
connected to the DR<n> signals coming from either the DCSAs or Data
Read maxes. Output latch circuits appear in a variety of forms, each serving
the needs of a specific application or architecture. The data path may con­
tain additional latches or circuits in support of special modes such as burst
operation.

5.1.9 Test Modes

Address compression and data compression are two special test modes
that are usually supported by the data path design. Test modes are included
in a DRAM design to extend test capabilities or speed component testing or
to subject a part to conditions that are not seen during normal operation.
Compression test modes yield shorter test times by allowing data from mul­
tiple array locations to be tested and compressed on-chip, thereby reducing
the effective memory size by a factor of 128 or more in some cases. Address
compression, usually on the order of 4x to 32x, is accomplished by inter­
nally treating certain address bits as "don't care“ addresses.
The data from all of the "don't care“ address locations, which corre­
spond to specific data input/output pads (DQ pin，are compared using spe­
cial match circuits. Match circuits are usually realized with NAND and
NOR logic gates or through P&E-type drivers on the differential DR<n>
buses. The match circuits determine if the data from each address location is
the same, reporting the result on the respective DQ pin as a match or a fail.
The data path must be designed to support the desired level of address com­
pression. This may necessitate more DCSA circuits, logic, and pathways
than are necessary fbr normal operation.
The second form of test compression is data compression: combining
data at the output drivers. Data compression usually reduces the number of
DQ pins to four. This compression reduces the number of tester pins
required for each part and increases the throughput by allowing additional
parts to be tested in parallel. In this way, xl6 parts accommodate 4x data
126 Chap. 5 Global Circuitry and Considerations

compression, and x8 parts accommodate 2x data compression. The cost of

any additional circuitry to implement address and data compression must be
balanced against the benefits derived from test time reduction. It is also
important that operation in test mode correlate 100% with operation in non­
test mode. Correlation can be difficult to achieve, however, because addi­
tional circuitry must be activated during compression, modifying noise and
power characteristics on the die.

5.2 ADDRESS PATH ELEMENTS

DRAMs have used multiplexed addresses since the 4kb generation. Multi­
plexing is possible because DRAM operation is sequential: column opera­
tions follow row operations. Obviously, the column address is not needed
until the sense amplifiers have latched, which cannot occur until some time
after the wordline has fired. DRAMs operate at higher current levels with
multiplexed addressing because an entire page (row address) must be
opened with each row access. This disadvantage is overcome by the lower
packaging cost associated with multiplexed addresses. In addition, owing to
the presence of the column address strobe (CAS*), column operation is
independent of row operation, enabling a page to remain open fbr multiple,
high-speed column accesses. This page mode type of operation improves
system performance because column access time is much shorter than row
access time. Page mode operation appears in more advanced forms, such as
EDO and synchronous burst mode, providing even better system perfor­
mance through a reduction in effective column access time.
The address path for a DRAM can be broken into two parts: the row
address path and the column address path. The design of each path is dic­
tated by a unique set of requirements. The address path, unlike the data
path, is unidirectional, with address information flowing only into the
DRAM. The address path must achieve a high level of performance with
minimal power and die area just like any other aspect of DRAM design.
Both paths are designed to minimize propagation delay and maximize
DRAM performance. In this chapter, we discuss various elements of the
row and column address paths.

5.2.1 Row Address Path

The row address path encompasses all of the circuits from the address
input pad to the wordline driver. These circuits generally include the
address buffer, CAS* before RAS* (CBR) counter, predecode logic, array
buffers, redundancy logic, phase drivers, and row decoder blocks. Row
decoder blocks are discussed in Section 2.3, while redundancy is addressed
in Section 4.2. We will now focus on the remaining elements of the row
Sec. 5.2 Address Path Elements 127

address path, namely, the row address buffer, CBR counter, predecode logic,
array buffers, and phase drivers.

5.2.2 Row Address Buffer

The row address buffer, as shown schematically in Figure 5.15, consists
of a standard input buffer and the additional circuits necessary to implement
the functions required for the row address path. The address input buffer
(inpBufi is the same as that used for the data input buffer (see Figure 5.1)
and must meet the same criteria for VIH and VIL as the data input buffer. The
row address buffer includes an inverter latch circuit, as shown in Figure
5.15. This latch, consisting of two inverters and any number of input maxes
(two in this case), latches the row address after RAS^ falls.
The input buffer drives through a mux, which is controlled by a signal
called row address latch (RAL). Whenever RAL is LOW, the mux is
enabled. The feedback inverter has low drive capability, allowing the latch
to be overwritten by either the address input buffer or the CBR counter,
depending on which mux is enabled. The latch circuit drives into a pair of
NAND gates, forcing both 7M<n> and RA *<n> to logic LOW states when­
ever the row address buffer is disabled because RAEN is LOW.

Figure 5.15 Row address buffer.

5.2.3 CBR Counter

As illustrated in Figure 5.15, the CBR (CAS^ before RAS*) counter con­
sists of a single inverter and a pair of inverter latches coupled through a pair
of complementary muxes to form a one-bit counter. For every HIGH-to-
LOW transition of C£K*, the register output at Q toggles. All of the CBR
128 Chap. 5 Global Circuitry and Considerations

counters from each row address buffer are cascaded together to form a CBR
ripple counter. The Q output of one stage feeds the CLK* input of a subse­
quent stage. The first register in the counter is clocked whenever RAS* falls
while in a CBR Refresh mode. By cycling through all possible row address
combinations in a minimum of clocks, the CBR ripple counter provides a
simple means of internally generating Refresh addresses. The CBR counter
drives through a mux into the inverter latch of the row address buffer. This
mux is enabled whenever CBR address latch (CBRAL) is LOW. Note that
the signals RAL and CBRAL are mutually exclusive in that they cannot be
LOW at the same time. For each and every DRAM design, the row address
buffer and CBR counter designs take on various forms. Logic may be
inverted, counters may be more or less complex, and maxes may be
replaced with static gates. Whatever the differences, however, the function
of the input buffer and its CBR counter remains essentially the same.

5.2.4 Predecode Logic

As discussed in earlier sections of this book, using predecoded address­
ing internal to a DRAM has many advantages: lower power, higher effi­
ciency, and a simplified layout. An example of predecode circuits fbr the
row address path is shown in Figure 5.16. This schematic consists of seven
sets of predecoders. The first set is used for even and odd row selection and
consists only of cascaded inverter stages. RA<Q> is not combined with
other addresses due to the nature of row decoding. In our example, we
assume that odd and even will be combined with the predecoded row
addresses at a later point in the design, such as at the array interfaces. The
next set of predecoders is fbr addresses RA<\> and RA <2>, which together
form RA 12<0:3>. Predecoding is accomplished through a set of two-input
NAND gates and inverter buffers, as shown. The remaining addresses are
identically decoded except fbr 7t4<12>. As illustrated at the bottom of the
schematic, RA<12> and RA*<12> are taken through a NOR gate circuit.
This circuit forces both RA<12> lines to a HIGH state as they enter the
decoder, whenever the DRAM is configured fbr 4k Refresh. This process
essentially makes 7t4<12> a "don't care" address in the predecode circuit,
forcing twice as many wordlines to fire at a time.

5.2.5 Refresh Rate

Normally, when Refresh rates change fbr a DRAM, a higher order
address is treated as a "don't care“ address, thereby decreasing the row
address space but increasing the column address space. For example, a
16Mb DRAM bonded as a 4Mb x4 part could be configured in several
Refresh rates: Ik, 2k, and 4k. Table 5.1 shows how row and column
Sec. 5.2 Address Path Elements 129

addressing is related to these Refresh rates fbr the 16Mb example. In this
example, the 2k Refresh rate would be more popular because it has an equal
number of row and column addresses or square addressing.
Refresh rate is also determined by backward compatibility, especially in
personal computer designs. Because a 4Mb DRAM has less memory space
than a 16Mb DRAM, the 4Mb DRAM should naturally have fewer address
pins. To sell 16Mb DRAMs into personal computers that are designed for
4Mb DRAMs, the 16Mb part must be configured with no more address pins
than the 4Mb part. If the 4Mb part has eleven address pins, then the 16Mb
part should have only eleven address pins, hence 2k Refresh. To trim cost,
most PC designs keep the number of DRAM address pins to a minimum.
Although this practice holds cost down, it also limits expandability and
makes conversion to newer DRAM generations more complicated owing to
resultant backward compatibility issues.

Table 5.1 Refresh rate versus row and column addresses.

Refresh Row Column

Rows Columns
rate addresses addresses

4K 4,096 1,024 12 10

2K 2,048 2,048 11 H

IK 1,024 4,096 10 12
130 Chap. 5 Global Circuitry and Considerations

RA*<Q> o-------------------------------------
A— --------- ►Even

RA <0> o-------------------------------------
0- --------- ►Odd

NAND Row
predecc►de address
DA <O> 4— — — ^^rivers
r-1A — ------------ aN412<3>

n --------- ARA12<2>
DA 〜_____— —
-D>
RA *<2> 〜 — --------- ►RA12<1>

L> ----------- »加12<0>

5 —
— _________

RA <4> 。，— 一, ， 一 NAND Row

-------- ► RA34<3>
R>4*<4> o------------------------------------ -------- ► RA34<2>
RA <3> o 一 ， 一... ，一 predecode address --------- ►RA34<1>
RA*<3> o------------------------------------ — drivers -------- A RA34<0>

RA <6> —， ・一 , ，， ， —— — NAND Row

-------- A RA56<3>
R4*<6> o------------------------------------ predecode address
-------- ► RA56<2>
RA <5> o — 一 ■ 一 --------- A RA56<1>
加 ---- --------------- drivers
-------- ► RA56<Q>

Figure 5.16 Row address predecode circuits.

5.2.6 Array Buffers

The next elements to be discussed in the row address path are anay
buffers and phase drivers. The array buffers drive the predecoded address
signals into the row decoder blocks. In general, the buffers are no more than
cascaded inverters, but in some cases they include static logic gates or level
translators, depending on row decoder requirements. Additional logic gates
Sec. 5.2 Address Path Elements 131

could be included for combining the addresses with enabling signals from
the control logic or for making odd/even row selection by combining the
addresses with the odd and even address signals. Regardless, the resulting
signals ultimately drive the decode trees, making speed an important issue.
Buffer size and routing resistance, therefore, become important design
parameters in high-speed designs because the wordline cannot be fired until
the address tree is decoded and ready for the PHASE signal to fire.

5.2.7 Phase Drivers

As the discussion concerning wordline drivers and tree decoders in
Section 2.3 showed, the signal that actually fires the wordline is called
PHASE. Although the signal name may vary from company to company,
the purpose of the signal does not. Essentially, this signal is the final address
term to arrive at the wordline driver. Its timing is carefully determined by
the control logic, PHASE cannot fire until the row addresses are set up in the
decode tree. Normally, the timing of PHASE also includes enough time for
the row redundancy circuits to evaluate the current address. If a redundancy
match is found, the normal row cannot be fired. In most DRAM designs,
this means that the normally decoded PHASE signal will not fire but will
instead be replaced by some form of redundant PHASE signal.
A typical phase decoder/driver is shown in Figure 5.17. Again, like so
many other DRAM circuits, it is composed of standard static logic gates. A
level translator would be included in the design if the wordline driver
required a PHASE signal that could drive to a boosted voltage supply. This
translator could be included with the phase decoder logic or placed in array
gaps to locally drive selected row decoder blocks. Local phase translators
are common on double-metal designs with local row decoder blocks.

5.2.8 Column Address Path

With our examination of row address path elements complete, we can
turn our attention to the column address path. The column address path con­
sists of input buffers, address transition detection circuits, predecode logic,
132 Chap. 5 Global Circuitry and Considerations

redundancy, and column decode circuits. Redundancy and column decode

circuits are addressed in Section 4.2 and Section 4.1, respectively.
A column address buffer schematic, as shown in Figure 5.18, consists
of an input buffer, a latch, and address transition circuits. The input buffer
shown in this figure is again identical to that described in Section 5 J fbr the
data path, so further description is unnecessary. The column address input
buffer, however, is disabled by the signal power column (PCOL*) whenever
RAS* is HIGH and the part is inactive. Normally, the column address bu伍
ers are enabled by PCOL* shortly after RAS* goes LOW. The input buffer
feeds into a NAND latch circuit, which traps the column address whenever
the column address latch (CAL*) fires LOW. At the start of a column cycle,
CAL* is HIGH, making the latch transparent. This transparency permits the
column address to propagate through to the predecode circuits. The latch
output also feeds into an address transition detection (ATD) circuit that is
shown on the right side of Figure 5.18.

--------- [>C------------------------- ^CA

Buffer Latch ATD

Figure 5J8 Column address buffer.

5.2.9 Address Transition Detection

The address transition detection (ATD) circuit is extremely important to
page mode operation in an asynchronous DRAM. An ATD circuit detects
any transition that occurs on a respective address pin. Because it follows the
NAND latch circuit, the CAL* control signal must be HIGH. This signal
makes the latch transparent and thus the ATD functional. The ATD circuit in
Figure 5.18 has symmetrical operation: it can sense either a rising or a fall­
ing edge. The circuit uses a two-input XNOR gate to generate a pulse of
prescribed duration whenever a transition occurs. Pulse duration is dictated
by simple delay elements: Il and Cl or 12 and C2. Both delayed and unde­
layed signals from the latch are fed into the XNOR gate. The ATD output
signals from all of the column addresses, labeled 7DJf*<n>, are routed to
the equilibration driver circuit shown in Figure 5.19. This circuit generates a
set of equilibration signals fbr the DRAM. The first of these signals is equil­
ibrate I/O (EQIO*), which, as the name implies, is used in arrays to force
Sec. 5.2 Address Path Elements 133

equilibration of the I/O lines. As we learned in Section 5.1.4, the I/O lines
need to be equilibrated to Vcc^th prior to a new column being selected by
the CS£L<n> lines. EQIO* is the signal used to accomplish this equilibra­
tion.
The second signal generated by the equilibration driver is called equili­
brate sense amp (EQSA), This signal is generated from address transitions
occurring on all of the column addresses, including the least significant
addresses. The least significant column addresses are not decoded into the
column select lines (CSEL), Rather, they are used to select which set of I/O
lines is connected to the output buffers. As shown in the schematic, EQSA is
activated regardless of which address is changed because the DCSA s must
be equilibrated prior to sensing any new data. EQIO, on the other hand, is
not affected by the least significant addresses because the I/O lines do not
need equilibrating unless the CSEL lines are changed. The equilibration
driver circuit, as shown in Figure 5.19, uses a balanced NAND gate to com­
bine the pulses from each ATD circuit. Balanced logic helps ensure that the
narrow ATD pulses are not distorted as they progress through the circuit.
The column addresses are fed into predecode circuits, which are very
similar to the row address predecoders. One major difference, however, is
that the column addresses are not allowed to propagate through the part until
the wordline has fired. For this reason, the signal Enable column (ECOL) is
gated into the predecode logic as shown in Figure 5.20. ECOL disables the
predecoders whenever it is LOW, forcing the outputs all HIGH in our exam­
ple. Again, the predecode circuits are implemented with simple static logic
gates. The address signals emanating from the predecode circuits are buff­
ered and distributed throughout the die to feed the column decoder logic
blocks. The column decoder elements are described in Section 4.1.

Figure 5.19 Equilibration driver.

134 Chap. 5 Global Circuitry and Considerations

Figure 5.20 Column predecode logic.

Sec. 5.3 Synchronization in DRAMs 135

5.3 SYNCHRONIZATION IN DRAMS[1]

In a typical SDRAM, the relationship when reading data out of the DRAM
between the CLK input and the valid data out time or access time tAC can
vary widely with process shifts, temperature，and operating clock frequency.
Figure 5.21 shows a typical relationship between the SDRAM, CLK input,
and a DQ output. The parameter specification tAC is generally specified as
being less than some value, for example, tAC < 5ns.
As clock frequencies increase, it is desirable to have less uncertainty in
the availability of valid data on the output of the DRAM. Towards this goal,
double data rate (DDR) SDRAMs use a delay-locked loop (DLL) [3] to
drive tAC to zero. (A typical specification for tAC in a DDR SDRAM is
土 0.75ns.) Figure 5.22 shows the block diagram for a DLL used in a DDR
SDRAM. Note that the data I/O in DDR is clocked on both the rising and
falling edges of the input CLK (actually on the output of the delay line) [4].
Also, note that the input CLK should be synchronized with the DQ strobe
(DQS) clock output.
To optimize and stabilize the clock-access and output-hold times in an
SDRAM, an internal register-controlled delay-locked loop (RDLL) has
been used [4, 5, 6]. The RDLL adjusts the time difference between the out­
put (QQs) and input (CLK) clock signals in SDRAM until they are aligned.
Because the RDLL is an all-digital design, it provides robust operation over
all process comers. Another solution to the timing constraints found in
SDRAM was given by using the synchronous mirror delay (SMD) in [7].
Compared to RDLL, the SMD does not lock as tightly, but the time to
acquire lock between the input and output clocks is only two clock cycles.

DRAM input clock signal

DRAM output signal

Figure 5.21 SDRAM CLK input and DQ output.

1 .This material is taken directly from [4].

136 Chap. 5 Global Circuitry and Considerations

Figure 5.22 Block diagram fbr DDR SDRAM DLL.

As DRAM clock speeds continue to increase, the skew becomes the domi­
nating concern, outweighing the RDLL disadvantage of longer time to
acquire lock.
This section describes an RSDLL (register-controlled symmetrical
DLL), which meets the requirements of DDR SDRAM. (Read/Write
accesses occur on both rising and falling edges of the clock.) Here, symmet­
rical means that the delay line used in the DLL has the same delay whether
a HIGH-to-LOW or a LOW-to-HIGH logic signal is propagating along the
line. The data output timing diagram of a DDR SDRAM is shown in Figure
5.23. The RSDLL increases the valid output data window and diminishes
DQS to DQ skew by synchronizing both the rising and falling edges of the
DQS signal with the output data DQ.
Figure 5.22 shows the block diagram of the RSDLL. The replica input
butler dummy delay in the feedback path is used to match the delay of the
input clock buffer. The phase detector (PD) compares the relative timing of
the edges of the input clock signal and the feedback clock signal, which
comes through the delay line and is controlled by the shift register. The out­
puts of the PD, shift-right and shift-left, control the shift register. In the sim­
plest case, one bit of the shift register is HIGH. This single bit selects a
point of entry fbr CLKIn the symmetrical delay line. (More on this later.)
When the rising edge of the input clock is within the rising edges of the out­
put clock and one unit delay of the output clock, both outputs of the PD,
shift-right and shift-left, go LOW and the loop is locked.
Sec. 5.3 Synchronization in DRAMs 137

'csdq Data strobe edge to output

Data edge, undefined area
tDQV Output data valid window
bsv Output data strobe valid window

Figure 5.23 Data timing chart for DDR DRAM.

5.3.1 The Phase Detector

The basic operation of the phase detector (PD) is shown in Figure 5.24.
The resolution of this RSDLL is determined by the size of the unit delay
used in the delay line. The locking range is determined by the number of
delay stages used in the symmetrical delay line. Because the DLL circuit
inserts an optimum delay time between CLKIn and CLKOutf making the
output clock change simultaneously with the next rising edge of the input
clock, the minimum operating frequency to which the RSDLL can lock is
the reciprocal of the product of the number of stages in the symmetrical
delay line with the delay per stage. Adding more delay stages increases the
locking range of the RSDLL at the cost of increased layout area.

5.3.2 The Basic Delay Element

Rather than using an AND gate as the unit-delay stage (NAND + inverter),
as in [5], a NAND-only-based delay element can be used. The implementa­
tion of a three-stage delay line is shown in Figure 5.25. The problem when
using a NAND + inverter as the basic delay element is that the propagation
delay through the unit delay resulting from a HIGH-to-LOW transition is
not equal to the delay of a LOW-to-HIGH transition (tp^L 工，plh)，Further­
more, the delay varies from one run to another. If the skew between tPHL and
tpLH is 50 ps, fbr example, the total skew of the falling edges through ten
stages will be 0.5 ns. Because of this skew, the NAND + inverter delay ele­
ment cannot be used in a DDR DRAM. In our modified symmetrical delay
element, another NAND gate is used instead of an inverter (two
NAND gates per delay stage). This scheme guarantees that tPHL = tPLH inde­
pendent of process variations. While one NAND switches from HIGH to
138 Chap. 5 Global Circuitry and Considerations

LOW, the other switches from LOW to HIGH. An added benefit of the two-
NAND delay element is that two point-of-entry control signals are now
available. The shift register uses both to solve the possible problem caused
by the POWERUP ambiguity in the shift register.

Shift-left

CLK
(to shift
register)

Shift-right

Faster,
shift-right Adjusted

CLKIn
CLKOut
CLKOut +
unit delay ,

Figure 5.24 Phase detector used in RSDLL.

5.3.3 Control of the Shift Register

As shown in Figures 5.25 and 5.26, the input clock is a common input
to every delay stage. The shift register selects a different tap of the delay
line (the point of entry for the input clock signal into the symmetrical delay
line). The complementary outputs of each register cell select the different
tap: Q is connected directly to the input of a delay element, and Q* is con­
nected to the previous stage of input B.
Sec. 5.3 Synchronization in DRAMs 139

Figure 5.25 Symmetrical delay element used in RSDLL.

Figure 5.26 Delay line and shift register fbr RSDLL.

From right to left, the first LOW-to-HIGH transition in the shift register
sets the point of entry into the delay line. The input clock passes through the
tap with a HIGH logic state in the corresponding position of the shift regis­
ter. Because the Q* of this tap is equal to a LOW, it disables the previous
stages; therefore, the previous states of the shift register do not matter
(shown as "don't careJX in Figure 5.25). This control mechanism guaran­
tees that only one path is selected. This scheme also eliminates POWERUP
concerns because the selected tap is simply the first, from the right, LOW-
to-HIGH transition in the register.

5.3.4 Phase Detector Operation

To stabilize the movement in the shift register, after making a decision,
the phase detector waits at least two clock cycles before making another
decision (Figure 5.24). A divide-by-two is included in the phase detector so
that every other decision, resulting from comparing the rising edges of the
140 Chap. 5 Global Circuitry and Considerations

external clock and the feedback clock, is used. This provides enough time
for the shift register to operate and the output waveform to stabilize before
another decision by the PD is implemented. The unwanted side effect of
this delay is an increase in lock time. The shift register is clocked by com­
bining the shift-left and -right signals. The power consumption decreases
when there are no shift-left or -right signals and the loop is locked.
Another concern with the phase-detector design is the design of the flip­
flops (FFs). To minimize the static phase error, very fast FFs should be
used, ideally with zero setup time.
Also, the metastability of the flip-flops becomes a concern as the loop
locks. This, together with possible noise contributions and the need to wait,
as just discussed, before implementing a shift-right or -left, may make it
more desirable to add more filtering in the phase detector. Some possibili­
ties include increasing the divider ratio of the phase detector or using a shift
register in the phase detector to determine when a number of-say,
four-shift-rights or - lefts have occurred. For the design in Figure 5.26, a
divide-by-two was used in the phase detector due to lock-time require­
ments.

535 Experimental Results

The RSDLL of Figure 5.22 was fabricated in a 0.21 pm, fbur-poly, dou­
ble-metal CMOS technology (a DRAM process). A 48-stage delay line with
an operation frequency of 125-250 MHz was used. The maximum operat­
ing frequency was limited by delays external to the DLL, such as the input
buffer and interconnect. There was no noticeable static phase error on either
the rising or falling edges. Figure 5.27 shows the resulting rms jitter versus
input frequency. One sigma ofjitter over the 125-250 MHz frequency range
was below 100 ps. The measured delay per stage versus Vcc and tempera­
ture is shown in Figure 5.28. Note that the 150 ps typical delay of a unit­
delay element was very close to the rise and fall times on-chip of the clock
signals and represents a practical minimum resolution of a DLL for use in a
DDR DRAM fabricated in a 0.21 gm process.
The power consumption (the current draw of the DLL when
Vcc = 2.8 V) of the prototype RSDLL is illustrated in Figure 5.29. It was
found that the power consumption was determined mainly by the dynamic
power dissipation of the symmetrical delay line. The NAND delays in this
test chip were implemented with 10/0.21 gm NMOS and 20/0.21 gm
PMOS. By reducing the widths of both the NMOS and PMOS transistors,
the power dissipation is greatly reduced without a speed or resolution pen­
alty (with the added benefit of reduced layout size).
Sec. 5.3 Synchronization in DRAMs 141

6o

5o

4o

(s
e 3o
』
忠二「

2o

1o

o
25
133 143 1 54 167 1 82 200
Input frequency (MHz)

Figure 5.27 Measured rms jitter versus input frequency.

5 0o

4 5o

4 0o

3 5o
(sd)

3 0o
0 6 E W A布

2 5o

2 0o
B Q

1 5o

1 0o

2.25 2.5 2.75 3 J.25 3.5 3.75 4

Figure 5.28 Measured delay per stage versus Vcc and temperature.
142 Chap. 5 Global Circuitry and Considerations

Freq (MHz)

Figure 5.29 Measured ICC (DLL current consumption) versus input frequency.

5.3.6 Discussion
In this section we have presented one possibility for the design of a
delay-locked loop. While there are others, this design is simple, manufac­
turable, and scalable.
In many situations the resolution of the phase detector must be
decreased. A useful circuit to determine which one of two signals occurs
earlier in time is shown in Figure 5.30. This circuit is called an arbiter. If SI
occurs slightly before S2 then the output SOI will go HIGH, while the out­
put SO2 stays LOW. If 52 occurs before SI, then the output SO2 goes HIGH
and SOI remains LOW. The fact that the inverters on the outputs are pow­
ered from the SR latch (the cross-coupled NAND gates) ensures that SOI
and SO2 cannot be HIGH at the same time. When designed and laid out cor­
rectly, this circuit is capable of discriminating tens of picoseconds of differ­
ence between the rising edges of the two input signals.
The arbiter alone cannot be capable of controlling the shift register. A
simple logic block to generate shift-right and shift-left signals is shown in
Figure 5.31. The rising edge of SO\ or SO2 is used to clock two D-latches
so that the shift-right and shift-left signals may be held HIGH for more than
one clock cycle. Figure 531 uses a divide-by-two to hold the shift signals
valid for two clock cycles. This is important because the output of the arbi­
ter can have glitches coming from the different times when the inputs go
back LOW. Note that using an arbiter-based phase detector alone can result
in an alternating sequence of shift-right, shift-left. We eliminated this prob­
lem in the phase-detector of Figure 5.24 by introducing the dead zone so
that a minimum delay spacing of the clocks would result in no shifting.
Sec. 5.3 Synchronization in DRAMs 143

Figure 5.30 Two-way arbiter as a phase detector.

Figure 531 Circuit for generating shift register control.

In some situations, the fundamental delay time of an element in the

delay line needs reduction. Using the NAND-based delay shown in Figure
5.26, we are limited to delay times much longer than a single inverter delay.
However, using a single inverter delay results in an inversion in our basic
cell. Figure 5.32 shows that using a double inverter-based delay element can
solve this problem. By crisscrossing the single delay element inputs the cell
appears to be non-inverting. This scheme results in the least delay because
the delay between the cell's input and output is only the delay of a single
inverter. The problems with using this type of cell over the NAND-based
delay are inserting the feedback clock and the ability of the shift register to
control the delay of the resulting delay line.
144 Chap. 5 Global Circuitry and Considerations

Figure 5.32 A double inverter used as a delay element.

Figure 5.33 shows how inserting transmission gates (TGs) that are con­
trolled by the shift register allows the insertion point to vary along the line.
When C is HIGH, the feedback clock is inserted into the output of the delay
stage. The inverters in the stage are isolated from the feedback clock by an
additional set of TGs. We might think, at first glance, that adding the TGs in
Figure 5.33 would increase the delay significantly; however, there is only a
single set of TGs in series with the feedback before the signal enters the
line. The other TGs can be implemented as part of the inverter to minimize
their impact on the overall cell delay. Figure 5.34 shows a possible inverter
implementation.

To output of
the delay line

Figure 533 Transmission gates added to delay line.

Finally, in many situations other phases of an input clock need generat­

ing. This is especially useful in minimizing the requirements on the setup
and hold times in a synchronous system. Figure 5.35 shows a method of
segmenting the delays in a DLL delay line. A single control register can be
used for all four delay segments. The challenge becomes reducing the
amount of delay in a single delay element. A shift in this segmented config­
uration results in a change in the overall delay of four delays rather than a
single delay. Note that with a little thought, and realizing that the delay ele­
ments of Figure 5.32 can be used in inverting or non-inverting configura­
tions, the delay line of Figure 5.35 can be implemented with only two
segments and still provide taps of 90°, 180°, 270°, and 360°.
Sec. 5.3 Synchronization in DRAMs 145

OUT

Figure 5.34 Inverter implementation.

Figure 5.35 Segmenting delays fbr additional clocking taps.

REFERENCES

[1] H. Ikeda and H. Inukai, ^High-speed DRAM architecture development,,5

IEEE Journal of Solid-State Circuits, vol. 34, pp. 685-692, May 1999.
[2] M. Bazes, “Two novel full complementary self-biased CMOS differential
amplifiers,0 IEEE Journal of Solid-State Circuits, vol. 26, pp. 165-168,
February 1991.
[3] R. J. Baker, CMOS: Circuit Design, Layout, and Simulation, 2nd ed.
Hoboken, NJ: IEEE Press, 2005.
[4] F. Lin, J. Miller, A. Schoenfeld, M. Ma, and R. J. Baker, “A register-controlled
symmetrical DLL for double-data-rate DRAM,“ IEEE Journal of Solid-State
Circuits, vol. 34, no. 4, 1999.
[5] A. Hatakeyama, H. Mochizuki, T. Aikawa, M. Takita, Y. Ishii, H. Tsuboi,
S. Y. Fiyioka, S. Yamaguchi, M. Koga, Y. Serizawa, K. Nishimura,
K. Kawabata, Y. Okajima, M. Kawano, H. Kojima, K. Mizutani, T. Anezaki,
M. Hasegawa, and M. Taguchi, “A 256-Mb SDRAM using a register-
controlled digital DLL," IEEE Journal of Solid-State Circuits, vol. 32,
pp. 1728-1732, November 1997.
[6] S. Eto, M. Matsumiya, M. Takita, Y. Ishii, T. Nakamura, K. Kawabata,
H. Kano, A. Kitamoto, T. Ikeda, T. Koga, M. Higashiro, Y. Serizawa,
K. Itabashi, 0. Tsuboi, Y. Yokoyama, and M. Taguchi, UA 1Gb SDRAM with
ground level precharged bitline and non-boosted 2.1V word line,“ in ISSCC
Digest of Technical Papers, pp, 82-83, February 1998.
146 Chap. 5 Global Circuitry and Considerations

[7] T. Saeki, Y. Nakaoka, M. Fujita, A. Tanaka, K. Nagata, K. Sakakibara,

T. Matano, Y. Hoshino, K. Miyano, S. Isa, S. Nakazawa, E. Kakehashi,
J. M. Drynan, M. Komuro, T. Fukase, H. Iwasaki, M. Takenaka, J. Sekine,
M. Igeta, N. Nakanishi, T. Itani, K. Yoshida, H. Yoshino, S. Hashimoto,
T. Yoshii, M. Ichinose, T. Imura, M. Uziie, S. Kikuchi, K. Koyama,
Y. Fukuzo, and T. Okuda, UA 2.5-ns clock access 250-MHz, 256-Mb SDRAM
with synchronous mirror delay/' IEEE Journal of Solid-State Circuits, vol. 31,
pp. 1656-1665, November 1996.
 Chapter
6

Voltage Converters
In this chapter, we discuss the circuitry for generating the on-chip voltages
that lie outside the supply range. In particular, we look at the wordline
pump voltage and the substrate pumps. We also discuss voltage regulators
that generate the internal power supply voltages.

6.1 INTERNAL VOLTAGE REGULATORS

6.1.1 Voltage Converters

DRAMs depend on a variety of internally generated voltages to operate
and to optimize their performance. These voltages generally include the
boosted wordline voltage VCcp, the internally regulated supply voltage VCCt
the VCCI2 cellplate and digitline bias voltage DVC2, and the pumped sub­
strate voltage VBB. Each of these voltages is regulated with a different kind
of voltage generator. A linear voltage converter generates VCCt while a
modified CMOS inverter creates DVC2. Generating the boosted supply
voltages VCcp and VBB requires sophisticated circuits that employ charge or
voltage pumps (a.k.a., charge pumps). In this chapter, we discuss each of
these circuits and how they are used in modem DRAM designs to generate
the required supply voltages.
Most modem DRAM designs rely on some form of internal voltage
regulation to convert the external supply voltage VCCx into an internal sup­
ply voltage FcC. We say most, not all, because the need for internal regula­
tion is dictated by the external voltage range and the process in which the
DRAM is based. The process determines gate oxide thickness, field device
characteristics, and diffused junction properties. Each of these properties, in
turn, affects breakdown voltages and leakage parameters, which limit the

147
148 Chap. 6 Voltage Converters

maximum operating voltage that the process can reliably tolerate. For
example, a 1 Gb DRAM built in a 45 nm CMOS process with a 14A thick
gate oxide can operate reliably with an internal supply voltage not exceed­
ing 1.2 V. If this design had to operate in a 3.3 V system, an internal voltage
regulator would be needed to convert the external 3.3 V supply to an inter­
nal 1.2 V supply. For the same design operating in a 1.2 V system, an inter­
nal voltage regulator would not be required. Although the actual operating
voltage is determined by process considerations and reliability studies, the
internal supply voltage is generally proportional to the minimum feature
size. Table 6.1 summarizes this relationship.

Table 6.1 DRAM process

versus supply voltage.

Process
vcc
Internal

90 nm 1.5 V

72 nm 1.35 V

45 nm 1.2 V

32 nm L0 V

All DRAM voltage regulators are built from the same basic elements: a
voltage reference, one or more output power stages, and some form of con­
trol circuit. How each of these elements is realized and combined into the
overall design is the product of process and design limitations and the
design engineer's preferences. In the paragraphs that follow, we discuss
each element, overall design objectives, and one or more circuit implemen­
tations.

6.1.2 Voltage References

The sole purpose of a voltage reference circuit is to establish a nominal
operating point for the voltage regulator circuit. However, this nominal
voltage is not a constant; rather, it is a function of the external voltage Vccx
and temperature. Figure 6.1 is a graph of Vcc versus VCCx for a typical
DRAM voltage regulator. This figure shows three regions of operation. The
first region occurs during a POWERUP or POWERDOWN cycle, in which
VCcx is below the desired Vcc operating voltage range. In this region, Vcc is
set equal to Vccx. providing the maximum operating voltage allowable in
the part. A maximum voltage is desirable in this region to extend the
Sec. 6.1 Internal Voltage Regulators 149

DRAM's operating range and ensure data retention during low-voltage con­
ditions.
The second region exists whenever Vccx is in the nominal operating
range. In this range, Vcc flattens out and establishes a relatively constant
supply voltage to the DRAM. Various manufacturers strive to make this
region absolutely flat, eliminating any dependence on We have found,
however, that a moderate amount of slope in this range for characterizing
performance is advantageous. It is critically important in a manufacturing
environment that each DRAM meet the advertised specifications, with some
margin for error. A simple way to ensure these margins is to exceed the
operating range by a fixed amount during component testing. The voltage
slope depicted in Figure 6.1 allows this margin testing to occur by establish-

Figure 6.1 Ideal regulator characteristics.

The third region shown in Figure 6.1 is used for component bum-in.
During bum-in, both the temperature and voltage are elevated above the
normal operating range to stress the DRAMs and weed out infant failures.
Again, if there were no VCcx and Vcc dependency, the internal voltage could
not be elevated. A variety of manufacturers do not use the monotonic curve
shown in Figure 6.1. Some designs break the curve as shown in Figure 6.2,
producing a step in the voltage characteristics. This step creates a region in
which the DRAM cannot be operated. We will focus on the more desirable
circuits that produce the curve shown in Figure 6.1.
lb design a voltage reference, we need to make some assumptions
about the power stages. First, we will assume that they are built as unbuf­
150 Chap. 6 Voltage Converters

fered, two-stage, CMOS operational amplifiers and that the gain of the first
stage is sufficiently large to regulate the output voltage to the desired accu­
racy. Second, we will assume that they have a closed loop gain of Av. The
value of Av influences not only the reference design, but also the operating
characteristics of the power stage (to be discussed shortly). For this design
example, assume Av = 1.5. The voltage reference circuit shown in
Figure 6.3 can realize the desired VCc characteristics shown in Figure 6.4.
This circuit uses a simple resistor and a PMOS diode reference stack that is
buffered and amplified by an unbuffered CMOS op-amp. The resistor and
diode are sized to provide the desired output voltage and temperature char­
acteristics and the minimum bias current. Note that the diode stack is pro­
grammed through the series of PMOS switch transistors that are shunting
the stack. A fuse element is connected to each PMOS switch gate. Unfortu­
nately, this programmability is necessary to accommodate process varia­
tions and design changes.

Figure 6.2 Alternative regulator characteristics.

Sec. 6.1 Internal Voltage Regulators 151

Figure 6.3 Resistor/diode voltage reference.

3.0 3.3 3.6 Vccx

Figure 6.4 Voltage regulator characteristics.

The reference temperature characteristics rely on establishing a proper

balance between VTH. mobility, and resistance variations with temperature.
An ideal temperature coefficient for this circuit would be positive such that
the voltage rises with temperature, somewhat compensating for the CMOS
gate delay speed loss associated with increasing temperature.
152 Chap. 6 Voltage Converters

Two PMOS diodes (M3—M4) connected in series with the op-amp out­
put terminal provide the necessary bum-in characteristics in Region 3. In
normal operation, the diodes are OFF. As Vccx is increased into the bum-in
range, Vccx will eventually exceed VREE by two diode drops, turning ON the
PMOS diodes and clamping VREF to Vth below Vccx. The clamping action
will establish the desired burn-in characteristics, keeping the regulator
monotonic in nature.
In Region 2, voltage slope over the operating range is determined by the
resistance ratio of the PMOS reference diode Ml and the bias resistor RI.
Slope reduction is accomplished by either increasing the effective PMOS
diode resistance or replacing the bias resistor with a more elaborate current
source as shown in Figure 65 This current source is based on a VTH refer­
enced source to provide a reference current that is only slightly dependent
on Vccx [1]. A slight dependence is still necessary to generate the desired
voltage slope.
The voltage reference does not actually generate Region 1 characteris­
tics for the voltage regulator Rather, the reference ensures a monotonic
transition from Region 1 to Region 2. To accomplish this task, the reference
must approximate the ideal characteristics for Region 1, in which
VCc = Vccx- The regulator actually implements Region 1 by shorting the
VCc and Vccx buses together through the PMOS output transistors found in
each power stage op-amp. Whenever Vccx is below a predetermined voltage
VI, the PMOS gates are driven to ground, actively shorting the buses
together. As Vccx exceeds the voltage level VI, the PMOS gates are
released and normal regulator operation commences. To ensure proper
DRAM operation, this transition needs to be as seamless as possible.

Figure 6.5 Improved voltage reference.

Sec. 6.1 Internal Voltage Regulators 153

6.1.3 Bandgap Reference

Another type of voltage reference that is popular among DRAM manu­
facturers is the bandgap reference. The bandgap reference is traditionally
built from vertical pnp transistors. A representative bandgap reference cir­
cuit is shown in Figure 6.6. As shown, it uses two bipolar transistors with an
emitter size ratio of 10:1. Because they are both biased with the same cur­
rent and owing to the different emitter sizes, a differential voltage will exist
between the two transistors. The differential voltage appearing across resis­
tor RI will be amplified by the op-amp. Resistors R2 and RI establish the
closed loop gain for this amplifier and determine nominal output voltage
and bias currents for the transistors [1].
The almost ideal temperature characteristics of a bandgap reference are
what make them attractive to regulator designers. Through careful selection
of emitter ratios and bias currents, the temperature coefficient can be set to
approximately zero. Also, because the reference voltage is determined by
the bandgap characteristics of silicon rather than a PMOS VTH, this circuit is
less sensitive to process variations than the circuit in Figure 6.3.
Three problems with the bandgap reference shown in Figure 6.6, how­
ever, make it much less suitable for DRAM applications. First, the bipolar
transistors need moderate current to ensure that they operate beyond the
knee on their I-V curves. This bias current is approximately 10-20 p.A per
transistor, which puts the total bias current fbr the circuit above 25)xA. The
voltage reference shown in Figure 6.3, on the other hand, consumes less
than 10 gA of the total bias current. Second, the vertical pnp transistors
inject a significant amount of current into the substrate-as high as 50% of
the total bias current in some cases. For a pumped substrate DRAM, the
resulting charge from this injected current must be removed by the substrate
pump, which raises standby current fbr the part. Third, the voltage slope for
a bandgap reference is almost zero because the feedback configuration in
Figure 6.6 has no dependence on Vccx- Seemingly ideal, this performance
is unacceptable because to perform margin testing on the DRAM a finite
voltage slope is needed.
154 Chap. 6 Voltage Converters

tR2

R5
(
Substrate
R3
R6
R7

Figure 6.6 Bandgap reference circuit.

6.1.4 The Power Stage

Although static voltage characteristics of the DRAM regulator are
determined by the voltage reference circuit, dynamic voltage characteristics
are dictated by the power stages. The power stage is therefore a critical ele­
ment in overall DRAM performance. To date, the most prevalent type of
power stage among DRAM designers is a simple, unbuffered op-amp.
Unbuffered op-amps, while providing high open loop gain, fast response,
and low offset, allow design engineers to use feedback in the overall regula­
tor design. Feedback reduces temperature and process sensitivity and
ensures better load regulation than any type of open loop system. Design of
the op-amps, however, is anything but simple.
The ideal power stage would have high bandwidth, high open-loop
gain, high slew rate, low systematic offset, low operating current, high
drive, and inherent stability. Unfortunately, several of these parameters are
contradictory, which compromises certain aspects of the design and necessi­
tates trade-offs. While it seems that many DRAM manufacturers use a sin­
gle op-amp for the regulator's power stage, we have found that it is better to
use a multitude of smaller op-amps. These smaller op-amps have wider
bandwidth, greater design flexibility, and an easier layout than a single,
large op-amp.
The power op-amp is shown in Figure 6.7. The schematic diagram for a
voltage regulator power stage is shown in Figure 6.8. This design is used on
a 256Mb DRAM and consists of 18 power op-amps, one boost amp, and
one small standby op-amp. The Vcc power buses fbr the array and periph­
eral circuits are isolated except fbr the 20-ohm resistor that bridges the two
together. Isolating the buses is important to prevent high-current spikes that
occur in the array circuits from affecting the peripheral circuits. Failure to
Sec. 6.1 Internal Voltage Regulators 155

isolate these buses can result in speed degradation fbr the DRAM because
high-current spikes in the array cause voltage cratering and a corresponding
slowdown in logic transitions.

Figure 6.7 Power op-amp.

With isolation, the peripheral Vcc is almost immune to array noise. Tb

improve slew rate, each of the power op-amp stages shown in Figure 6.8
features a boost circuit that raises the differential pair bias current and slew
rate during expected periods of large current spikes. Large spikes are nor­
mally associated with Psense-amp activation. To reduce active current con­
sumption, the boost current is disabled a short time after Psense-amp
activation by the signal labeled BOOST. The power stages themselves are
enabled by the signals labeled ENS and ENS* only when RAS* is LOW and
the part is active. When RAS^ is HIGH, all of the power stages are disabled.
The signal labeled CLAMP* ensures that the PMOS output transistor is
OFF whenever the amplifier is disabled to prevent unwanted charging of the
VCc bus. When forced to ground, however, the signal labeled PWRUP shorts
the VCcx and VCc buses together through the PMOS output transistor. (The
need fbr this function was described earlier in this section.) Basically, these
two buses are shorted together whenever the DRAM operates in Region 1,
as depicted in Figure 6.1. Obviously, to prevent a short circuit between VCcx
and ground, CLAMP* and PWRUP are mutually exclusive.
156 Chap. 6 Voltage Converters

CLAMPF*
ENS clamp- Qower op-amp (3X)
ENS* ENS*

V^VRUP PWRUP
BOOST
BOOSTF
CLAMPF*
cm/wp* Power op-amp (3X)
ENS
ENS* ENS*
V„,
V F^VRUP
v REF BOOST
VPWRUP
BOOSTF
clamp* Power op-amp (12X) Vcc Af
ENS*
“20 W
PWRUP
BOOST
PUMPBOOST Vcc Periph
PUMPBOQST*

Standby op-amp

DISABLEA

Figure 6.8 Power stage.

The smaller standby amplifier is included in this design to sustain the

Vcc supply whenever the part is inactive, as determined by RAS*. This
amplifier has a very low operating current and a correspondingly low slew
rate. Accordingly, the standby amplifier cannot sustain any type of active
load. For this reason, the third and final type of amplifier is included. The
boost amplifier shown in Figure 6.8 is identical to the power stage amplifi­
Sec. 6.1 Internal Voltage Regulators 157

ers except that the bias current is greatly reduced. This amplifier provides
the necessary supply current to operate the VCCP and VBS voltage pumps.
The final element in the voltage regulator is the control logic. An exam­
ple of this logic is shown in Figure 6.9. It consists primarily of static CMOS
logic gates and level translators. The logic gates are referenced to VCc-
Level translators are necessary to drive the power stages, which are refer­
enced to VCcx levels. A series of delay elements tune the control circuit rel­
ative to Psense-amp activation (ACT) and RAS* (RL*) timing. Included in
the control circuit is the block labeled VCcx level detector [1]. The reference
generator generates two reference signals, which are fed into the compara­
tor, to determine the transition point VI between Region 1 and Region 2
operation for the regulator. In addition, the boost amp control logic block is
shown in Figure 6.9. This circuit examines the VBB and VCcp control signals
to enable the boost amplifier whenever either voltage pump is active.
158 Chap. 6 Voltage Converters

6.2 PUMPS AND GENERATORS

Generation of the boosted wordline voltage and negative substrate bias volt­
age requires voltage pump (or charge pump) circuits. Pump circuits are
commonly used to create voltages that are more positive or negative than
available supply voltages. Two voltage pumps are commonly used in a
DRAM today. The first is a VCCP pump, which generates the boosted word­
line voltage and is built primarily from NMOS transistors. The second is a
VBB pump, which generates the negative substrate bias voltage and is built
from PMOS transistors. The exclusive use of NMOS or PMOS transistors
in each pump is required to prevent latch-up and current injection into the
Mbit arrays. NMOS transistors are required in the VCcp pump because vari­
ous active nodes would swing negative with respect to the substrate voltage
VCCP, Any n-difliision regions connected to these active nodes would for­
ward bias and cause latch-up and injection. Similar conditions mandate the
use of PMOS transistors in the VBB pump.

6.2.1 Pumps
Voltage pump operation can be understood with the assistance of the
simple voltage pump circuit depicted in Figure 6.10. For this positive pump
circuit, imagine, for one phase of a pump cycle, that the clock CLK is
HIGH. During this phase, node A is at ground and node B is clamped to
Vcc-Vth by transistor Ml. The charge stored in capacitor Cl is then

Q\ = C\ (Vcc- VTH) coulombs

(6#
During the second phase, the clock CLK will transition LOW, which brings
node A HIGH. As node A rises to Vcc, node B begins to rise to VCc +
(Vcc-Vth)^ shutting OFF transistor Ml. At the same time, as node B rises
one VTH above transistor M2 begins to conduct. The charge from
capacitor Cl is transferred through M2 and shared with the capacitor CLOAD.
This action effectively pumps charge into C^oad and ultimately raises the
voltage Vom. During subsequent clock cycles, the voltage pump continues
to deliver charge to CLOad until the voltage VO[JT equals ZVcc-VthTthz,
one VTH below the peak voltage occurring at node B. A simple, negative
voltage pump could be built from the circuit of Figure 6.10 by substituting
PMOS transistors for the two NMOS transistors shown and moving their
respective gate connections.
Schematics fbr actual VCCP and VBS pumps are shown in Figures 6.11
and 6.12, respectively. Both of these circuits are identical except for the
changes associated with the NMOS and PMOS transistors. These pump cir­
Sec. 6.2 Pumps and Generators 159

cuits operate as two phase pumps because two identical pumps are working
in tandem. As discussed in the previous paragraph, note that transistors Ml
and M2 are configured as switches rather than as diodes. The drive signals
for these gates are derived from secondary pump stages and the tandem
pump circuit. Using switches rather than diodes improves pumping effi­
ciency and operating range by eliminating the VTH drops associated with
diodes.

Figure 6.10 Simple voltage pump circuit.

Two important characteristics of a voltage pump are capacity and effi­

ciency. Capacity is a measure of how much current a pump can continue to
supply, and it is determined primarily by the capacitor's size and its operat­
ing frequency. The operating frequency is limited by the rate at which the
pump capacitor Cl can be charged and discharged. Efficiency, on the other
hand, is a measure of how much charge or current is wasted during each
pump cycle. A typical DRAM voltage pump might be 30-50% efficient.
This translates into 2-3 milliamps of supply current for every milliamp of
pump output current.
In addition to the pump circuits just described, regulator and oscillator
circuits are needed to complete a voltage pump design. The most common
oscillator used in voltage pumps is the standard CMOS ring oscillator. A
typical voltage pump ring oscillator is shown in Figure 6.13. A unique fea­
ture of this oscillator is the multifrequency operation permitted by including
mux circuits connected to various oscillator tap points. These muxes, con­
trolled by signals such as PWRUP, enable higher frequency operation by
reducing the number of inverter stages in the ring oscillator.
Typically, the oscillator is operated at a higher frequency when the
DRAM is in a PWRUP state, since this will assist the pump in initially
160 Chap. 6 Voltage Converters

charging the load capacitors. The oscillator is enabled and disabled through
the signal labeled REGDIS*. This signal is controlled by the voltage regula­
tor circuit shown in Figure 6.14. Whenever REGDIS* is HIGH, the oscilla­
tor is functional, and the pump is operative. Examples of VCCP and %
pump regulators are shown in Figure 6.14 and 6.15, respectively.

Figure 6.12 VBB pump.

REGDIS'
PWRUP o-------- [>> --------- - 4>o->oI-►CLK

a
Figure 6.13 Ring oscillator
Sec. 6.2 Pumps and Generators 161

冬
二
. .

一
del
r

wr
Figure 6.14 vCcp regulator.

Oscillator
ENABLE*

Feedback

Figure 6.15 VRB regulator.

These circuits use low-bias current sources and NMOS/PMOS diodes to

translate the pumped voltage levels, VCcp and to normal voltage levels.
This shifted voltage level is fed into a modified inverter stage that has an
output-dependent trip point (the inverter has hysteresis). The trip point is
modified with feedback to provide hysteresis for the circuit. Subsequent
inverter stages provide additional gain for the regulator and boost the signal
to the full CMOS level necessary to drive the oscillator. Minimum and max­
imum operating voltages for the pump are controlled by the first inverter
stage trip point, hysteresis, and the NMOS/PMOS diode voltages. The
clamp circuit shown in Figure 6.14 is included in the regulator design to
limit the pump voltage when VCc is elevated, such as during burn-in.
A second style of voltage pump regulator fbr controlling VCCP and VBS
voltages is shown in Figure 6.16 and 6.17, respectively. This type of regula­
tor uses a high-gain comparator coupled to a voltage translator stage. The
translator stage of Figure 6.16 translates the pumped voltage VCCP and the
reference voltage VDD down within the input common-mode range of the
comparator circuit. The translator accomplishes this with a reference current
162 Chap. 6 Voltage Converters

source and MOS diodes. The reference voltage supply VDD is translated
down by one threshold voltage (VTH) by sinking reference current from a
current mirror through a PMOS diode. The pumped voltage supply VCcp is
similarly translated down by sinking the same reference current with a
matching current mirror through a diode stack. The diode stack consists of a
PMOS diode, matching that in the VDD reference translator, and a pseudo-
NMOS diode. The pseudo-NMOS diode is actually a series ofNMOS tran­
sistors with a common gate connection.
Transistor quantity and sizes included in the pseudo-NMOS diode are
mask-programmable. The voltage drop across this pseudo-NMOS diode
determines the regulated voltage for KecP. Accordingly,

Vqqp = KpD+ Vndiode

(6.2)

The voltage dropped across the PMOS diode does not affect the regulated
voltage because the reference voltage supply VDD is translated through a
matching PMOS diode. Both translated voltages are fed into a comparator
stage, enabling the pump oscillator whenever the translated VCcp voltage
falls below the translated VDD reference voltage. The comparator hysteresis
dictates the amount of ripple present on the regulated VCCP supply.
The VBB regulator in Figure 6.17 operates in a similar fashion to the
VCCP regulator of Figure 6.16. The primary difference lies in the voltage
translator stage. For the VBB regulator, this stage translates the pumped volt­
age VBB and the reference voltage Vss up within the input common mode
range of the comparator: The reference voltage Vss is translated up by one
threshold voltage (VTH) by sourcing a reference current from a current mir­
ror through an NMOS diode. The regulated voltage VSB is similarly trans­
lated up by sourcing the same reference current with a matching current
mirror through a diode stack. This diode stack contains an NMOS diode that
matches that used in translating the reference voltage Vss. The stack also
contains a mask-adjustable, pseudo-NMOS diode. The voltage across the
pseudo-NMOS diode determines the regulated voltage for VBB such that

VBB = -Vndiode

(6.3)

The amount of ripple present on the regulated VBB supply is dictated by

comparator hysteresis.
Sec. 6.2 Pumps and Generators 163

Figure 6.16 Vccp differential regulator

Figure 6.17 VBl
5 differential regulaton
chap* 6 v o y a g e Converters

Comparator
Sec. 6.3 Discussion 165

6.2.2 DVC2 Generator

With our discussion of voltage regulator circuits concluded, we can
briefly turn our attention to the DVC2 generation. As discussed in
Section 1.2, the memory capacitor cellplate is biased to VCCI2. Furthermore,
the digitlines are always equilibrated and biased to 喂/2 between anay
accesses. In most DRAM designs, the cellplate and digitline bias voltages
are derived from the same generator circuit. A simple circuit fbr generating
Pcc/2 (DVC2) voltage is shown in Figure 6.18. It consists of a standard
CMOS inverter with the input and output terminals shorted together. With
correct transistor sizing, the output voltage of this circuit can be set pre­
cisely to Vee。V. Taking this simple design one step further results in the
actual DRAM DVC2 generator found in Figure 6.19. This generator con­
tains additional transistors to improve both its stability and drive capability.

6.3 DISCUSSION
In this chapter, we introduced the popular circuits used on a DRAM fbr volt­
age generation and regulation. Because this introduction is far from exhaus­
tive, we include a list of relevant readings and references in the Appendix
for those readers interested in greater detail.
166 Chap. 6 Voltage Converters

Modified
inverter

Driver
bias DVC2

ENABLE*

Figure 6.19 DVC2 generator.

REFERENCES

[1] R. Baker, CMOS: Circuit Design, Layout, and Simulation, 2nd ed. Hoboken,
NJ: IEEE Press, 2005.
[2] B. Keeth. 1994. Control circuit responsive to its supply voltage level. US
Patent 5,373,227, issued December 13,1994.
 Chapter
7

An Introduction to
High-Speed DRAM
Historically, the semiconductor industry has been the slave to two masters:
cost and performance. Cost, or more specifically cost reduction, is a matter
of pure economic survival for most companies in this industry. It is cost
after all that drives the industry to aggressively shrink process technology
year after year. Memory chips, due to their status as a commodity product,
are highly sensitive to cost. This fact fosters additional incentives for mem­
ory companies to push their process technology development faster than
most other segments of the semiconductor industry. The historical record is
littered with the names of companies that produced memory products at one
time or another, but subsequently dropped out of the marketplace. A key
contributor to this attrition is the ability or inability of any given company
to remain cost competitive. Market forces, not supplier controls, determine
the basic selling price for commodity memory-most notably, the balance
or imbalance in supply and demand. This leaves cost as the only component
of a balance sheet that can be controlled. Control costs and survive. Lose
control and fade away.

7.1 THE PERFORMANCE PARADIGM

While a complete analysis of the cost issues facing memory companies
would be interesting, the remainder of this book deals with a topic more fas­
cinating and dear to engineering professionals: performance. The word per­
formance^ much like the words quality or value, is rather ambiguous. Yet,
when we talk about performance, most people assume we are referring to
high performance. This is a natural and appropriate assumption, as we
rarely strive for low performance, let alone write books about it. Just to set

167
168 Chap. 7 An Introduction to High-Speed DRAM

the record straight, the remaining chapters deal entirely with high-perfor­
mance DRAM.
Now, DRAM may not be what first comes to mind when you hear the
phrase high performance. We agree with that sentiment as images of race
cars and fighter aircraft flash through our minds. But, when we come back
to reality and refocus our attention on the semiconductor industry, DRAM
memory devices still don't come to mind. Rather, we are more likely to con­
sider microprocessors. Clearly, these devices are on the leading edge of high
performance, and they benefit from the use of the latest transistor, metal,
and packaging technology in the quest for speed and processing muscle.
Memory, on the other hand, tends to limp along with slower transistors,
a minimum of metal layers, and inexpensive packaging. The goal at DRAM
companies has always been to meet the necessary speed targets while keep­
ing manufacturing costs to an absolute minimum. This doesn't mean that
the DRAM processes are not highly advanced. They clearly are advanced.
The real difference between microprocessor and DRAM process technology
has been the focus of process development. While the focus for micropro­
cessors has been performance, the focus in the memory sector has always
been cost. For DRAM, this focus has become a myopic quest for smaller
and smaller process geometries, with the ultimate goal of yielding more die
per wafer and, hence, at a lower cost than the competition.
Only recently has the need for higher performance DRAM become a
pressing concern. There are two sectors of the memory industry that are
driving this issue. The first is high-performance desktop systems; the sec­
ond is high-end graphics cards. Both of these sectors are extremely compet­
itive, and they are highly prized fbr the pricing margins and marketing
mileage available to the best in class. With processor speeds racing into the
multi-gigahertz range there are greater incentives fbr the memory sub­
system to try and keep pace. Pricing pressures are lower in this sector,
which gives DRAM manufacturers more latitude with their process technol­
ogy, design architectures, and packaging in order to realize desired perfor­
mance gains. These performance gains and the underlying technology
eventually trickle down, over time, to the higher volume desktop mar­
ket-increasing performance levels fbr the mainstream market, but ulti­
mately raising the bar again in the high-performance desktop market.
DRAM devices developed fbr the graphics card markets follow a simi­
lar path, except at a more accelerated pace. High-end graphics cards create a
need for very high-performance DRAM devices. Given lower pricing pres­
sure, DRAM manufacturers can improve their process technology, designs,
and packaging to significantly increase performance. Ultimately, these
improvements trickle down to other portions of the graphics market and
Sec. 7.2 Performance for DRAM Memory Devices 169

even into those main memory devices being developed for the high-perfor­
mance desktop market.

7.2 PERFORMANCE FOR DRAM MEMORY DEVICES

So, how do we define high performance for DRAM memory devices? His­
torically, asynchronous, fast page mode-style DRAM device performance
was dictated by timing specifications, such as 3 (access time from column
access), tcAC (access time from CAS*), (access time from RAS^), and
tpc (Read or Write cycle time). Note that these specs, save the final one, are
all measures of access time, which is the amount of time it takes for the
device to drive out data following the transition of the column address, the
column address strobe (C/S*), or the row address strobe (RAS*). Only the
final spec, tpc，relates to the cycle time, that is, the ability of the device to
deliver data at a sustained rate from one column address to anothe匚 There
were many other specifications on an asynchronous DRAM data sheet, but
these four, far and away, defined performance.
What was so important about these particular specifications? With
regard to 3, tcAG and Qc，these specs ail defined some sort of delay:
delay that the memory controller must endure every time that new data is
needed from the DRAM. The shorter the delay, the sooner the memory con­
troller could ultimately deliver data to the processor. Whenever the proces­
sor waits fbr data, its performance, and that of the system, suffers. Any
improvement in these critical DRAM timing specifications ultimately leads
to improvements in system performance. As a result, the memory industry
naturally migrates towards the development of devices with faster access
and shorter cycle times.
Following the advent of synchronous operation, DRAM device specifi­
cations began a slow migration towards performance metrics related to
clock frequency. The most important timing parameters became t^c (access
time from clock), tcK (clock cycle time), CL (CAS latency), and t^cD
(active to read or write delay). Most of these new parameters are specified
in terms of nanoseconds, except for CAS latency, which is specified in num­
ber of clock cycles. Again, all, save tcK，are a measure of delay. While the
Uc parameter is somewhat important, because it defines data availability
relative to clock transitions, real performance is measured by CAS latency
and These two parameters define how fast the DRAM device can cycle
(Jck) and how many clock cycles the memory controller, and ultimately the
system, must wait fbr data following a request for data (a Read operation).
Of course, system designers want their DRAM devices to have the low­
est CAS latency possible at any given operating frequency. This combina­
170 Chap. 7 An Introduction to High-Speed DRAM

tion yields the best system performance, but reality being what it is, won't
result in acceptable device yields for the memory manufacturer. Generally,
both tcK and CAS latency incrementally improve with successive genera­
tions. Both memory manufacturers and system builders manage device per­
formance, and the inevitable improvements, through a system known as
speed grading. Basically, a speed grade is a set of timing parameters associ­
ated with device operation at a specific data bus frequency. At higher bus
frequencies, these timing parameters become more demanding and harder
to meet. Fortunately for the memory industry, devices that are unable to
meet all of the timing parameters for a specific speed grade can be down-
binned to a slower speed grade and sold into other market segments. Speed
grades are not only important to ensure that components from different sup­
pliers are compatible, but also to support down-binning strategies that
improve overall production yields.
Synchronous DRAM (SDRAM), despite notable improvements in per­
formance, represents only an evolution in technology. When SDRAM first
appeared, manufacturers essentially converted existing asynchronous
designs by merely changing I/O circuits to accommodate the synchronous
interface. Even the device specifications remained rooted in asynchronous
behavior. The real advantage gained by synchronous operation occurred in
the output timing specs. Asynchronous designs had very limited data rates
due to the fact that the memory controller could not issue a new data request
(Read operation) until it successfully captured data from the previous Read.
The controller drove CAS* LOW to initiate the Read operation and held
CAS* LOW until it was able to capture the Read data. This meant that col­
umn cycle times included not just the DRAM access delay, but also the I/O
transfer delays, including flight time. Accordingly, column cycle time was
not optimized, because the controller was precluded from issuing a subse­
quent read command until it had captured data from the current read com­
mand.
Extended Data Out (EDO) asynchronous DRAMs reduced column
cycle time by hiding some of the I/O transfer delay. While traditional asyn­
chronous devices, such as Page Mode (PM) or Fast Page Mode (FPM), only
drive Read data while CAS* is LOW, EDO devices continue to fire Read
data after CAS^ transitions HIGH. This allows the memory controller to
begin to cycle the column for a subsequent Read while concurrently captur­
ing data from the cunent Read. This seemingly small change allowed a rea­
sonable gain in system performance.
SDRAM represents another step in the steady evolution of DRAM tech­
nology towards higher performance. The synchronous interface specifica­
tion is important, as it essentially divides the Read operation into two
separate components: the column access phase and the data output phase.
Sec. 7.3 Underlying Technology Improvements 171

Essentially, the data output operation was pushed into a separate clock cycle
from that of the array access. This meant that device speed and the associ­
ated clock frequency were now only limited by how fast the memory array
could perform column accesses. Synchronization ultimately increased per­
formance because the data output operation became a scheduled event, iso­
lated from the array access.
Double data rate (DDR)-style DRAM devices and their faster cousins,
graphics double data rate (GDDR) devices, represent an evolution in tech­
nology compared to synchronous DRAM (SDRAM). For each Read com­
mand, they drive two bits of data for every clock cycle: one bit on the rising
edge and a second bit on the falling edge. This simple change, dating from
technology developed in the early 1980s, effectively doubles data transfer
rates to and from the DRAM [1]. The various permutations of DDR and
GDDR, such as DDR2, DDR3, GDDR2, GDDR3, and GDDR4, encompass
evolutionary advances in technology to achieve higher overall data transfer
rates. These advances are enabled by changes in burst length, improvements
in signaling technology and packaging, and significant advances in circuit
design.

7.3 UNDERLYING TECHNOLOGY IMPROVEMENTS

In the balance of the book we discuss the underlying technology improve­
ments that enable high-speed DRAM. Many of these improvements don't
show up all at once, but begin to appear, in various forms, at each step of the
high-speed staircase. Some of the advances are somewhat delayed, if not
outright restrained, by legacy baggage that gets carried along from one gen­
eration to the next. We discuss the importance of this issue as well in
upcoming chapters.

REFERENCE

[1]M. F. Novak, K. M. Guttag, and D. J. Redwine. 1987. Control of data access to

memory fbr improved video system. US Patent 4,688,197, issued Aug. 18,
1987.
 Chapter

High-Speed Die Architectures

8.1 INTRODUCTION: OPTIMIZING DRAM ARCHITECTURE FOR
HIGH PERFORMANCE
Given the fact that the DRAM market is a commodity market, manufactur­
ers struggle to squeeze out profit in the only way available to them, by both
controlling and reducing cost, whenever and wherever possible. Because
die size is a leading indicator of cost, designers have always placed signifi­
cant emphasis on producing highly efficient and minimized die architec­
tures. As the saying goes, less is more. Standard die architectures are highly
optimized as a result of this quest for lower cost. At higher clock and data
rates, die architectures must undergo a transformation or re-optimization,
which places somewhat greater value on performance. While cost remains
an overriding concern, the quest fbr performance places new demands on
the design. When comparing die photographs of standard and high-speed
DRAM devices, as shown in Figure 8.1, the difference in relative size may
be easy to spot, but detailed changes in architecture may be harder to detect.
In point of fact, the differences can be subtle, but nevertheless necessary to
permit sustained operation at higher clock and data rates. In this chapter, we
explore high-speed die architectures and, in the process, gain insight into
many high-speed design issues.

173
Figure 8.1 a) Standard DDR2 die photo and b) high-speed GDDR3 die photo.
lliiillliHHiiiliiililiiHHilil

7卯闾才4 •

chap8
■删唧黜螂:
出|册朋那曲眠池常

H _ g h speed Die
A—
「
c h = e c t u es
「
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 175

8.2 ARCHITECTURAL FEATURES: BANDWIDTH, LATENCY,

AND CYCLE TIME
To gain an understanding of features specific to a high-speed die architec­
ture, let's first examine the limitations of standard die architectures. Exactly
what improvements are needed to reach the next performance level on a
standard die? The three device characteristics most closely associated with
high-speed performance are bandwidth, latency, and cycle time. Bandwidth
is the rate at which data can be written to or read from the memory device.
Typically it is measured in bits per second, per pin. A standard DDR2
device, for instance, is designed to go as high as 800Mbps/pin. GDDR3
memory is designed to reach speeds of 1.6Gbps/pin and GDDR4 upwards
of 2.5Gbps/pin. Latency is the delay between a request and a response. Read
latency, accordingly, is the delay from the time that a Read command is
issued by the host to the time that the Read data is made available to the
host. Cycle time, in the context of memory, is the rate at which array activi­
ties can be repeated. Column cycle time, specifically, is the rate at which
back-to-back, column-specific commands, such as Reads or Writes, can be
supported within a single memory block. So, as expected, a high-perfor­
mance device should be significantly better in one or more of these perfor­
mance metrics than what we might call a standard memory device.
However, try to keep in mind that performance and speed are in a state of
perpetual migration. Today's high-performance parts eventually become
tomorrow's standard-performance parts. Everything is relative.
While die size and cost are of paramount importance, most DRAM
designs start with specific targets for bandwidth, latency, and cycle time.
Let,s take a look at each of these metrics to see how they can be impacted
by die architecture.
Typical synchronous DRAM devices achieve higher data rates and
bandwidth by fetching multiple bits of data per data pin (£)0 for every
Read operation. For instance, a DDR2 device with eight data pins (X8)
fetches or pre-fetches 32 bits of data during each Read operation. These 32
bits are divided amongst the eight DQ pins so that a total of four bits of data
are driven out serially over two full clock cycles: one data bit per clock
edge. So what limits a DDR2 device to only 800Mbps/pin data rates for
Read operations? Basically, any circuit that is called into action during a
Read operation is capable of limiting performance, including command and
address receiver circuits, capture circuits, command decoders, column
address path, column decoders, array data path, secondary sense amplifiers,
Read data path, clock generation circuits, clock trees, serialization circuits,
and output drivers. Lefs assume however, that there are no limitations
resulting from circuit shortcomings and focus only on architectural limiters.
176 Chap. 8 High-Speed Die Architectures

This leaves us with the array data path, Read data path, and any limitations
caused by the locality of circuit blocks.

8.2.1 Architectural Limiters: The Array Data Path

The array data path is the heart of any DRAM. It consists of the sense
amplifier 7/0 transistors, local I/O lines, global I/O lines, and any DC sense
amplifiers or helper flip-flops (HFF) along with Write drivers for Write
operations. So, in order to increase DRAM bandwidth, the array data path
must cycle faster. Essentially, bandwidth is equal to the burst length divided
by the column cycle time as shown.
Bandwidth = Burst length / Column cycle time

(8.1)
Hence, a shorter column cycle time yields a higher data rate (bandwidth).
Note that the column cycle time is expressed as 及3 亩 DRAM device data
sheets. This relationship leads to the conclusion that higher bandwidths
require faster column cycle times, which brings us right back to the array
data path. Historically, column cycle time hasn't ventured below the 5 ns
mark for the majority of designs. In fact, burst length typically increases
from one interface technology to the next to maintain the column cycle time
above 5 ns. Table 8.1 shows the relationship between column cycle time,
burst length, and bandwidth for a variety of DRAM technologies.

Table 8.1 DRAM technology and speed relationships.

Min Column
Peak
Technology Burst Length Cycle Time
Bandwidth
(tcco)

SDRAM 1 6 ns 167Mbps/pin

DDR 2 5 ns 400Mbps/pin

DDR2 4 5 ns 800Mbps/pin

DDR3 8 5 ns 1.6Gbps/pin

GDDR3 4 2.5 ns 1.6Gbps/pin

GDDR4 8 3.2 ns 2.5Gbps/pin

The apparent exceptions to a 5 ns limit are GDDR3 and GDDR4 graph­

ics DRAM. So, what allows a graphics device to cycle faster than other
technologies? Lefs explore a GDDR3 array data path and find out.
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 177

To begin, we'll review what happens within a DRAM array data path
during a typical column cycle. As shown in Figure 8.2, the array data path
consists of a column decode circuit; column select lines; I/O transistors,
which connect each sense amplifier to local I/O lines; global I/O lines; and
helper flip-flop (HFF) circuits. In a column cycle, as shown in Figure 8.3,
the column select line fires HIGH, and the I/O and HFF equilibration cir­
cuits are released. This permits the sense amplifier to drive data onto the
local and global I/O lines, thus forcing a voltage differential to develop on
the lines. This differential signal drives into the HFF, shown in Figure 8.4,
which is fired to both amplify and latch the data for subsequent handling by
the Read data path. Finally, the I/O equilibration circuit fires, and the col­
umn select line drives LOW. In a GDDR3 device, all of this action takes
place in as little as 2.5 ns.

Figure 8.2 Array data path block diagram.

178 Chap. 8 High-Speed Die Architectures

3.0
2.8
2.6
2.4
2.2
2.0
1.8
1.6
1.4
1.2
1.0
.8
.6
.4
.2
0

1.8
1.6
1.4
1.2
1.0
.8
.6
.4
.2
0

Figure 8.3 Array data path 2.5 ns column cycle waveforms (0.5 ns/division).

To maximize the performance of the array data path, all of its compo­
nents must be carefully designed fbr high-speed operation. The column
decoder circuit and the associated redundancy logic must be designed to
cycle very quickly. This may require adapting different types of logic styles
into the design and modifying how redundant columns are handled. Tradi­
tional column redundancy techniques push out column cycle time to allow
fbr address matching and decoder delays. Alternative schemes, such as col­
umn steering or data suppression, may need to be considered in place of the
more popular address suppression technique. Also, the global column select
lines, which connect the column decoder to the sense amplifier I/O gates,
must be carefully constructed to minimize the RC time constant. Finally, the
I/O transistors must be properly sized to strike an optimal balance between
minimum loading on the column select line (via smaller gate area) and max­
imum Read/Write performance (via low source/drain resistance). Necessary
compromises such as this are common in the high-speed design realm.
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 179

The 2.5 ns column operation shown in Figure 8.3 is a good example of

finely tuned column cycle timing. Note that the I/O equalization signal
IOEQ^ is active LOW for no more than 600 ps. This timing is tightly syn­
chronized with HFF equalization as indicated by the HFFEg* signal in Fig­
ure 8.3. The column select (CSEL) transition is carefully placed within the
equalization envelope, while exhibiting very similar rise and fall times to
IOEQ* itself. Given the short timing intervals involved, the I/O lines must
split quickly to adequately separate the Data lines before the flip-flop fires,
as indicated by the HFFSense* driving LOW. Flip-flop firing is held off to
develop as much separation (signal) as possible for correct sensing. How­
ever, adequate time must be left in the cycle to transfer data into a secondary
latch before the I/O lines and HFF equalize in preparation for a subsequent
column cycle.
A mzyor consideration, which is not apparent in the timing waveforms
shown, is that the various circuits involved in column cycle operation can be
distributed over large distances. Also, there is generally a multitude of dis­
tributed helper flip-flops and sense amplifier strips involved, creating an
even more challenging timing problem, especially for tight column cycle
180 Chap. 8 High-Speed Die Architectures

times. One good technique fbr managing the spatial nature of this problem
is to design the signal nets with identical flight times. For instance, refening
back to Figure 8.2, the flight times of the column select lines (CSEL) must
match the flight times of the various signals controlling the helper flip­
flops. This requires us to carefully manage loading, wire widths, and drivers
so as to match both the propagation delay and the rise/fall times of these
critically timed signals. Any carelessness or shortcut taken in this regard
pushes out column cycle times. Likewise, all of the timing control circuits
that drive the column decoder and array data path must be similar in struc­
ture and design to ensure that they track over temperature and voltage varia­
tions. This attention to detail is necessary so that the timing relationships,
necessary fbr fast cycle times, are consistent despite anticipated changes in
the environment.
Column cycle time can be significantly affected by architectural choices
and by the construction of the array data path. These choices include array
block size, array partitioning, column decoder configuration and location,
global column select design, HFF configuration and location, routing metal
utilization, local I/O line length, global I/O line configuration, and any deci­
sion to share resources between array banks. Let's walk through a simple set
of examples to see the impact that each of these choices may have on speed.
Start by examining Figure 8.5, which shows a simplified block diagram
of a memory array and an array data path. This example is typical of that
found in a DDR or DDR2 SDRAM device. It consists of a pair of array
blocks that are each subdivided by sense amplifier strips into four array
cores. Each array block adjoins column and row decoder blocks. At each
wordline and column select intersection we gain access to four bits of data,
two bits per sense amplifier strip. In this example, the HFF block is shared
between two array blocks. The shared HFF block necessitates the addition
of multiplexers onto each HFF so that the I/O lines can be selected from one
array block or the other. These multiplexers increase loading on the I/O
lines and invariably impact cycle time. The sizing of these multiplexers as
well as of the I/O equilibration devices is critical in order to keep column
cycle times minimized.
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 181

I/O lines

Wordline
I/O lines
」

sdo-"®」
岩 。。昭

Japooop
M
®
B

Moa
s
® 。

o
工

Memory array

Memory array

Sense amp strip Sense amp strip

Column decoder Column decoder

Figure 8.5 Array block diagram.

The array design as shown in Figure 8.5 is capable of accessing only

four bits of data at a time. As a result, the HFF block needs to contain only
four helper flip-flops. Because there are actually five sense amplifier strips
that can feed data into the helper flip-flops, an additional level of multiplex­
ing would be needed to fully implement the array data path. Once again,
these multiplexers will impact column cycle time if their design is not care­
fully implemented. To access more than four data bits per column cycle
would require basic architectural changes to the design.
The simplified array block diagram shown in Figure 8.6 is modified to
permit two wordline segments to fire rather than one. With this change,
eight bits of data are available from every column access, which doubles the
number of helper flip-flops needed in the HFF block. Because any given
HFF services only half of the array block, as compared to the configuration
shown in Figure 8.5, its I/O lines can be shortened by almost fifty percent.
Also, the I/O line multiplexers do not need to connect with as many sense
182 Chap. 8 High-Speed Die Architectures

amplifier strips, which simplifies their design and reduces their effective
loading on the I/O lines. These changes to the I/O lines and multiplexers
support reduced column cycle time and higher bandwidth performance. All
of this results from a simple change to the array architecture.

-' I/O lines

Wordline

I/O lines
」

」
①
o)po
p o 。8 P M o a

° I/O lines

。
3P M o a
Word line

I/O lines

Memory array Memory array

Sense amp strip Sense amp strip

Column decoder Column decoder

Figure 8.6 Array block diagram, double wordline activation.

Another possible architectural change to the array architecture is shown

in Figure 8.7. In this embodiment, two column select lines are fired across a
single wordline segment for every array access. This configuration also
delivers eight bits of data for every access, requiring a like number of helper
flip-flops in the HFF block. However in this design, the number of I/O lines
running through the sense amplifier strips is doubled from that shown in
Figure 8.5 and Figure 8.6. There is no relief with regard to the length of the
I/O lines or the level of multiplexing needed to support the entire anay
block. So, while this configuration delivers the amount of data required by
our specification, its speed is comparable to the array shown in Figure 8.5,
and it is larger because the number of I/O lines has doubled. While this
approach may not seem preferable to the double wordline activation
approach in Figure 8.6, it still finds application since many device designs
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 183

have wordline activation restnctions imposed upon them by either the spec­
ification or by simple power constraints, lypically, most designs use some
combination of multiple wordline and multiple column activation to satisfy
data access requirements.
I/O lines

Wordline

I/O lines
--

s
wpoog Moa

0
do

5fd
e 8
1
」 不工

&

UJ
s
s
o o

Memory array Memory array

Sense amp strip Sense amp strip

Column decoder Column decoder

Figure 8.7 Array block diagram, double column activation.

8.2.2 Architectural Limiters: The Read Data Path

The next architectural limiter to higher bandwidth is the Read data path.
The Read data path essentially starts off where the array data path ends: at
the helper flip-flops. The Read data path can also limit device bandwidth if
it suffers from cycle time limitations. A block diagram of a typical Read
data path is shown in Figure 8.8. It consists of the helper flip-flops, one or
more secondary latches, the data pipeline, the data serializer, and the output
driver circuits. The Read data path design is generally constrained by both
the array architecture and the locations of the bond pads. Lefs take a look at
how these constraints affect the Read data path and its performance.
184 Chap. 8 High-Speed Die Architectures

Figure 8.8 Read data path block diagram.

Figure 8.9 shows a simplified floor plan for a high-speed GDDR3 DRAM
device. Graphics memory devices, typified by GDDR3, typically have 32
DQ pins (X32). GDDR3 and GDDR4 standards, as defined by JEDEC
(Joint Electron Device Engineering Council), divide these 32 DQ pins into
four byte groups with eight DQ pins in each group. As shown in
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 185

Figure 8.10, the GDDR3 package is pinned out so that there is a byte group
located in each of the four comers, with the command and address pins
(C/A) located in the center. This unique feature permits the floor plan of Fig­
ure 8.9 to be divided into four quadrants, where each quadrant services a
separate byte group.
一_--------------

,J Byte group \"、

Figure 8.9 GDDR3 floor plan diagram.

To help facilitate discussion, one of the four identical byte groups is cir­
cled in Figure 8.9. Each byte group consists of a 32-bit data path that oper­
ates on four bits of data for each of the eight DQ pins associated with the
group. The byte groups are fairly compact owing to the very pinout that
gives rise to them. Package pinouts fbr SDRAM, DDR, DDR2, and DDR3,
on the other hand, locate the DQ pins and C/A pins on opposite ends of the
package. While this arrangement simplifies module routing, it penalizes the
on-die data path by stretching it out along the entire length of the die. The
compact GDDR3 byte group of Figure 8.9 produces shorter routing lengths
and better timing control to enable faster cycle times and higher overall per­
formance. The GDDR3 device implements two parallel pad rows. This fea­
ture permits all address, control, and timing circuits to exist between the
pads. This in turn allows these circuits to be centered between all the byte
groups, while keeping address and control timing tight and signal routing
short.
186 Chap. 8 High-Speed Die Architectures

Byte group Byte group

X 、
<^DDQ
VDO VSS \ZQ /MF VSS VDD VQDQ

VS*
VSSQ DQO DQ1 VS5Q VSSQ DQ9 DQ8
VDD4
\
c VDDQ DQ2 DQ3 VDDQ DQ11 DQ10 vooq
vssj
D VSSQ WDQ2C RDQ20 VSSQ RDQ21 WDQ21 vssq
VD&
\E VDDQ DQ4 DMO VDDQ DM1 DQ12 VDqQ
"DQ5^ 以S*
VDD D96- \cs* 't>Q43 DQ14 痂0
VS^ AZSSQ. -0(37 BAO BAI^ 'BQ15. VSS

H yREF A1 RAS* CKE WE* BA2 A5 \ VREF

*SA C/A pins

RFU RFU VDDQ VDDQ CK* CK / VSSA
J
K VDE|A A10 A2 AO A4 A5 VDDA
“
yss 0025s \A11 D9^ VSS^b 、VSS
/
VDD DQ24 DQ27- DQ19 DQ16 加
VDDQ DQ26 DM3 VDDQ\ VDDQ DM2 DQ15 VDDp

P VSSQ /VDQ28 RDQ28 VSSQ VSSQ RDQ22 WDO22 VSSC

VDDQ DQ28 DQ29 VDDQ/ VDDQ DQ21 DQ20 VDC)Q

T VSSQ DQ30 DQ31 VS^Q VSSQ DQ23 DQ22 V^SQ

/
V VDO VSS/ /SEN VSS VDP^ VDDQ

Byte group Byte group

Figure 8.10 GDDR3 pinout diagram.

Additionally, Read data path components sit outside the two pad rows
allowing the overall byte group to be even more compact since alt of the cir­
cuits associated with that byte group are in close proximity to the array and
the array data path. It should be apparent by now that cycle time and speed
are improved by keeping the layout very compact and the associated routes
as short as possible. For a GDDR3 device, as shown in Figure 8.9, the
highly optimized placement of bond pads, array, array data path, and Read
data path produce very short cycle times and excellent overall performance.

8.2.3 Architectural Limiters: Latency

The remaining performance characteristic to be discussed in this chap­
ter is latency. Read latency is fairly sensitive to changes in die architecture.
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 187

Most of these changes can be sorted into one of two cost classifications:
minor and major. Minor changes or low-hanging fruit are simple design
changes and optimizations that result in minor improvements to Read
latency. Major changes basically trade off die size for significant improve­
ments in Read latency. Lefs discuss each of these cost classifications in
turn.
Latency is a by-product of delay accumulation that can be broadly clas­
sified as gate delay, analog delay, or synchronization delay. While a large
number of circuits are called into action by a DRAM Read operation, only a
portion of these circuits actually impact Read latency. These particular cir­
cuits form the Read latency critical timing path. All three delay types are
usually present in this critical timing path. However, the critical timing path
can actually change to a new path as individual circuits are tuned to mini­
mize delay. This result stems from the fact that Read latency is detennined
by the longest-not the shortest-electrical path involved in a Read opera­
tion.
For the sake of discussion, we will define the Read latency critical cir­
cuit path for an exemplary device as comprised of C/A receivers; C/A cap­
ture circuits; a command decoder; a column address path; column
redundancy; a column decoder; an array data path, including HFF; a Read
data path, including pipeline and serialization circuits; and an output driver.
This list of circuits is foirly representative of a critical timing path for most
SDRAM designs in production and is characterized by gate delays, synchro­
nization points, and various wire and analog delays. Of these delays, analog
delays and the synchronization points are the most important.
Synchronization (sync) points are unavoidable in an SDRAM Read
latency control path. This is due to the fact that latency is defined in clock
cycles, not in nanoseconds. So, at the very least, the final stage must be syn­
chronized to the clock. Additional sync points can help overcome timing
uncertainties and assist with latency management. Sync points can hide sig­
nificant timing variability. However these results come at a cost. Each sync
point adds a clock cycle to Read latency. At shorter t(jK (higher frequency),
this may be a small concern, but as tcK pushes out at lower frequencies,
these clock cycles can translate into large latencies.
So, lefs return to the topic of Read latency and minor design changes.
The first consideration when designing the critical timing path is striking the
proper balance of delay types. Better performance is normally achieved by
minimizing the number of sync points to only those required to ensure
deterministic latency. Additional sync points generally contribute to higher
tcK dependency. The real goal is to produce the lowest Read latency possi­
ble, independent of frequency. The memory array and array data path are
highly asynchronous. The timing of events within the array data path is gen­
188 Chap. 8 High-Speed Die Architectures

erally controlled by either timing models or delay circuits. Timing models

attempt to mimic the behavior of actual circuits or networks to safely pre­
dict when to trigger subsequent events. Their timing is designed to track any
timing changes inherent in the circuit that is being modeled. While these
models are important fbr accurately timing events within asynchronous
blocks, they nevertheless introduce timing variability into the overall Read
latency critical path. A pipeline circuit is placed downstream of these cir­
cuits to absorb this variability and ensure deterministic latency. The output
of the pipeline feeds into a parallel-to-serial converter to ultimately syn­
chronize the data to the Read clock.
In addition to balancing delay types within the critical timing path, it is
important to keep the entire critical timing path tracking well over PVT. For
instance, if two different signals are designed for a specific relationship at
one operating comer, then they should maintain that relationship, as much
as possible, at all other operating comers. If this timing relationship changes
across comers, then the timing path sacrifices both timing margin and per­
formance. As stated before, the best way to maintain timing relationships
across comers is through careful design, with an emphasis placed on match­
ing topology, loading, and circuitry.
Larger, more significant improvements to latency are still possible.
However, these improvements are not free, being typically paid for with
increased die size. An example of this is evident in reduced latency DRAMs
(RLDRAMs). Figure 8.11 shows a typical Read operation for an RLDRAM.
A row activation cycle with Read and Precharge takes less than 15 ns from
start to finish, as shown. These devices are built to serve applications that
previously resorted to using SRAM. RLDRAMs have a traditional SRAM
broadside address interface but are otherwise DRAMs built fbr low cycle
times and low Read latency. Given the broadside addressing, however,
RLDRAM Read latency is essentially the concatenation of both row and
column access delays. To achieve low latency, both pieces of this must be
thoughtfully constructed with minimal delay in mind.
Sec. 8.2 Architectural Features: Bandwidth, Latency, and Cycle Time 189

Figure 8.11 RLDRAM Read operation wavefonns.

Figure 8.12 is an RLDRAM die photo exemplifying a memory architec­

ture that is capable of both low cycle times and low latency The most
important feature of this architecture is the use of very small memory subar­
rays within larger memory mattes. A small block of subarrays is serviced
along one edge by a local row decoder and along an orthogonal edge by a
local column decoder. Both the wordlines and column select lines are kept
short to keep rise and fall times short. For the device shown in Figure 8.12,
each wordline segment crosses only 512 addressable digit pairs, while each
digitline crosses only 256 addressable wordlines. Thus each subarray is only
128Kbits (131.072Kbits actual). The global column select lines span eight
subarrays or 2,048 wordlines. This is significantly fewer than the 8,192
wordlines crossed by a column select line in the GDDR3 part of Figure
8.1b. Therefore, the length of the column select line is much shorter, which
helps reduce RLDRAM cycle times and latency.
190 Chap. 8 High-Speed Die Architectures

Figure 8.12 RLDRAM die photo.

It should be readily apparent that smaller subarrays are achieved by

subdividing the memory array to a higher degree. Additional subdivision, in
turn, requires more sense amplifier strips, larger row decoder strips, and
more die area for additional column decoder blocks. So the benefit of better
performance is paid for with added die size and its corresponding higher
cost. In fact, the memory array and associated circuits are approximately
40-80% larger on RLDRAM than on a comparable DDR2 device. This is a
necessary and reasonable trade-off for RLDRAM. In this instance, new
market opportunities were created. However, if these opportunities aren't
accompanied by a higher average selling price (ASP), then the device and
its place in the market will evaporate.

8.3 CONCLUSION: DESIGNING FOR HIGH PERFORMANCE

In this chapter, we briefly explored how DRAM architecture impacts device
performance. Architectural features, both large and small, can affect device
performance fbr better or worse. A handful of these features can be modi­
fied or tuned late in the design cycle to squeeze out additional performance.
Optimum performance cannot be achieved, however, without considering
Sec. 8.3 Conclusion: Designing for High Performance 191

perfdrmance/cost architectural trade-offs at the beginning of a design

project. Only then, before the schematics and layout take shape, can major
architectural changes be fully woven into the design fabric. After all, high
performance is no accident.
 Chapter
9

Input Circuit Paths

As part of the I/O interface, input paths manage incoming data and com­
mand and address (C/A) and store them for subsequent processing. Input
paths may include receivers, buffers, and latches fbr data capture and distri­
bution. In this chapter, we explore various receiver circuits.
For higher frequency operation, input paths also need clocks to latch
and retime received data and C/A. After retiming, any distortion and noise
introduced by the memory channels can be substantially removed before
propagating further down the path. We discuss two commonly used meth­
ods for retiming in high-speed DRAM: matched routing and input timing
adjustment. Both techniques involve clock distribution fbr multiple I/Os,
which are spread out over a long distance. Finally, to meet stringent timing
specifications and tight setup and hold at capture latches, current-mode
logic (CML) is introduced trading off better timing performance with power
and design complexity.

9.1 INTRODUCTION
For synchronous DRAM (SDRAM), clocks and data are transmitted
together with a fixed relationship: they are source synchronous (SS). This
essentially means that the clock edges used to launch transmitted data are
available at the receive end fbr capture. While double data rate (DDR)
SDRAM uses both clock edges to latch data, C/A uses only a single edge
(single data rate or SDR). The relationships between the clock, data, C/A,
and reference voltage (Vref) for DDR SDRAM are shown in Figure 9.1
[1-3]. Because data is running twice as fast as C/A, additional clocks called
data strobes can be included in the I/O interface and dedicated fbr high­
speed data capture. In other words, strobes can be treated as a special type
of data as they are bundled and tightly linked (source synchronous) with

193
194 Chap. 9 Input Circuit Paths

data. Strobes save power because they are active only in the presence of
data, unlike clocks, which are active continuously. Furthermore, strobes,
either single-ended or differential, provide predictable timing relationships
when synchronized with system clocks (CLK/CLK*). High-speed data cap­
ture is more reliable with dedicated strobes than with non-source-synchro-
nous system clocks.
Clock and data alignment are also critical fbr data capture. Clocks and
strobes (DQS) are usually center aligned with incoming C/A and data,
respectively. Center alignment creates the largest timing margin fbr the cap­
ture latches. Timing parameters, such as data setup (r§) and hold (勿)are
illustrated in Figure 9.1 along with input voltage margins (匕小心・ As data
rates continue to increase (from 133 Mb/s for SDR to 1,600 Mb/s fbr
DDR3) and supply voltages scale down (from 3.3 V fbr SDR to 1.5 V for
DDR3), the valid data windows (or data “eyes")shrink due to clock jitter,
duty-cycle distortion, inter-symbol interference (ISI), simultaneous switch­
ing noise (SSN), impedance mismatch, reference noise, and supply noise, to
name a few. To address these signal integrity (SI) issues, we investigate sev­
eral circuit techniques in Chapter 10, such as on-die termination (ODT) and
data bus inversion (DBI).

clk*-、

v/h(ac)min
v/h(dc)min
qf(dc)
VIL (DC) MAX
■ VIL (AC) MAX
-V
VSSQ
Figure 9.1 a) Setup and hold timing diagram fbr C/A.
Sec. 9.1 Introduction 195

Figure 9.1 b) Setup and hold timing diagram fbr data.

Timing uncertainty is another challenge fbr data capture and processing.

A clock path that runs between the clock (or strobe) pads and capture
latches may include a clock receiver, timing adjust circuits, and a clock dis­
tribution network (or clock tree). Data and C/A paths, on the other hand, can
be much shorter with only input buffers. With different electrical lengths,
the input paths are no longer source synchronous internally. Furthermore,
process, voltage, and temperature (PVT) variations introduce excessive jit­
ter, and even post-production trimming may not be sufficient as the timing
margin shrinks. One way to retain the clock and data relationship is by
matched routing all the way from the pads to the capture points. On-chip,
closed-loop synchronization or input path “de-skewing” is another
approach. For instance, data training, either slaved to the host memory con­
troller or conducted inside the DRAM itself, can greatly improve the cap­
ture timing and ensure reliable data capture over PVT variations. Timing
alignment methods are addressed in Section 9.3 and Section 9.5.
Besides increasing I/O speed, prefetching multiple bits of data can
improve memory throughput significantly without sacrificing internal array
cycle time. A 4n- or 8n-Prefetch widens the internal bus compared to the
external data bus. SIPO (series in parallel out) circuits and pipelines can
cooperate with input circuits to tailor the incoming data into the internal
buses based on the decoded commands and programmable Write latency.
Control logic fbr the internal data path is discussed in Chapter 12. This
chapter starts with input receiver design.
196 Chap. 9 Input Circuit Paths

9.2 INPUT RECEIVERS

Due to pin limitations in DRAM, data buses are usually configured bi-direc­
tionally, while command and address (C/A) buses are unidirectional. Data
and C/A signals are single-ended with reference to a common voltage
(Vref)* Clocks and/or strobes, on the other hand, come in pairs for differen­
tial signaling and are more immune to common mode and coupling noise.
Input signals like clock, data, and C/A are first captured by receivers or
buffers. Two primary criteria should be met when designing the input
receivers: 1) input signals must convert from external to internal signaling
levels and 2) information presented in the original signals must be main­
tained.
Input signal imperfections make signal conversion difficult. Figure 9.2
shows delay variations based on Figure 9.2a signaling levels, Figure 9.2b
slew rate, and Figure 9.2c common-mode range (eg, center terminated ver­
sus Vdd terminated) and reference voltage. As a consequence, timing jitter
can occur, which greatly impacts data capture. One way to minimize timing
variation is to retain the same input receiver characteristics for data, C/A,
and the clocks, which is implied in the second design criterion. The other
way to minimize delay variations and extend the effective data eye is by
using 以底/ calibration.
Different voltage swing

Figure 9.2 Input receiver delay variations based on a) signaling levels and b) slew rate.
Sec. 9.2 Input Receivers 197

Figure 9.2 c) Input receiver delay variations

under different common-mode range or reference voltage.

Several input buffers or receivers based upon differential amplifiers

(diff-amps) and self-bias techniques have been shown in Figures 5.3, 5.4,
and 5.5. Other examples of data receivers employing different current
sources, voltage clamps, and input pairs are given in Figure 9.3. The trip­
point of the second-stage amplifier (inverters in Figure 9.3a) becomes criti­
cal because it needs to complete the conversion of the input signal to inter­
nal CMOS levels. DifFerent rise and fall times reduce the valid data window.
The high-speed comparator shown in Figure 9.3b can be used as a receiver
having a pre-amplifier (current-mode logic or CML) and decision circuit.
This receiver has less propagation delay and is less sensitive to PVT varia­
tions. Although designed for fully differential signaling, one of the two
inputs for this differential receiver comes from a reference voltage (嗓e尸).
This is commonly referred to as pseudo-differential signaling. To achieve
the best input timing, Vref can be trained on-die or externally by the mem­
ory controller.
198 Chap. 9 Input Circuit Paths

* Data

a)

Figure 93 a) Differential receivers and

b) comparator with differential CML as input receivers.

An important concern with low-voltage design is the need for an

improved common-mode range. Two complementary diff-amp stages can
be placed in parallel to extend the usable operating range fbr the composite
Sec. 9.2 Input Receivers 199

receiver, as shown in Figure 9.4. Notice that in order to get a better voltage
margin, not all of the devices are biased in the saturation region. For folly
differential clock receivers, the reference input (Vref) is replaced by the
complementary clock input (CLK*).
%
--------------------- d-------------------------------

Figure 9.4 Wide-swing difi-amps fbr input receiver.

200 Chap. 9 Input Circuit Paths

Other considerations for input receiver design include electrostatic dis­

charge (ESD), PVT sensitivity, capacitive loading, power, and layout area.
Between input devices and pads, a thick oxide pass-gate with a higher gate
voltage (Vccp in Figure 10.23) can provide protection for the thin gate
oxide present in the receiver transistors. For device protection, voltage
clamps are also used. Low-pin capacitance is crucial to achieving higher
data rate. Unfortunately, higher gain and lower offset generally require large
input devices. Power consumption, slew rate, and the overall complexity of
the receiver design must also be considered.

9.3 MATCHED ROUTING

As mentioned, maintaining the original timing relationship (i.e., source-syn­
chronous between clock and data) is important for data capture. Matched
routing is the most effective solution for establishing and preserving data
capture timing. The quality or property most important for matched routing
is delay. Delay mismatch produces timing skew that can negatively impact
input or output timing specifications. Sensitivities introduced by process,
voltage, and temperature (PVT) variations are highly undesirable, too, and
can impact yield and potential tcK migration. Sensitivity to PVT can be
reduced or mitigated by eliminating all sources of potential mismatch-in
essence, every single element of the matched routes must be “exactly identi­
cal.Being exactly identical necessitates that the following elements match:
• Conductor width, spacing, and length fbr every segment of the routes.
Each route, including the conductor layer, must match segment-for-
segment.
• Driver topology and size.
• Load topology and size.
• Shielding.
• Route topology.
• Quantity of bends and their location.
A simple point-to-point routing topology is shown in Figure 9.5. A path
with two identical inverters and metal 1 (Ml) and metal2 (M2) routes is
drawn in both schematic and layout views. A "matched” path is placed side-
by-side fbr comparison.
Sec. 9.3 Matched Routing 201

Metal
length/width

M2 M1
100pm/4pm 1000pm/2pm 100pm/4pm

100pm/4pm 1000pm/2pm 100pm/4pm

Inverters Inverter Inverter

Figure 9.5 Point-to-point matched routings.

The placement (locality) of input paths can greatly impact the length
and shape of the route and should be floor-planned in the early design
phases. Shielding further improves matching by eliminating in-phase or out-
of^phase coupling. However, shielding also increases the overall layout area
and capacitive loading.
Three routing styles, with and without shielding, are seen in Figure 9.6:

1. Size maze: Efficient for adding length to horizontal routes. (Style can
be rotated fbr vertical routes.)
2. Trombone: For adding length to vertical routes.
3. Accordion: For adding length, but requires excessive comers.
202 Chap. 9 Input Circuit Paths

1_____ u
C)

Figure 9.6 Matched routing styles: a) size maze, b) trombone, and c) accordion.

Matching the clock/strobe path to the data path is another challenge.

Instead of point-to-point, the clock path must branch or tree out to multiple
locations. Routing distances are longer for the clock path than for the data
path. Extra buffer or repeater stages are required to drive the clock over a
long distance with a minimum propagation delay. Loading difference intro­
duces another delay variation. A tree-typed clock distribution network
(CDN) may be used for matched routing. An example for data and clock
matching is shown in Figure 9.7, with both point-to-point and tree-typed
routes.
Sec. 9.4 Capture Latches 203

In order to properly size clock drivers, logical effort [4] can be evalu­
ated based on logic style, loading, and propagation delay. Extra buffer
stages and routes are added on each data or C/A path to match the
clock/strobe path. Input path delay, denoted d in Figure 9.7, represents the
delay from the input pins to the capture latches, including the receiver,
buffer, and matched path delays. One measurement for the quality of
matched routing is the difference of £加・ For well-matched input paths, less
than 10 ps timing skew can be achieved across simulation comers with lay­
out back-annotation. Less than 10 ps skew across the data bus was con­
firmed in silicon for a GDDR3 part. However, when matched routing is
applied, a trade-off must be made between capture timing and the power
and layout area consumed by the routes.
From a layout point of view, matched routing is rarely accomplished by
automated tools due to its complexity and restrictions. Before preparing full
custom layout, a preliminary-matched route design must be budgeted and
floor-planned. Basic route topologies, routing distances, materials, dimen­
sions, buffers, and loads must be determined and tuned to obtain the desired
signal quality. The initial design phase relies on using parasitic resistor and
capacitor (PRC) models to estimate routing parasitics, which enable dimen­
sional and electrical tuning of the routes. Pad order and pad pitch also have
a direct impact on the route cells. An automated tool is helpful fbr checking
the final quality of matched routing by comparing conductors, topologies,
bends, and sizes. SPICE back-annotated simulation provides timing verifi­
cation at the final design phase.

9.4 CAPTURE LATCHES

The next logical step after matched routing is data capture. A capture latch
acts like a storage element controlled by clock edges. High-speed data
requires two capture latches: one fbr rising clock edges; the other, falling
edges. C/A typically needs only one latch in response to only rising clock
edges. After the capture latch, data is split into two parts, running at half the
input data rate fbr each part. C/A also gets extended to a full cycle (tcid for
a valid window. Figure 9.8 shows a capture timing diagram, where Iqq is
the clock to Q delay of the capture latch, and % is the input path delay,
which includes the receiver and matched routing.
204 Chap. 9 Input Circuit Paths

DQ C/A (pin)

CLK (pin)
A
-----A
J \
DQ C/A (latch)

CLK (latch)

Captured DQ/C/A

Figure 9.8 Timing diagram for input capture latches.

The implementation of a capture latch is straightforward if the clock

and data relationship (center aligned in this case) can be preserved. A static
edge-triggered flip-flop (FF) is shown in Figure 9.9, with master and slave
latches and transmission gates (TG). Complementary clocks sample the
inputs and control the latches. Set-up time for this FF is in the order of hun­
dred picoseconds, depending on PVT comers.

Figure 9.9 A static edge-triggered flip-flop as a capture latch.

A dynamic latch with tens of picoseconds set-up time is shown in Fig­

ure 9.10. Compared to a TG and latch in the static FF, loading for the input
data is only a transistor. For it to switch, the input need only overcome the
threshold voltage 0fMp1, making it significantly faster compared to static
FF. Sampling clocks, CLKS1/2. have much less loading compared to static
FF. CLKS2 is one gate delay version of CLKS1 to reduce overlapping cur­
rent. A keeper circuit at the output tries to hold the value during the Pre­
charge phase (clock is HIGH). This domino-typed dynamic FF is useful for
high-speed capture latches.
Sec. 9.4 Capture Latches 205

For high-speed data capture, sense-amplifier latches can also be

applied, as shown in Figure 9.11. With cross-coupled devices, they are fast
and have high gain. They can sense small voltage differences at the inputs
in response to clock edges, which makes them ideal even as input receivers.
A second stage, like the one shown in Figure 9.10, is necessary to store and
convert the sensed value. To reduce offset and ensure proper sense function,
careful layout of the input pairs is imperative.

Figure 9.11 Sense amplifiers as front stage of capture latches.

206 Chap. 9 Input Circuit Paths

9.5 INPUT TIMING ADJUSTMENTS

Write data is received and captured by the Write strobe (WDQS)y although
the data timing essentially remains in the WDQS domain. The Write com­
mand is decoded via command and address (C/A) paths, which are timed by
and remain in the system clock (CLK/CLK*) domain. Therefore, two clock
domains exist: one fbr the Write command and the other for Write data. The
timing difference between system clocks and data strobes at pins is defined
as S°ssand is tightly controlled by the system to limit the timing variation
to less than 0.25 依,酎 example. Internal mismatch between the data path
and the Write command path makes this timing uncertainty even greater
under PVT variations. Timing adjust circuits, like delay-locked loop (DLL)
or phase-locked loop (PLL), can be useful to reduce timing uncertainty,
cross clock domains, and ensure deterministic Write latency. Block dia­
grams with and without C/A DLL are shown in Figure 9.12a and
Figure 9.12b, respectively.

Figure 9.12 Input clock domain crossing with a) C/A DLL.

Sec. 9.5 Input Timing Adjustments 207

Figure 9.12 Input clock domain crossing with b) matched routing.

As shown in Figure 9.12b, a clock path that includes a clock divider and
Write clock tree, is not strictly matched to the input paths. The Write clock
tree is routed to both the command and data paths, which makes it longer
than the input matched network. However, the Write command must be
placed precisely in order to hit the setup and hold window in the Write pipe­
line sequencer, which is controlled by the Write strobe. So, clock domain
crossing and timing adjustment may require extra buffers and delays in the
input matched paths, which increase latency, power, area, and sensitivity to
PVT variations. Data training can provide better alignment between strobes
and clocks during initialization. This open-loop ”de-skewing” exercise con­
ducted by the memory controller can help meet Sqss，Sss，and 如“ speci­
fications. This is referred to as Write leveling.
Instead of training or matching, an on-die timing adjustment circuit
(ie, DLL) can be used. With the input C/A DLL shown in Figure 9.12a, the
Write clock tree delay can be hidden by synchronizing CACLK with the out­
puts (CAFB) of the clock tree. Capture latch delay can also be modeled in
the feedback path. The C/A DLL provides flexible tuning over PVT varia­
tions and permits on-die Write leveling. However, the DLL also adds some
timing uncertainty compared to strictly matched routing. There are also
concerns about extra power, area, and "wake-up" time.
208 Chap. 9 Input Circuit Paths

As signaling speeds continue to increase, the corresponding valid data

window shrinks due to intersymbol interference (ISI), random jitter (RJ),
SSO, duty-cycle distortion, supply noise, and cross talk. The eye diagrams
shown in Figure 9.13 compare an ideal case (Figure 9.13a) running at 5
Gbps to a more realistic system (Figure 9.13b) with all of the additive noise
and jitter. Distributing data and clock at such high speeds on-die is a chal­
lenge. Matched routing may not be suitable for data rates greater than
3Gbps. The clock distribution must be redesigned to minimize jitter as well
as clocking power. One way to relieve this problem is to divide the clock
down, for example, so that it runs at a quarter of the data rate. Another prac­
tical approach is to recover the dock on-die based upon the data stream.
These circuits are referred to as clock recovery circuits (CRC) in most com­
munication systems [5]. Since clocks recovered by CRC are not source syn­
chronous with incoming data, it is critical to generate lowjitter local clocks,
which usually require a phase-locked loop (PLL). To provide enough transi­
tions so as to avoid "frequency wandering" by PLL oscillators, data scram­
ble is mandatory. However, the added complexity and cost make it
unattractive for a memory interface.

1.2
1.15
1.1
1.05
1
s 0.95

I 0.9
0.85
0.8
0.75
07
0.65
0.6
0 20 40 60 80 100 120 150 160 180 200
Picoseconds

Figure 9.13 Eye diagrams fbr 5Gbps received data signals:

a) ideal eye with ISL
Sec. 9.5 Input Timing Adjustments 209

1.2
1.15
1.1
1.05
1
s
:
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0 20 40 60 80 100 120 150 160 180 200
Picoseconds

Figure 9.13 Eye diagrams for 5Gbps received data signals:

b) ideal eye with ISI, RJ, SSO, DCD, and clock jitter.

A compromise of source synchronous and CRC systems is shown in

Figure 9.14. A continuous clock is still sent source synchronous with data
(DQ) from the memory controller (TX), although they may run at different
rates. To reduce routing and pin capacitance, data receiver or latch, in this
case, is placed close to the DQ pad. Clocks at the capture latch are re­
aligned by a DLL/PLL with reference to the external clock. Quarter-rate
clocks can simplify the design of the CDN, whereas internal multi-phase
clocks still keep up with the data rate or unit interval (UI), A set of sense
amplifiers shown in Figure 9.11 can function as an input receiver (RX) sam­
pled by internal multi-phase clocks. The external clock can be edge aligned
with the data, while the data eye is centered on internally distributed clocks
during initialization or training. A timing diagram drawn in Figure 9.15
describes the relationship of the clocks.

Figure 9.14 A PLL/DLL-based transceiver with clock alignment.

210 Chap. 9 Input Circuit Paths

3
I I ]« I __ —] — — — — 1 — — 一》I

Figure 9.15 Timing diagram fbr input clock alignment.

In Figure 9.15, DQ represents incoming data at pins, while CLK and

CLK* are differential source synchronous clocks running at one quarter of
the data rate. One clock cycle «ck) equals four unit intervals (Uis). CLKO,
CLK90, CLK180, and CLK210 represent internal, four-phase clocks, one UI
apart from each other. Referred to the CLK/CLK*, CLKO is delayed by t^ugn
due to the input PLL/DLL and the clock trees. Although only a 2-67 delay
is shown for tAlign in the timing diagram, in reality more than 6-UI delay
should be considered for the overall timing budget, depending on the data
rate and circuit topologies. The。物 parameter can be defined as

Ulign = tRx + Udjfee =…UI

(9.1)
where 垢 is the delay fbr clock receiver;。力，fbr PLL/DLL delay; and tTree,
fbr CDN. The receive clock path delay can also be rounded up in terms of
unit intervals (UI). Therefore, the clock edge used to capture the data is n-
bit-shifted compared to the edge that launches it at the transmit site (TX).
When the clock and data are out of phase, such n，UI delay amplifies jitter at
certain noise frequencies. To reduce the impact of the n・UI input path delay,
clock path circuits should be insensitive to PVT variations and efforts
should be made to minimize。恸.Step size and adjustment range also steer
the timing circuit design toward an analog approach.
One problem associated with traditional CMOS logic is its sensitivity to
PVT variations, especially to variations in process and supply voltages. A
plot of delay versus PVT comers fbr two standard NAND gates is shown in
Figure 9.16. The delay can vary from 103 ps (fast process-FF, 1.5 V, and
Sec. 9.5 Input Timing Adjustments 211

0℃) to 268 ps (slow process-SS, 1.2 V, and 120℃), which is a 160%

increase. If these CMOS buffers were used in the critical timing path, like
digital DLL, or clock trees, the accumulated jitter would be so great that
clock alignment would be impossible.
Another problem with CMOS buffers is duty-cycle distortion and cross
talk, which can severely impact signal quality. Common-mode noise
injected from the reference voltage, power supplies, and substrate intro­
duces dynamic jitter and requires careful design for power bussing, shield­
ing, and isolation. The trade-offs for better signal quality at higher speeds,
though, are excessive power and area. An alternative logic style must be
investigated to mitigate these problems.

1.2 1.25 1.3 1.35 1.4 1,45 1.5

%(V)

Figure 9.16 CMOS buffer delay versus PVT comers.

212 Chap. 9 Input Circuit Paths

9.6 CURRENT MODE LOGIC (CML)

Current mode logic (CML), shown in Figure 9.17, is also referred to as
MOS CML (MCML). A fully differential CML buffer in Figure 9.17a is
basically a differential pair with resistive loads (Rd) and constant bias cur­
rent (Ifau). Differential signaling makes CML much more immune to com­
mon-mode noise. Instead of swinging from rail-to-rail, CML outputs swing
(in Figure 9.17a) only from 匕m t0，O£)- Rd • hail- The small swing not
only makes CML fast, but also reduces electromagnetic interference (EMI)
and cross talk.

Figure 9.17 Current-mode logic: a) CML buffer and its I-V characteristics.
Sec. 9.6 Current Mode Logic (CML) 213

Figure 9.17 Current-mode logic: b) MOS current-mode logic.

With clock frequencies higher than 2 GHz, the power consumption of

CMOS logic becomes unmanageable, while the power consumption of
CML remains constant across the entire frequency range [5]. Delay (D),
power (P), and energy (E) products between CML and CMOS are tabulated
in Table 9.1 [5], where N is the logic depth of the critical path, I is the con­
stant tail current for CML, k is the transconductance of MOS devices
(/iCoxW/L), C is the load capacitance, A V is the output swing, Vt is the
threshold voltage, and a is a parameter approximately 1.3.
It is interesting to see that the propagation delay of CML logic has no
relationship with supply voltage (%)). If bias current (7) and voltage swing
are kept constant over PVT comers, a constant delay can be achieved,
which keeps PVT sensitivity at a minimum. Now, as an example, well
explore the design of a clock distribution network (CDN) using both CML
and CMOS buffers.
CDN is very important fbr I/O circuit design. Its primary goal is to
maintain the same propagation delay across all destinations under PVT vari­
ations. With the input clock alignment depicted in Figure 9.14, a CDN is
implied fbr distributing multi-phase clocks to various DQ and C/A latches.
Minimizing delay and sensitivity through the tree is essential fbr this
scheme to work.
214 Chap. 9 Input Circuit Paths

Table 9.1 Delay, power, and energy products comparison.

CMOS CML

D 二 NxCxVdd
D = NRC = ------ y------

P = NxCx^dd^ P = NxIxVdd

P D = Nx/x Fdd P D = /xCx/Px Poo

N3xC2xVddxA^
ED = 7V2x2x^-x
ED -----------------------------
k公-匕 I

A simulation setup for both CMOS and CML clock trees is shown in
Figure 9.18. Two buffers drive two parasitic RC models (PRC) for metal
routes, each with a distance of 1,500 pm, a width of 0.5 gm, and a space of
1 |nm in an 80 nm DRAM process. A resistor is used fbr a CML load with
improved bandwidth. The CML buffer is biased to get a constant swing of
approximately 400 mV at 1.35 Vdd- For CMOS inverters, the width of
PMOS is 50 jim and for NMOS, 25Pm. A minimum length is applied to all
devices except the tail NMOS (MN3 in Figure 9.17a) fbr the CML buffer,
which has twice the minimum length. The input clocks (CLK/^ and
CLKinb) are differential and run at 1 GHz. Rise and fall times for the CML
and CMOS buffers are 135 ps and 108 ps, respectively, to maintain reason­
able duty cycle and signal quality. The simulation is set up at a typical pro­
cess with PRC, L35 V, and 95℃ comers.

Figure 9.18 Simulation setup fbr CML (top) and CMOS (bottom) clock trees.
Sec. 9.6 Current Mode Logic (CML) 215

Simulation results for voltage and process sensitivities are shown in

Figure 9.19, where the difference of CMOS propagation delay (Zp) is six
times greater than it is for CML buffers. The worst-case absolute delay is
also halved for the CML tree. In Figure 9.20, temperature sensitivity is
comparable between the two designs with approximately 26 ps delay differ­
ence per 100℃.

2u u
(sd)

QJ 17 5
A
g1
a
『q

Figure 9.19 a) CML and b) CMOS clock trees delay time

under process and voltage variations @ 95 ℃.
216 Chap. 9 Input Circuit Paths

Temperature (℃)
b)
Figure 9.20 a) CML and b) CMOS clock trees
under process and temperature variations @ 1.35 V

Indeed, the CML clock tree is immune to voltage and process varia­
tions, as predicted in Table 9.1. Near-constant delay makes the CML buffer
ideal for designing a critical timing path, especially when using a clock
alignment scheme. (Timing circuits using the CML buffer are discussed in
Chapter 11.) The CML buffer is also a natural fit for clock receiver circuits,
as shown in Figure 9.3b. Extensive use of CML logic makes input circuits
capable of handling much higher data rates and more aggressive scaling.
So, what are the obstacles preventing designers from using CMLs in
DRAM timing paths as opposed to using simple CMOS inverters?

1. A more complex circuit topology and differential signaling require a

larger area and more restricted matching.
Sec. 9.6 Current Mode Logic (CML) 217

2. Dedicated bias generation and distribution is essential to maintain a

constant delay over PVT variations. A low-voltage bandgap circuit
can generally fit in this role.
3. A signal conversion circuit (differential-to-single-ended translator) is
needed to interface with CMOS logic. Duty-cycle distortion and
PVT sensitivity may be exposed at this stage. Multi-phase clock gen­
eration or duty-cycle correction is inevitable and complicates the
design.
4. Increased static power is another concern. A CML buffer consumes
current regardless of whether it is switching or not. For the example
described, the average current consumed by the CML tree is approx­
imately 6.8 mA, compared to 2.2 mA for a CMOS tree. On the other
hand, the current profile for a CML tree is fairly constant with small
spikes, which translates to less switching noise and quieter buses.

Looking forward, as speed in memory systems continuously acceler­

ates, memory interface design becomes formidable fbr timing closure.
Teamwork is necessary fbr analyzing signal integrity, packaging, transmis­
sion line, power, and clock delivery. The critical timing path must be identi­
fied and modeled early in the design phase followed closely by a timing
analysis. Performance and cost must be balanced when designing high­
speed I/O circuits.

REFERENCES

[1] Micron technical note, TN-47-06,4tUpdated JEDEC DDR2 specifications.”

[2] Micron technical note, TN-47-02, <4DDR2 offers new features/functionality
introduction?,
[3] Micron technical note, TN-47-11,4<DDR2 differential DQs introduction.”
[4] I. E. Sutherland, R. F. Sproull, and D. Harris, Logical Effort: Designing Fast
CMOS Circuits. San Diego, CA: Academic Press, 1999.
[5] B. Razavi, Monolithic Phase-locked Loops and Clock Recovery Circuits,
Theory and Design. Piscataway, NJ: IEEE Press, 1996.
[6] A. Tanabe, M. Umetani, I. Fujiwara, K. Kataoka, M. Okihara, H. Sakuraba, T.
Endoh, and F. Masuoka, u0.18-|im CMOS 10-Gb/s multiplexer/ demultiplexer
ICs using current mode logic with tolerance to threshold voltage fluctuation/,
IEEE Journal ofSolid-State Circuits, vol. 36, pp. 988-996, June 2001.
[7] M. Mizuno, M. Yamashina, K. Furuta, H. Igura, H. Abiko, K. Okabe, A. Ono,
and H. Yamada, “A GHz MOS adaptive pipeline technique using MOS
current-mode logic," IEEE Journal of Solid-State Circuitsi vol. 31, pp.
784-791, June 1996.
 Chapter
10

Output Circuit Paths

Serving as the driving engine of the memory data path, the output path mul­
tiplexes the internal data and sends it out in a timely manner, based on CAS
latency and burst length. Timing control, which is addressed in Chapters 11
and 12, is critical to determining when to send data out after receiving a
Read command. Besides data, strobes similar to those found on the input
paths, can be sent out source-synchronously (SS) with data to the memory
controller (MC). SS strobes can greatly simplify data capture at the control­
ler site. And, strobes can be bi-directional like data or unidirectional like
clocks, in which separate Read and Write strobes are needed. The latter can
improve the Read/Write turn-around time, albeit at a higher pin count.
Trade-ofls between performance and cost must be considered when design­
ing the I/O interface.
As I/O speeds push toward and beyond the gigahertz range, signal
integrity (SI) becomes increasingly important to I/O design. Transmission
line, bus structure, termination strategy, load capacitance, inductance, and
packaging technology all directly or indirectly impact I/O design. To meet
design targets and I/O specifications, a variety of issues, such as impedance
calibration, simultaneous-switching noise (SSN) reduction, slew-rate con­
trol, signal return path analysis, and electrostatic discharge (ESD), must be
examined. In this chapter, we start by exploring these issues and their impli­
cations on design. Then we move to examine high-speed, parallel-to-serial
conversion, and the interface between the internal data path and output driv­
ers. Finally, we conclude by describing some features of emerging memory
I/O.

219
220 Chap. 10 Output Circuit Paths

10.1 TRANSMISSION LINE, IMPEDANCE, AND TERMINATION

Signals can be carried via wires, coaxial cables, or microstrip or stripline
transmission lines. For signals transmitted on board at higher frequencies,
PCB (printed-circuit board) traces can no longer be treated as just wires:
they become radio frequency (RF) transmission lines. Signal integrity can
be affected in two ways. The first is transmission line (TL) effects, such as
ringing and reflections. The second is the interaction with other elements,
such as cross talk and noise on the power plane. Drive impedance control
and on-die termination (ODT) tend to reduce reflections, slew-rate control
can reduce ringing, and signal return-path analysis can help to reduce cross
talk and noise on the power planes (return path).
Let's start the discussion about signal integrity with basic transmission
line modeling and characterization. A simple model of a transmission line is
shown in Figure 10.1. The signal trace is on the top (shaded) with a return
path at bottom. The trace can be divided into unit length portions, each with
identical electrical characteristics.

Figure 10.1 Transmission line model.

The single-most important feature of transmission lines is characteris­

tic impedance Zc. Zc as derived from this model is given in Equation (10.1),
where R, Gt C, and L are the resistance, conductance, capacitance, and
inductance per unit length, respectively, and/is the signal frequency:

=加+J2ML =1R+j2矶
c NG+j2nfC N jlnfC

(10.1)

As shown, the characteristic impedance is frequency dependent and

specified in ohms. The conductance per unit length G is due to loss through
the dielectric material separating the two traces. For nearly all types of con­
ductors, it is safe to assume that G=0 at frequencies below the gigahertz
range. Higher frequencies introduce some AC conductance, which may not
Sec. 10.1 Transmission Line, Impedance, and Termination 221

be negligible [1]. The resistance R can also be assumed to be zero, except at

high frequencies where skin effect comes into play. Therefore, with
assumptions that both G and R are zero, the transmission line could be con­
sidered lossless, and the characteristic impedance, in this case Z。, would
become frequency independent.

lim Zc

(10.2)
The skin effect is a phenomenon in which resistance increases as fre­
quency increases due to the fact that the signal is pushed into a very thin
layer of the conductor surface. The thickness of this layer is inverse propor­
tional to the square root of frequency (f). The skin resistance 街J of the
transmission line is given in Equation (103) along with AC resistance
(H/C), where p is permeability, o is conductivity, and W and I are the width
and length of the conductor, respectively.

(10.3)
At the frequency that some of the point-to-point DDR systems are run­
ning, the skin depth is down to a few tenths of a mil (0.0002 inches) for the
conducting layer, which, in turn, increases the AC resistance. The skin
effect is the main reason that the resistance generally cannot be ignored in
calculating the characteristic impedance of a transmission line. The high
frequency loss due to skin effect can be compensated through equalization
to reduce inter-symbol interference (ISI).
When a transmission line is terminated at both ends with its characteris­
tic impedance (Z0), the line is said to be matched, as shown in Figure 10.2.
Being matched means that no reflections should occur on the line, which is
desirable for high-frequency signal transmission. Without termination and
impedance control, reflections on the line distort the signal, which, if the
distortion is large enough, prevents reliable detection by the receiver.
In Figure 10.2, at the driver side, the matched impedance (Z0) is equal
to the output driver impedance (e.g., 18 ohm) plus the termination imped­
ance (eg, 45 ohm). On-die termination (ODT) can provide a matched
impedance on the receiver end. An ODT pin is added to DDR2 parts, as
shown in Figure 10.3, to control the termination impedance as needed. For
DDR2, ODT is required only for the data, data mask, and strobes. For clock
222 Chap. 10 Output Circuit Paths

frequencies passing into the gigahertz range, as in GDDR4 parts, ODT is

also required for command and address signals.

{Z0 = 63G}
Driver R Series R T Line
Source)........... 节外一
18 Q 45 Q 63 Q

Figure 10.2 Matched transmission line (two-end termination).

DQ pins

ODT pin Control

Figure 10.3 Functional diagrams for on-die termination (ODT).

Examples of both Read and Write ODT configurations in a dual-ranked

module system are shown in Figure 10.4. Compared to the stub series termi­
nated logic (SSTL) bus of Figure 5.2, a bus with ODT demonstrates far bet­
ter signal quality fbr both Read and Write operations. Eye diagrams with
data running at 555Mbit/s fbr both configurations are shown in Figure 10.5.
Obviously, the ODT system produces increased voltage margin, larger data
eyes, and improved slew rate even under heavily loaded conditions.
Sec. 10.1 Transmission Line, Impedance, and Termination 223

Active module Standby module

50 Q
System
50 C
controller

a)
Active module Standby module

b)
Figure 10.4 Typical ODT configurations for DRAM (two modules populated)
5
a) Writes and b) Reads.
1
1.9
1
工

£
>)
4 > 13
)
①
— —

6S4 0)6
L
OA0 o,19
忠

~n00 07
oAlndu-

d£0 O.O
5
3
01 0 1
-0.1 -0.1
0 0.75 1.5 2,25 3 375 0 075 1.5 2,25 3 3,75
Time (nanoseconds) Time (nanoseconds)

a) Writes with SSTL (left) and ODT (right) terminations.

r5
>)rJ >)
①
6SL
「

OA0 O)6
5o
5d.97 >
u-0,6s
Indu-

40 075 1.5 2.25 3 3.75 ’ 0 0 75 1 5 2.25 3 3.75

Time (nanoseconds) Time (nanoseconds)
b) Reads with SSTL (left) and ODT (right) terminations.

Figure 10.5 Comparisons between SSTL and ODT for dual-ranked module.
224 Chap. 10 Output Circuit Paths

10.2 IMPEDANCE CONTROL

So, how do we control impedance in order to match the characteristic
impedance of the transmission line? An impedance calibration circuit is
generally the answer.
A typical calibration system using an external precision resistor
eg, 240 Q 土 1%） is shown in Figure 10.6. can be set to m times Zq to
improve accuracy and reduce current during calibration. For example, if Zq
=40 Q and R^q = 240 R m=6, the driver can be divided into six identical
legs, each calibrated to 240 Q and placed in parallel to achieve an equiva­
lent impedance of 40 Q This configuration is also helpful for ODT and
drive strength control.

ZQ calibration control
J
Digital filter Control logic < <SComp
<P>J SAR logic
P Code gen. VREF generator
dn,

35
一

Figure 10,6 Impedance calibrations using external resistor R^q.

On-die impedance calibration, also known as ZQ calibration, uses a

dedicated ZQ pin. During initialization, the calibration usually starts after
the system is powered up and the clock is stable. The resistance of one leg
of the output driver, ie, the pull-up driver in Figure 10.6, with a combina­
tion of transistors and resistors, compared to the external resistor Rzq. The
Rzq is placed between the ZQ pin and ground （七s。）. The transistors are
binary weighted to provide a certain resolution, ie, 5-10% of Z°, and a tun­
Sec. 10.2 Impedance Control 225

ing range depending on process, voltage, and temperature (PVT) variations.

In Figure 10.6, five-bit digital codes (Vq^ or Vqi) generated by ZQ control
circuits are used to select or deselect different pull-up or pull-down transis­
tors. A total of 32 steps are available with a quantization error of one LSB
(least significant bit). The resistor (Rpu/RpD)出 parallel with the transistors
improves the linearity of the eflfective impedance. Rpu / Rpd must be
greater than Rzq to allow for transistor tuning. Due to large current flow
through these drivers, wider metal and sufficient vias or contacts must be
used in the layout to avoid metal electromigration problems.
The first step of the calibration sequence starts with the signal Eni*
transitioning LOW. The voltage (嗫0) is tapped and fed in to the ZQ circuit
for comparison. An internal reference voltage (Vref) is generated and con­
nected to the other input of the comparator. A simple resistor divider can be
used as the Vref generator. Its voltage level can be adjusted if necessary.
For 50% Vddq calibration (嗓+尸二0.5 the calibrated pull-up
impedance (Rcal.up) is equal to Rzq with one LSB error (Rulsb)- The next
step is to calibrate the pull-down impedance by using another pull-up driver
and the generated Fq；/codes (Vqh* <0：4>). To start this step, signal En2*
transitions LOW, in this case. A voltage (匕tapped between the pull-up
and pull-down drivers is fed back to the ZQ circuit for comparison. The
final pull-up impedance (Rcal,UP)and pulkdown impedance (Rcal,DN)are
given in Equation (10.4) and Equation (10.5), respectively, with regard to
the driver multiplication factor m, pull-down LSB error Rdlsb, and external
resistance Rzq.

Real、up = (Rzq 土 1• Rulsb"171

(10.4)

Real, DN =Real, U尸 ±(1 • 掰=(RZQtRULSB土RDLSB)/

(10.5)

Compared to the pull-up impedance RcaltUP 出 Equation (10.4), the

pull-down impedance (Rcal,DN)may have much greater calibration error.
Given the previous examples, Rzq = 242.4 Q (+1%), zn = 6, and Rulsb =
Rdlsb = 6 0, then Rcal,UP = 414 Q, and Rcal,DN = 42.4 Q which have a
mismatch of 3.5% and 6%, respectively. The 6% error may not be accept­
able for a high-speed interface, especially for open-drain configurations
where only NMOS pull-down drivers are used. Improving the pull-up
impedance calibration can help to tighten up the specs. Instead of five bits,
one extra bit can be added for the pull-up transistors and the LSB reduced to
226 Chap. 10 Output Circuit Paths

3 Q, for example. Therefore, the calibrated pull-up and -down impedances

would be 40.9 Q and 41.9 Q respectively, with a maximum error of 4.75%.
Longer calibration time is one drawback of such an improvement.
Depending on the searching algorithm, the ZQ calibration could take
thousands of cycles during initialization. A binary search algorithm using
successive-approximation registers (SAR) is usually fast and well suited for
ZQ calibration. A diagram for a four-bit binary search algorithm is shown in
Figure 10.7, with the code initially set to “1000”(MSB, most significant
bit). A maximum of three steps are needed to complete the search, com­
pared to eight steps without it.

Figure 10.7 Diagram fbr binary search algorithm (4-bit).

Usually, an analog comparator can be used to discriminate the differ­

ence between Vzd匕erm and V^f. A large input device may be used for the
pre-amplifier to reduce offset and increase gain. A decision circuit with pos­
itive feedback and an output buffer can be used to improve the sensitivity
and generate digital control signals (i.e., “Increase" or "Decrease") fbr digi­
tal filters. Digital filters are used to ensure the right direction fbr adjustment
under noisy environments, An up/down counter can be used to log the deci­
sions made by the comparator*. After certain continuous “Increases” or
“Decreases,“ a true adjustment is made to the outputs, and the comparator is
reset fbr the next comparison. The initial calibration time Qcal,出。for either
a pull-up or pull-down impedance can be approximated as

tCalJni =("1)/F」SCK

(10.6)

iCal, Ei = " ,孙• bcK

(10 7)
Sec. 10.2 Impedance Control 227

where n is number of bits, xp is the digital filter factor for continuous

counts, and t^cK is the sampling clock period fbr the digital filter. The cali­
bration starts at midpoint. Equation (10.6) represents a SAR binary search
whereas Equation (10.7) is for counting step-by-step. The speed of the com­
parator can really slow down initial ZQ calibration at certain PVT comers.
Time-out and lock detect circuits must be included in the filter in case the
input voltage is close to Vref and the comparator foils to make a decision
within a predetermined period.
An alternative way to make a decision is by using a voltage-to-time
conversion circuit (VTC), as shown in Figure 10.8. The input voltage of the
VTC is converted to a current and charges or discharges an internal capaci­
tor. When the trip point of the output buffer is reached, the output of VTC is
switched to create a transition. When two identical VTCs are used, with one
hooked to Vref and the other to V^q or Vterm, the voltage difference can be
converted to a timing difference. A phase detector (PD) can then make the
judgment and send the results to the digital filter. Quick turn-around time
makes this attractive for some applications.

Figure 10.8 ZQ calibrations using digital comparator.

Due to voltage and/or temperature (VT) drift, on-die impedance may

require readjust fbr better matching after initial calibration. Such ZQ
updates may take a shorter time by changing only a couple of LSBs. This
event can be controlled by an up-down counter and triggered by on-die tim­
ing circuits, like a delay-locked loop (DLL). When the DLL makes continu­
ous timing adjusts after initial lock, a potential impedance mismatch caused
by VT drift can occur. The information is fed back to the ZQ calibration cir­
cuit to start the ZQ update. Periodic updates are also possible depending on
DRAM operations. For instance, ZQ updates can be scheduled during
DRAM Auto Refresh. Besides internal updates, an external calibration
scheme controlled by a memory controller is possible.
Layout can be another challenge fbr precise impedance control. Repli­
cated drivers should be used fbr ZQ calibration as well as for ODT. The
228 Chap. 10 Output Circuit Paths

same orientation and power buses should be applied to all drivers. To avoid
the physical limitations caused by tuning transistors, a minimum-sized tran­
sistor may need to be stacked (connected in serial) to provide better resolu­
tion. Because of mobility diflerence, a different number of legs can be used
in parallel for pull-up or pull-down calibrations. For instance, two pull-up
legs versus one pull-down leg results in 120 Q Rcal,DN compared to 240 Q
Rcal.Up' It is also important to isolate critical analog circuits such as the
comparator from the digital control logic. To minimize noise and cross talk,
analog signals Vref,七。and Vterm must be shielded and de-coupled. To
reduce errors, V^q and Vtern} should be tapped in the middle of the connec­
tion (small circles in Figure 10.6). Finally, layout extraction and simulation
are essential to ensure that the design meets the specs across the PVT cor­
ners.
Now, following design and layout, lefs talk about how to characterize
the output impedance. The output drive impedance can easily be approxi­
mated with the dV/dl curves during Test and Characterization. The curve is
obtained by sweeping the Vout from 0 to Vddq and measuring 101tt. An
example of GDDR3 output impedance versus temperature is plotted in Fig­
ure 10.9, with both simulated and silicon data running at Vj)£)q = 1.8 V and
at typical process comers. To extract the driver impedance from the curves,
divide the delta Kby delta I in the output driver's operating range. The pull-
up impedance shows tighter spread than the pull-down impedance in a two-
step ZQ calibration. Nonlinearity is also observed when Vout is approaching
0 for pull-down impedance or Vddq 命r pull-up impedance.
Pull-down and pull-up curves Pull up/down impedance, 40 ii

%“(V)

a) Simulation data fbr output l-V curves and driver impedance.

Figure 10.9 GDDR3 output I-V curves and impedances versus temperature.
Sec. 10.2 Impedance Control 229

k(40 Q) versus (40 Q) versus 2go

b) Silicon data fbr output I-V curves.

Zour Pull-up (40 Q) versus 匕g-%, 「
Z0y Pull-down (40 C) versus V00a
60 00

55.00 55.00

£ 50.00 50 00

? 45.00
4500

40.00 40.00

35.00 35.00

30 00 30.00

c) Silicon data fbr output impedances.

Figure 10.9 GDDR3 output I-V curves and impedances versus temperature.

Other impedance control schemes are reported in [2-3] with different

termination circuits, as shown in Figure 10.10. Passive resistors are binary
weighted fbr both pull-up (UPVR) and pull-down (DNVR) impedances.
Similar two-step self-calibration is adopted fbr impedance and ODT con­
trols. Seven-bit digital codes (pull-up and pull-down code) generated by the
up-down counters select different impedances. Compared to transistor-only
tuning, as shown in Figure 10.6, this scheme can achieve better linearity
because there is a resistor in each tuning path. However, temperature and
process variations of tuning resistors must be addressed. Matching and a
potentially large area are the primary concerns.
Instead of digital-controlled calibration, an analog impedance control
scheme is shown in Figure 10.11 [4]. The impedance is adjusted by chang­
ing the bias current, which is controlled by an op-amp. Proper compensation
(Rc and Cc) stabilizes the op-amp. The benefits are good linearity and con­
tinuous impedance adjustment.
230 Chap. 10 Output Circuit Paths

Figure 10.10 Digital-controlled impedance calibration [2-3].

ODT unit in feedback loop

Figure 10.11 Analog impedance controls [4].

10.3 SIMULTANEOUS SWITCHING NOISE (SSN)

Simultaneous switching noise (SSN) is a serious concern for high-speed I/O
design. SSN is referred to as I/O power supply noise generated by simulta­
neously switching output drivers. In order to drive data from the chip, out­
put drivers must supply large currents, usually more than ten milliamps
depending on the DQ supplies and Zo, When there are numerous (N) simul­
taneous switching outputs (SSO) connected to the same pair of power and
ground pins, a substantial discharge current 依)flows through the power
supply bus causing voltage bounce (右“ or SSN, given as
Sec. 10.3 Simultaneous Switching Noise (SSN) 231

di«
vs = LP~dl

(10.8)

In Equation (10.8), a lumped effective inductance Lp models the para­

sitic components of the power supply bus, pad, bonding wire, and package.
As N is increased, the voltage bounce increases, although the effect on
driver performance is not linear but rather has a square-root relationship due
to the gate-to-source voltage dependency.
There are several ways to reduce SSN as a result of high-speed opera­
tion. The first is to add more power and ground pins, which effectively
reduces parasitics and improves noise immunity. A 136 FBGA (fine ball
grid array) GDDR4 ball-out diagram is shown in Figure 10.12 with 22
Vddq pins and 24 V^SQ pins dedicated to I/Os. Every two DQs share a pair
of Vddq"ssq，which is essential for power delivery and SSN reduction.
Separated power pins (i.e., Pdl/Pss and ydda^ssa) also reduce noise cou­
pling from the I/O supplies into other sensitive timing circuits, like the DLL
and clock distribution network (CDN), for example. Power bus simulation
(Chapter 13) can be useful for analyzing SSN with power grid and package
models.
More advanced packaging techniques can further reduce effective Lp
and efficiently reduced SSN. Substrate layouts fbr both single-layer and
multi-layer packages are compared in Figure 10.13 for four DQs and their
associated power supplies. The black dots are for Vssq and the rest are fbr
DQO-3 unless otherwise stated. Notice that the DQ traces are matched.
Also, each DQ is sandwiched by power and ground, an effective way to
reduce cross talk and inductance.
The measured impedances for these two packages (value normalized by
dividing by 2nf) are shown in Figure 10.14. The normalized impedance val­
ues allow us to easily compare the effective low-frequency inductance (in
nH) of the two package designs. The resonance in the multi-layer package is
due to the parallel resonance between plane capacitance and ball induc­
tance. The resonance can be suppressed by adding a decoupling capacitor
(%f) between power and ground. Based on the data, over InH reduction
can be achieved with a multi-layer package.
232 Chap. 10 Output Circuit Paths

8x17 balls (0.8X, 0.8Y mm pitch)

1 2 3 4 9 10 11 12

A VDDQ VDD VSS ZQ MF VSS VDD VDDQ

B VSSQ DQO DQ1 VSSQ VSSQ DQ9 DQ8 VSSQ

C VDDQ DQ2 DQ3 VDDQ VDDQ DQ11 DQ10 VDDQ

D VSSQ WDQSC RDQSO VSSQ VSSQ RDQS1 WDQS1 VSSQ

E VDDQ DQ4 DM0 VDDQ VDDQ DM1 DQ12 VDDQ

F VDD DQ6 DQ5 VSSQ VSSQ DQ13 DQ14 VDD

G VSS VSSQ DQ7 CAS* cs* DQ15 VSSQ VSS

H VDDQ RAS* CKE* B0/A1 B1/A5 WE* NC VREF

VREFD
J VSSA RFU RFU VREFC CK* CK VSSA
/NC
K VDDA A10/A0 A12/A2 VSS VSS B2/A6 A5/A4 VDDA

L VSS VSSQ DQ25 A11/A3 A9/A7 DQ17 VSSQ VSS

M VDD DQ24 DQ27 VSSQ VSSQ DQ19 DQ16 VDD

N VDDQ DQ26 DM3 VDDQ VDDQ DM2 DQ18 VDDQ

P VSSQ /VDQS3 RDQS3 VSSQ VSSQ RDQS2 WDQS2 VSSQ

R VDDQ DQ28 DQ29 VDDQ VDDQ DQ21 DQ20 VDDQ

T VSSQ DQ30 DQ31 VSSQ VSSQ DQ23 DQ22 VSSQ

V VDDQ VDD VSS SEN RES VSS VDD VDDQ

Figure 10.12 GDDR4 x32 136 FBGA ball-out diagram.

Figure 10.13 Layout of a) wire bond and b) multi­

layer.
Sec. 103 Simultaneous Switching Noise (SSN) 233

a
p 3.0E-9
apmcSBE (&/Z

m
a 2.5E-9
e
E 2.0E-9

{PQ/Z
1 5E-0
1 .OE-9

0.0 0 5 1.0 1.5 2.0 2.5 3.0 3.5 4 0 4.5 5.0

a) Power-ground (left) and signal-power (right) normalized loop impedance.

3.0E-9 3.0E-9
(Dp
m2.5E-9
c 2.5E-9
6 Multi-layer
B2.0E-9
E 2.OE-9
1.5E-9

1.0E-9

10.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4 0 4.5 5.0
Freq., GHz Freq., GHz

b) Signal-ground normalized loop impedance without Cfat (left) and with CfQt (right).

Figure 10.14 Measured loop impedance fbr different packages.

To investigate the package impact on SSO, a simulation was set up

based on the packages just described with the following:
♦ Open-drain drivers, with three DQs (DQ1, DQ2, and DQ3) drivers as
aggressors and 1 DQ (DQ0) driver as victim.
• Aggressors run PRBS pattern (same phase among aggressors) and vic­
tim runs clock pattern.
• Data running at 1.333Gbps with a rise time of 300Ps.
• 1000 mil transmission line (60 Q characteristic impedance) with 60 Q
termination to K77：
The simulation results are shown in Figure 10.15. Timing jitter induced
by SSO noise is reduced by 53% using the multi-layer package.
234 Chap. 10 Output Circuit Paths

Jitter=101ps MaxSlew=3.80 V/ns MinSlew=1.98 V/ns

Jitter=47ps MaxSlew=3.20 V/ns MinSlew=2.77 V/ns

Figure 10,15 )SSO simulation results for a 60 ohm load:

a) wire bond package and b) multi-layer package.

Advanced package technologies reduce effective inductance at a higher

cost. From a circuit point of view, slew-rate control is a cost-effective way
to reduce SSN and output ringing. Slew-rate control is the technique of con­
trolling large transient cunent spikes during signal switching, i.e,, dis/ dt in
Equation (10.8). As mentioned, the last stage of the output driver usually
contains large-sized devices in order to drive the load (pin capacitance, in
several pico-farad ranges). A tapered pre-driver is usually needed to scale
up transistor sizes prior to the output driver. The delay of the pre-driver can
be adjusted to avoid crowbar cunent and to achieve a certain slew rate from
the output driver. The output driver can also be divided into several sub-
buflers (legs), which are connected in parallel to achieve the desired imped­
ance and drive strength. The distributed pre-drivers must be matched to
ensure identical timing for each output path.
Sec. 103 Simultaneous Switching Noise (SSN) 235

Moreover, by staggering the tum-on time of the three legs of the output
driver, shown in Figure 10.16, peak switching currents are reduced, result­
ing in less ringing and lower SSN. Gate delay or RC delay can be used to
create a small delta % the value of which depends on the targeted rise/foll
times and data rate. The range of T can be easily tailored from 3Ops to

Minimizing I/O pin capacitance (Cjq) is another way to improve speed

and reduce both active current (C[q * dv/dt) and SSN. Lower Cjq can be
achieved by using advanced package technologies. All DDR2 SDRAM
require FBGA packages versus TSOP used for SDR/DDR SDRAM. Dia­
grams for these two packages are shown in Figure 10.17. Table 10.1 lists
some example electrical parameters for TSOP and FBGA packages as used
in commodity DDR parts. AC performance can be improved even further
by using flip-chip and build-up package technologies.

Table 10.1 Pin RLC results for TSOP and FBGA packages.

FBGA TSOP % Change

L(nH) 1.4 4.1 -65%

C(pF) 0.4 1.0 -60%

R(mQ) 20 200 -90%

236 Chap. 10 Output Circuit Paths

Figure 10.17 Lead-frame-based (TSOP) packages on the left versus fine balkgrid array­
based (FBGA) board-on-chip (BOC) packages on the right.

From a circuit point of view, different termination methodologies, like

center-referenced (K77) or Kp£)g-referenced termination, impacts driver lin­
earity and pin capacitance differently. Generally a 夕Kerminated push-pull
driver is capable of producing smaller pin cap than an open-drain driver for
Vddq termination [3], Consolidating an output driver with the ODT circuits
can be a good way to reduce pin capacitance, which occurs naturally with
the previously discussed impedance control scheme. To avoid excessive
current, timing control for ODT is essential for P77 termination. You can
also reduce parasitics by placing the receiver and driver very close to the
pad.
Last but not least, data-bus inversion (DBI) can be used to reduce SSN.
By analyzing the data-stream pattern, I/O data can be manipulated inter­
nally to improve signal integrity. A DBI signal can be sent out via an inac­
tive data mask or dedicated pin to indicate whether or not the data state is
inverted. Three DBI algorithms are investigated fbr a high-speed graphics
memory system:

1. For an NMOS open-drain configuration, the driver is Fp^g-termi-

nated. We can reduce the maximum power by minimizing zeros
since zeros dissipate more power than ones. This algorithm is real­
ized by simply counting zeros. For a group of eight DQs, if more
than four zeros are detected, the DBI bit is set and the data is
inverted.
2. In a minimum transition algorithm we reduce SSN by reducing the
number of signals transitioning. A side benefit of this algorithm is
reduced cross talk (both capacitive and inductive) between parallel
signal paths. The transition detection is made by comparing current
data to the previous data through XOR operations. If more than four
transitions in a byte group are detected, the DBI bit is set and the data
is inverted.
3. In a balanced output algorithm we minimize the difference between
the number of ones and zeros that are transmitted simultaneously.
This technique measures the balance of zeros versus ones across the
transmitted data. If it appears unbalanced by a specific amount, we
flip half of the data prior to transmission. The effect is to make the
average current drawn from the supplies more consistent from cycle-
Sec. 103 Simultaneous Switching Noise (SSN) 237

to-cycle and hence reduce irregularities in the transmission (timing

and voltage variations)^

A simulation is set up with a five-inch MicroStrip channel. Open-drain

drivers with slew-rate control are modeled in Figure 10.18. The staggered
output enabling (匕〃力，心电 and 为〃曲)is controlled by a resistor string (RC
delay). The eye diagram results for all three DBI algorithms are summa­
rized in Table 10.2. With the minimum added complexity, data bus inver­
sion can readily reduce SSO and improve signal quality.

Figure 10.18 Output driver model used for DBI simulation.

Table 10.2 Average data eye parameters of four lanes with or without DBL

No DBI DBI Improvement

Algo­ Avg Avg Avg Avg Avg Avg

rithm height jitter PP height jitter PP height jitter PP
(mv) (PS) (mv) (ps) (%) (%)
1 237 137.8 270 101 15.24 -26.3

2 237 137.8 260 94.4 10.8 -31

3 237 137.8 273 110.3 16.1 -18.9

238 Chap. 10 Output Circuit Paths

10.4 SIGNAL RETURN PATH SHIFT (SRPS)

Besides the signal path, there is always a signal return path, which is gener­
ally through a reference power plane, either Vddq qt &S0 When a signal is
pulled up or down by the drivers at the near end or far end, its reference
plane may change during transmission and cause a return path shift (SRPS).
The impact and potential solution fbr SRPS are discussed in this section.
A simplified GDDR3 I/O system with a driver, ODT models (at the
receiver side), transmission line, and Vddq/VSSQ power planes is shown in
Figure 10.19. Lumped inductances fbr Vddq, Vssq，and I/O pins are added,
respectively, along with two decoupling capacitors: one (Cpiane) between
power planes at both ends and the other (Cfa^ placed close to the I/O. These
decoupling capacitors play an important role in reducing ground bounce and
supply noise caused by SSO.

Figure 10.19 Diagram fbr a simplified GDDR3 I/O system

with signals referenced to ground.

An example of SRPS with or without Cfat is shown in Figure 10.20a

and Figure 10.20b, respectively. The signal is referenced to the ground
plane and driven HIGH in this case. Cases fbr Vddq referenced and signal
driven LOW can be analyzed accordingly. Notice that the decoupling cap
(Cpiane) offers little help fbr SRPS, as indicated by Figure 10.20b.
Sec. 10.4 Signal Return Path Shift (SRPS) 239

Driver

Ground

a) Signal return path with Cfah no shift (抬sq return).

b) Without Cfab signal return path is shifted between V[)DQ and Vssq*

Figure 10.20 Signal return path shift (SRPS) a) with or

b) without Cfab given output driving HIGH.

An experimental system is set up with eight DQ lines (T1 to T8) spaced

12 mils apart, with Zp = 60 Q. Termination uses a 60 Q fixed resistor, A
four-layer, stack-up package is used with a signal-to-plane substrate thick­
ness of 4.5 mils and a plane-to-plane thickness of 54 mil. Inductance per
power or ground pin (Lp) is 0.5nH and 2.5nH per DQ pin. Data is running at
1500Mbps with 25 PRBS pattern. With one active aggressor (Tl) and seven
quiet victims, the cross talk is plotted in Figure 10.21, either with or without
Cpiane. Eye diagrams for one victim (T5) and seven aggressors are shown in
Figure 10.22.
240 Chap. 10 Output Circuit Paths

a)
Cross talk (wlh 100 nF decoupling capacitor。口后阪)

0 12.5 62.5 125 250 1000 1500 2250

C旬(PF)
b)
Figure 10.21 Measured cross talks with one aggressor:
a) no decoupling capacitor and b) 100 nF decoupling capacitor.

Two observations from Figure 10.21 follow:

L When the victim trace is close to the aggressor, cross talk is the com­
bination of SRPS (Signal Return Path Shift) noise and coupling
noise.
2, When the victim is far away from the aggressor (>60 mil), cross talk
is due only to the SRPS noise.
Sec. 10.4 Signal Return Path Shift (SRPS) 241

1 98

1.69

1.40

1.10

081

-0.13 —~1~11~~~~~~
0.00 0.27 0.53 0 80 1.07 1.33 0.52
0.00 0.27 0.53 0.80 1.07 1.33
T5, 1191 mV w/o Cf |t Lp=,5 nH T5, 304 mV, CIa =250pF, Lp=.5 nH
MaxV=2.453 V/ns MinV= 1.262 V MaxV=1.929 V/ns MinV=1.625 V

0.00 0.27 0.53 0 80 1.07 1.33

T5,1280 mV w/o Cfat, Lp=2 nH T5, 337 mV, C(a =250pF, Lp=2 nH
MaxV=2.496 V/ns MinV=1.216 V MaxV=1.955 V/ns MinV=1.618 V
2.60

2.28

1.96

1.64

1.32

1.00

Figure 10.22 Eye diagrams fbr the experiment system,

no Cplane (7 aggressors and 1 victim, T5).

The main SRPS issues are summarized as follows:

• Signal integrity of the main path degrades.
• Cross talk between lines increases, even with widely separated lines.
• Decoupling capacitors not located in close proximity to the SRPS
points have little effect on reducing cross talk.
, With or without Cpiane, the effects of SRPS can be significantly
reduced by connecting the power ground planes with capacitors (C^z)
close to the SRPS points.
• The more SSO, the more cross talk.
242 Chap. 10 Output Circuit Paths

, Noise can be reduced by lowering the inductance per pin.

, The amount of Cfat required for reducing noise varies for different sys­
tems and data rates. Noise can be managed with 250pF Cfat per driver in
the exemplary GDDR3 system.

10.5 ELECTROSTATIC DISCHARGE (ESD)

Reliability issues such as ESD impose certain constraints in I/O circuit
design [5]. ESD is a high-current, short duration stress. Three different
models can be used to characterize ESD stresses:

1. Human body model (HBM)

2. Machine model (MM)
3. Charge device model (CDM)

The damages caused by ESD events include oxide breakdown for

CMOS input gates, drain-substrate junction breakdown and contact punch­
through for CMOS output buffers, and post-ESD leakage current increase
between internal power pins. When designing an ESD circuit, the following
rules should be kept in mind.
• Clamp the ESD voltage to shunt the ESD stress current.
• Turn on fast (less than 1 ns).
• Carry large currents > 2 amps or more for 150 ns.
• Occupy the minimum area at the bond pad.
• Introduce the minimum capacitance and series resistance.
• Retain immunity to process drifts.
• Offer protection for various ESD stress models.
• Do not interfere with the function, testing, or operation of the IC.
• Survive and remain inactive during bum-in tests.
Different ESD devices and circuits provide different protection. Care­
fully choosing ESD devices and circuits is very important to prevent any
damage to the I/O circuits and, at the same time, achieve better perfor­
mance. ESD devices can be grouped into two categories: breakdown
devices, such as SCR (silicon-controlled rectifier) or GCNMOS (gate-cou­
pled NMOS), and non-breakdown devices, such as diodes and MOSFETs.
In general, the breakdown devices are very area efficient but difficult to
design due to their more complicated behavior.
An ESD I/O protection circuit diagram is drawn in Figure 10.23. An
input resistor (R加pass gate and CDM voltage clamps are applied to
Sec. 10.5 Electrostatic Discharge (ESD) 243

protect input gate oxide. The GCNMOS device is for the power clamp. ESD
diodes (ESDPN and ESDNP) serve as primary cunent clamps for the I/O
supplies. A large decoupling capacitor (FatCap) between power and ground
can shunt some ESD stress current and improve ESD performance.
Primary I/O supply

ODT

Figure 10.23 Diagram fbr I/O ESD protection circuits.

The output driver transistor can be built to act like a stand-alone ESD
circuit or as a high-speed driver. If we lay out the driver to meet the high
current requirements that ESD requires (i.e., 3 amps), then this would
impact the resolution of impedance matching and high-speed performance
of the driver. Instead, ESD and the driver can be separated into two different
circuits. The tuning transistors for impedance and ODT control would be
chosen to meet the desired LSB without any ESD considerations. The ESD
circuit would be then added in parallel as the best possible ESD solution.
This attempt can maximize the perfbnnance of both circuits independently.
However, this approach works up to a certain bandwidth like 3 GHz.
Beyond that, the drivers are built as both driver and ESD circuit in an
attempt to allocate design space fbr the ESD circuit function.
As mentioned before, Cjo reduction is beneficial for high-speed opera­
tions. However, bidirectional I/O with push-pull drivers and ODT circuits
add more capacitance per I/O pad. From the DDR2 Go spec, only 2.5-3pF
is allowed, which leaves little room fbr the ESD circuits. On the other hand,
the diodes fbr primary power and ground current clamps can add up to more
than 0.3pF. If the ESD circuits are too small, then there will be a yield pen­
alty as Assembly and Test still generate static electricity through their back­
end process and handling. ESD circuits using a breakdown device like an
SCR can reduce capacitance with a turn-on time limitation. A modified
SCR for twin-well technology is reported in [3] with a maximum 0.45pF
Cjo reduction. A diode-triggered SCR (DTSCR) [6] can achieve 500ps
turn-on time with only 0.2pF pin capacitance.
Noise coupling from the ESD diodes among different power domains
must be addressed. When the power bus networks are broken into sections,
244 Chap. 10 Output Circuit Paths

i.e., Vdd，and Vddq as in Figure 10.12, the ESD discharge current

event can cause voltage spikes on one power bus network and not the other.
Those circuits that receive the signal from one domain to another are at
severe risk of failing. This issue is already present with 32 A gate oxides at
the 90nm DRAM process. This problem will become worse for more
advanced processes. The receiver gate oxides are at risk of seeing the high
stress voltages, and they will either trap charge on their gates or be totally
damaged, changing their switching speeds. One way to minimize this risk is
to minimize the number of circuits at these power domain interfaces. The
other option is to reduce the fan-out after a signal is sent from one power
domain to another. For noncritical speed paths, the receiver circuit can be
designed so that its gate oxide is thick and not thin. For critical speed paths,
the circuits can be put in one specific area with very small ESD power
ciamps between and Vddq，instance. One example would be in the
pre-driver of the output path. These small ESD devices limit the voltage
transients without affecting the high-speed signals.
Layout is another critical issue when designing high-speed I/O circuits.
It has been found that the metal routing from the bond pad to the I/O driv­
ers, ODT, and ESD circuits can add a large amount of parasitic fringe metal
capacitance. Routing one metal into a circuit and then using a different level
of metal to route out of the circuit helps decrease parasitic metal capaci­
tance. By laying out the ESD circuits on the side of the bond pad, the rout­
ing to them from the pad has the smallest distance to travel.
Moreover, ESD circuits are designed to carry high current; conse­
quently, metal interconnect resistance is a critical parameter. As the technol­
ogies have scaled, the contact aspect ratio has become significant. When the
contact is formed, the stack height reduces the contact area. Adding more
contacts becomes inevitable to drive extremely high current through these
small contact areas. Unfortunately, adding more contacts widens the metal
and increases the fringe capacitance.

10.6 PARALLEL-TO-SERIAL CONVERSION

After discussing the signal integrity issues and their influence on output
driver design, we now investigate parallel-to-serial conversion.
As mentioned in Chapter 9, the memory internal data bus is fairly wide
in order to pre-fetch multiple bits at the same time, enhance parallelism, and
relieve the burden of array cycle time. Read data is typically loaded into
pipelines prior to stream out. This internal data must be serialized and ulti­
mately converted back to the external data rate. An output path is shown in
Figure 10.24, with a parallel-to-serial (P2S) converter as the starting point.
P2S generally employs a high-speed multiplexer controlled by the clock
Sec. 10.6 ParalleUto-Serial Conversion 245

edges. The outputs of P2S are fed into pre-drivers and finally output drivers
with proper slew-rate and impedance controls.

Figure 10.24 Block diagram of typical output path.

The overall output timing is denoted as touh which is from the internal
Read clock (i.e., the DLL and clock distribution network) to the DQ at the
pad. The output path is generally modeled in the timing control circuit (i.e,
DLL) and latency control logic to maintain clock synchronization and the
desired programmable CAS latency. Several timing parameters can be
derived from or related to the output timing:
• Oc as DQ output access time from external system clocks CLK/CLK*
• Iqqsck as Read strobe (DQS) access time from CLK/CLK*

• ([)qsq as DQS-DQ skew (DQ timing relative to DQS)

The first two parameters are linked to the performance of the on-die
timing control circuits, like DLL. Due to power and area constraints, the
model used in the DLL is not a straight copy of the output path. Instead, a
scaled down version of the output path is used, especially for pre-drivers
and output drivers. Moreover, the noisy DQ supplies (VD[)q/Vssq) cannot
be mapped into the output model, which could result in excessive timing jit­
ter. On the other hand, DQS-DQ skew is a measure of the matched routing
of output paths and SSO-induced jitter. For high-speed DRAM interface,
toQSQ is more stringent than tAQ and tDQSCK since strobes are used to cap­
ture the data at the memory controller. DQS is generally treated as a special
data signal having a one-zero toggle pattern. For DDR2-800, the maximum
tDQSQ is only 175ps, which is half of the tDQSCK spec. For GDDR4,
is reduced to only 80Ps.
A circuit block diagram for a P2S is shown in Figure 10.25. Four identi­
cal latches can be used with outputs wire OR'd together for multiplexing
four bits of data. Multiplexer fan-in can seriously limit the bandwidth of the
output path [7]. In Figure 10.26, a dynamic latch with domino logic can be
used for the Read latch. The dynamic latch has speed advantages over the
static latch because the load for input (D) is only one NMOS transistor
246 Chap. 10 Output Circuit Paths

(WN39). The internal sample clock CLKsp is single phase with light load­
ing. A keeper circuit bleeds a small current to counter the leakage through
an internal high-impedance node, d*, Sizing the output stage of this latch
and buffer stage in the P2S is critical to maintaining good duty cycle.
Data is first sampled by the rising edges of multi-phase internal Read
clocks, which run at lower speed for less distortion and better clock distri­
bution. A clock sampling window for bit 0 is shown in Figure 10.27, where
CO is at logic C” and C90 at logic "0." CO and C90 are connected to CLK1
and CLK2 signals of the Read latch, respectively. Clock routes and loads are
matched inside the P2S to maintain the phase relationship. Implementation
of the phase generator is discussed in Chapter 11.
An output enable (OE) is generated from the latency control logic to
start the serialization. The OE signal passes through the same latch and
selects either the latch output, ODT, or tri-state signal. Then proper pull-up
or pull-down signals are generated and fed into the pre-divers and output

Figure 10.25 Circuit diagram of a parallel-to-serial converter.

Figure 10.26 Dynamic Read latch fbr high-speed P2S.

Sec. 10.6 Parallel-to-Serial Conversion 247

Figure 10.27 Four-phase clocking and sampling window for P2s.

Some speed limitation is imposed by the P2S because the data is serial­
ized and running full-speed from this point. Due to the bandwidth limitation
of the CMOS gates, it is difficult to push the bit time below 400ps (2.5Gbps
data rate) without encountering some signal distortion. One way to handle
this is to push the serialization out to the output drivers [8]. This output-
multiplexed architecture requires multiple copies of the output driver, which
is power- and area-hungry. Current-mode logic (CML) is another solution
for the output path when the speed is beyond 2.5Gbps. With reduced volt­
age swing and fully differential signaling, CML offers superior performance
over traditional CMOS logic with the added cost of constant current draw
and a more complicated design. A CML-based output path [7] is shown in
Figure 10.28. Pass transistors controlled by different phases of the clock are
used for input multiplexing. True and complementary signals are required
for CML, which effectively doubles the logic.

V
Figure 10.28 CML-based output paths [7].
248 Chap. 10 Output Circuit Paths

10.7 EMERGING MEMORY I/O FEATURES

As I/O speeds continue to increase, circuit techniques generally found in
communication systems will start penetrating into the memory interface
domain. We discuss a few of these techniques in this section.

I. Differential signaling. Single-ended (SE) signaling is the standard fbr

data in current DRAM interfaces due to pin restrictions and band­
width requirements. SE signaling is susceptible to SSN and com­
mon-mode noise. Differential signaling, on the other hand, can reject
common-mode noise, reduce ground bounce and SSN, minimize
EMI, and double the slew rate. To justify the increasing pin count,
however, data must run twice as fast. High-speed differential logic,
like CML is well suited for differential signaling.
2. Unidirectional I/O. Bidirectional I/O generally becomes a burden fbr
high-speed data operations. To avoid bus contention, timing restric­
tions are needed, which hurts Read and Write turnaround. Pin capac­
itance is another hurdle for speed increases. Separate inputs and
outputs are generally beneficial except for the added cost of higher
pin counts. Simultaneous bidirectional (SBD) signaling [9-10] is
appealing if the bandwidth and pin counts are truly a concern. With
limited swing and multiple reference points, SBD signaling would be
a challenge fbr both Design and Test.
3. Transceiver equalization and pre-emphasis. From transmission line
(TL) theory described in Section 10.1, attenuation of the TL
increases with frequency due to the skin-effect resistance defined in
Equation (10.3). The digital signal used in memory systems includes
both low-frequency and high-frequency components. The superposi­
tion of non-attenuated, low-frequency signal components with atten­
uated, high-frequency signal components causes inter-symbol
interference (ISI). This interference degrades noise margins and
reduces the maximum frequency at which the system can operate.
Equalization fbr both the transmitter and receiver can be incorpo­
rated to boost the high-frequency component and eliminate the prob­
lem of frequency-dependent attenuation. The desired frequency
response should be flat over the entire operating range. Compared to
receiver equalization, transmitter pre-emphasis is much easier to
implement with a multi-tap FIR (finite impulse response) filter [7-8].
However, adding an equalization path increases the power and area
of the output drivers.
4. Data encoding has long been used in communication systems fbr bit
error correction. To achieve better bit-error-rate (BER), error-correct­
ing code (ECC) or cyclic-redundant code (CRC) can be applied to a
Sec. 10.7 Emerging Memory I/O Features 249

memory system. However, data encoding would make memory

design much more complicated.

Obviously, any of these circuit techniques to improve I/O performance are

not free, especially considering added cost for test and validation. Neverthe­
less, the path toward higher performance always necessitates greater com­
plexity as evidenced by the evolution of circuits used in SDR, DDR, and
DDRII devices. For a specific memory technology, good engineering judg­
ment is needed to balance performance needs against complexity and cost.

REFERENCES

[1] H. Johnson and M. Graham, High-Speed Signal Propagation, Advanced Black

Magic. Prentice Hall, Upper Saddle River, NJ, 2002.
[2] H. Y. Song, S. J. Jang, J. S. Kwak, C. S. Kim, C. M. Kang, D. H. Jeong, Y. S.
Park, M. S. Park, K. S. Byun, W. J. Lee, Y. C. Cho, W. H. Shin, Y. U. Jang, S.
W. Hwang, Y. H. Jun, and S. I. Cho, "A 1.2Gb/s/pin double data rate SDRAM
with oivdie-tennination,“ ISSCC Digest of Technical Papers, pp. 314-315,
2003.
[3] C. Park, H. Chung, Y. Lee, J. Kim, J. Lee, M. Chae, D. Jung, S. Choi, S. Seo,
T. Park, J. Shin, J. Cho, S. Lee, K. Song, K. Kim, J. Lee, C. Kim, and S. Cho,
“A 512-MB DDR3 SDRAM prototype with Cj0 minimization and self­
calibration techniques,IEEE Journal of Solid-States Circuits, vol. 41, April
2006.
[4] Y. Fan and J. E. Smith, uOn-die termination resistors with analog impedance
control for standard CMOS technology/' IEEE Journal ofSolid-State Circuits,
vol. 38, Feb. 2003.
[5] S. Dabral and T. J. Maloney, Basic ESD and I/O Design. Wiley & Sons, Inc.,
New York, 1998.
[6] J. D. Sarro, K. Chatty, R. Gauthier, and E. Rosenbaum, uStudy of design
factors affecting tum-on time of silicon-controlled rectifiers (SCRS) in 90 and
65mn bulk CMOS technologies,in IEEE 44th Annual International
Reliability Physics Symposium, 2006, pp. 163-168.
[7] M. E. Lee, W. J. Dally, and P. Chiang, ^Low-power area-efficient high-speed
I/O circuit techniques,IEEE Journal of Solid-State Circuits^ vol. 35,
November 2000.
[8] W. J. Dally and J. Poulton, ^Transmitter equalization for 4Gb/s signaling,,,
Proc. Hot Interconnect, pp. 29-39, Aug. 1996.
[9] E. Yeung and M. Horowitz, UA 2.4 Gb/s/pin simultaneous bidirectional
parallel link with per-pin skew compensation/' IEEE Journal of Solid-State
Circuits, vol. 35, pp. 1619-1628, Nov. 2000.
[10] J. Kim, S. Kim, W. Kim, J. Choi, H. Hwang, C. Kim, and S. Kim, “A 4-
Gb/s/pin low-power memory I/O interface using 4-level simultaneous bi­
250 Chap. 10 Output Circuit Paths

directional signaling,'' IEEE Journal of Solid-State Circuits, vol, 40, pp.

89-101, Jan. 2005.
 Chapter
11

Timing Circuits
As mentioned in Chapter 5, timing circuits, such as those used fbr clock
synchronization, are an important class of circuits. In this chapter, we
explore various types of timing circuits and their implementation. Among
these, delay-locked loop circuits (DLL), both digital and analog, are cov­
ered in detail due to their simplicity and wide use in high-speed DRAM.

11.1 INTRODUCTION
Timing circuits, in general, are associated with a clock. A system clock,
which is generated by the memory controller, synchronizes the input and
output (I/O) operations of the DRAM. It makes sense then that two of the
biggest obstacles for high-speed I/O operations are clock-related: clock
skew and jitter. Clock skew is any timing difference between the external
system clock and the internal DRAM operating clocks. A propagation delay
from the I/O circuits and local buffers (to drive long interconnects for clock
distribution) can skew the internal clock relative to the external clock.
Delay mismatch among different clock paths and duty-cycle distortion can
impact the timing budget even further. Dynamic clock jitter, on the other
hand, can be introduced by supply noise; cross talk; and voltage and tem­
perature (VT) variations. Timing variations (both skew and jitter) along the
critical data path could cross several clock cycles as speeds increase making
it difficult to predict and track latency. To reduce such variations, timing cir­
cuits such as clock synchronization or de-skewing circuitry are generally
used. These play an increasingly important role in high-speed DRAM cir­
cuit design.
To achieve better performance, the critical timing path of the memory
device from input to output must be carefully analyzed and modeled. The
critical timing path includes the input path (Chapter 9), output path

251
252 Chap. 11 Timing Circuits

(Chapter 10), clock distribution networks, and one or more de-skewing cir­
cuits. Latency control (Chapter 12) and power delivery (Chapter 13) must
also be considered when designing the timing circuits. Although high-speed
circuit design for de-skewing is the focus of this chapter, each piece along
the critical timing path inevitably interacts with the core sync circuits and
contributes to the overall timing budget. Several critical timing specs, such
as Uc，fDQSQ^ tDQSS，CAS latency, data setup and hold, etc., can be deter­
mined through the analysis. Therefore, the design of clock synchronization
circuits are integral to the overall system; design modifications and
improvements are necessary to accommodate increasing clock speed.
So, how is the design of synchronization circuits impacted by increas­
ing clock frequencies? First, the timing margins shrink with a shorter clock
cycle time. In turn, the clock and data alignment must be more precise and
tolerate less jitter. Because both the rising and falling edges are utilized,
duty-cycle distortion jeopardizes the accuracy of data processing. Second,
clock domain crossing becomes more of a problem due to timing uncer­
tainty arising from clock skew and short cycle time. With the inherently
slow transistors and minimum number of metal layers of a DRAM process,
clock and power delivery is a challenge: ground bounce, voltage bumps,
and noise from cross coupling degrade timing. Last but not least, impedance
mismatch and inter-symbol interference (ISI) degrade signaling, which
necessitates on-die calibration and training. Advanced packaging technol­
ogy must be considered and modeled in the design flow to analyze timing
and power delivery. Since cost remains the primary concern fbr DRAM
manufacturers, design trade-offs must be made between cost and perfor­
mance, simplicity, and scalability.
Historically, sync circuitry first appeared in DDR SDRAM, in which
the clock frequency was approximately 66-133 MHz. A simple digital
delay-locked loop was used to synchronize the clock with output data.
High-performance GDDR4 DRAMs have pushed speeds up to multi-giga­
hertz range, while scaling cycle times down to less than 1 ns. Achieving
such high speeds requires enormous complexity in the sync circuit and sys­
tem design. However, variability in the lock time, tuning range, resolution,
jitter, power, and layout can greatly impact overall system performance.
Based on various design choices, sync circuitry can be implemented
using digital or analog, open or closed loop methodologies or any combina­
tion thereof. The two most common sync circuits in use today are the delay-
locked loop (DLL) and the phase-locked loop (PLL). The DLL is widely
used in memory interfaces for its simplicity and good jitter performance. An
alternative approach is the PLL, which is usually found in communication
systems and microprocessors fbr clock data recovery and frequency synthe­
sis. Both the DLL and PLL are closed-loop systems with a reference and
Sec. 11.1 Introduction 253

feedback signal. The biggest difference between a DLL and PLL is that a
DLL merely passes the reference signal through a variable delay line, while
a PLL generates a new signal that is both phase- and frequency-locked to
the reference signal.
Although both the DLL and PLL are good for clock de-skewing, the
PLL is rarely used in high-speed DRAM due to the following:

1. Increased complexity
2. Stability issues inherent in a higher order system
3. Need for frequency locking
4. Longer lock time

Unlike the PLL, jitter in a DLL is small and easy to control due to the
lack of jitter circulation, which occurs in a PLL oscillator. However, any
input jitter may directly pass to the DLL output without any filtering. With a
low-jitter reference signal, a well-designed DLL can achieve the same jitter
performance as a PLL fbr clock synchronization.
An analog approach generally includes a voltage-control delay line
(VCDL) and a charge-pump phase detector to generate a control signal [3].
However, an all-digital DLL (ADDLL) can be implemented fbr greater scal­
ability and portability across DRAM process generations. Because logic
gate delay remains fairly constant as feature size and supply voltage scale,
an all-digital design can be reused with little modification and is also easy
fbr test and debug, an issue which cannot be over-emphasized for modem
DRAM design. Good stability over process, voltage, and temperature (PVT)
variations is another plus because it is a Oth-order system without integra­
tion. A register-controlled DLL [1, 2] is discussed in Chapter 5, and
ADDLL design is addressed in detail in Sections 11.2 of this chapter.
A higher performance DRAM generally supports a variety of power­
saving modes. To resume normal operation after exiting power-saving
modes, the sync circuit must be relocked, resulting in some start-up delay.
Lock time of a sync circuit becomes an important design issue, and digital
implementations usually surpass analog counterparts in this area. Some
open-loop architectures such as synchronous mirror delay (SMD) [5] and
ring counter delay (RCD) can achieve lock within several cycles although
they lack fine timing resolution. A combination of DLL and SMD, called
measure-controlled delay (MCD), has been proposed to take advantage of
quick initialization with a closed-loop structure for better voltage and tem­
perature tracking after initial lock. Voltage and temperature variations intro­
duced by the power saving in DRAM can greatly impact critical timing
needed fbr high-speed operation. Power-saving modes may need to be
traded for better overall timing performance.
254 Chap. 11 Timing Circuits

As mentioned, clock and power delivery are big challenges for high­
speed DRAM circuit design. To maintain proper slew rate and duty cycle,
the external clock must be shaped and re-driven over a long distance inter­
nally. Clock power for the distribution network can represent as much as
70% of the entire timing path and even higher as the speed increases. One
solution to this problem is scaling down the internal clock frequency by
using a clock divider. Except for the I/O circuits, a single-phase, low-speed
clock signal can be distributed globally across the die with less attenuation
and power. However, the high-speed interface still processes data in full
data rate (twice the clock frequency), and an edge recovery circuit must be
added to transform the internally divided-down clock into multi-phase
clocks that include ali missing edge information. A mixed-mode DLL with
an analog phase generator can be configured to suit this purpose. It is obvi­
ous that increased speeds demand more complex clocking system design.
Layout is another challenge fbr high-speed clock synchronization
design. Full custom layout is generally required for better matching, rout­
ing, sizing, and shielding. The delay line (for delay adjustment) is without
question the most important piece as it carries a clock signal. For mixed­
mode layout, partitioning sensitive analog blocks from digital circuits with
proper isolation, bussing, and placement is key for better performance.
A circuit design is not complete without built-in test capability. Design-
fbr-test (DFT) is essential for clocking in high-speed DRAM products.
Wafer-level probing is a challenge due to power delivery, high-speed clock
and signal delivery, and availability of limited Probe resources. Trouble­
shooting can be extremely time-consuming if not impossible. Along with
circuit design, we address various test issues related to timing circuits.

11.2 ALL-DIGITAL CLOCK SYNCHRONIZATION DESIGN

Designing clock synchronization circuits starts with basic timing analysis,
as shown in Figure 11.1, which depicts the critical timing path. The typical
timing path includes an input path (d),a DLL with an adjustable delay line,
a phase detector (PD) and feedback model (tmodei)-> clock distribution Mee),
and a data output path itou^. An all-digital DLL (ADDLL) is used as an
example to generate an internal clock fbr latching the data (from the array)
and synchronizing outputs (the DQs) with the external system clock
(XCLK). The major building blocks of the ADDLL are discussed with
respect to different design aspects, including layout and test issues. Design
modifications and improvements fbr high-speed, all-digital implementation
are the fbcus of this section.
Sec. 112 All-Digital Clock Synchronization Design 255

Figure 11.1 Timing parameters of a typical all-digital DLL used in a memory system.

11.2.1 Timing Analysis

Intrinsic delay is defined as the delay between the clock input
(XCLK) and the data outputs (D0s), excluding the actual insertion delay
(tocD provided by the DLL delay line. The intrinsic delay is also referred to
as intrinsic forward path delay. From Figure 11.1, tiNTRl can be expressed as

OnTRI = 4〃 + huff^ hD + hree + 'out

where tin is the delay of the clock receiver and input buffer, and e崂 is the
delay reserved for tuning after initial lock due to voltage and temperature
variations. Parameter 勿 includes DLL buffer delays and a fine adjustment
delay, if any. The delay of the data output path tollt includes the DQ latches
and DQ drivers. For low-speed operations, the cycle time «ck) is usually
greater than “nt肛 As speed increases, Sk may be a fraction of "ntri，and
tight synchronization is needed to reduce timing skew. In addition to the for­
ward path, a DLL loop also includes a feedback path. An I/O model takes an
input from the clock tree and generates an output signal fed back to the PD.
The I/O model tries to track the delay of input and output paths and can be
defined as

model 〜hn + 'out

(11.2)
256 Chap. 11 Timing Circuits

When the DLL is locked, the total forward path delay, troTAL^ is equal to m
times the clock period "ck)，where m is an integer number greater than
zero. The relationship can be given as

，TOTAL = m ' lCK = fINTRI + Sock产

(H.3)
where is the insertion delay when the DLL is locked. Assuming a digi­
tal delay line with a total of N stages where the delay per stage is Zp (unit
delay) and n is the number of delay stages involved in the delay line when
the loop is locked, the tuning range can be given as

hock = ntD^ 1

(114)
Generally, it is important to know the timing difference caused by pro­
cess, voltage, and temperature (PVT) variations. Table 11.1 and Table 11.2
present data based on an exemplary timing analysis. A digital delay line
with fine and coarse tuning is shown as part of the DLL intrinsic delay in
Table 11.1. Different implementations produce different DLL intrinsic
delay. The clock input buffer also includes a clock divider for high-speed
operation. The sum of the intrinsic forward path delay is shown as the total
delay.

Table 11.1 Timing data under PVT variations.

PVT Input Output DLL Clock Tree Total

Corners “tt Intrinsic hree Delay

Slow, 1.3 1685 1615 3240 1967 8507

V, 85℃
(ps)

Fast, 1.9V, 721 780 1816 850 4167

0℃ (ps)

A(ps) 964 835 1424 1117 4340

% of delay +57% +51% +44% +56.8% +51%

change

% of total 17~20% 19% 38-43% 20-23% 100%

delay
Sec. 11.2 All-Digital Clock Synchronization Design 257

The data in Table 11.1 show the sensitivity of major timing blocks ver­
sus PVT variations. The clock input path, clock distribution, and data output
path represents 60% of the overall forward path delay at approximately 20%
for each. Larger die size and additional data pins require a larger clock dis­
tribution network and produce longer delay. The ADDLL intrinsic delay
accounts for the remaining 40% of the forward-path delay. Worst-case PVT
variation could introduce 60% timing skew (five cycles of variation) for the
total forward delay after lock for a 1GHz external clock.
Given a typical process, the voltage and temperature (VT) variations
depicted in Table 11.2 produce up to 2.1 ns (40%) timing skew (1.6 ns is
due to the 600 mV supply change). For every 100 mV drop in the nominal
1.5 V supply, the total forward path delay (excluding DLL delay) can
change by 360 ps, almost an entire bit time at a 2.5GBit/s data rate. Com­
pared to lowering the voltage, increasing the voltage has less impact on
delay. On the other hand, a 100℃ temperature change may vary the delay
by half a nanosecond with higher voltages creating even greater temperature
sensitivity.

Table 11.2 Timing data under voltage and temperature variations with a typical process.

Conditions Input Output CLK Tree lotal Delay

1.5 V； I1O℃ 3966

1298 1196 1472
(ps)

1.5 V,0℃(ps) 1130 1057 1287 3474

1.2 V 110℃ 5040

1665 1519 1856
(ps)

L2V,0℃(ps) 1499 1386 1669 4554

1.8 V. 110%： 3440

1124 1028 1288
(ps)

1.8 V, 0℃ (ps) 951 885 1087 2923

% of the total 100%

33% 30% 37%
delay

Delay change
157 130 183 470
(ps)/+100℃

Delay change
123 109 128 359
(ps)Z-O.l V

Delay change
-59 ~57 -64 -180
(ps)/+0.I V
258 Chap. 11 Timing Circuits

Supply noise sensitivity is the number one cause of timing variations in

high-speed, low-voltage DRAM. Slower transistors make it worse. All-dig­
ital implementation likely takes a much bigger hit due to the lack of supply
noise rejection. Supply noise can be triggered by array activation, coupling,
and IR drop. Delay variation due to supply noise can accumulate and trans­
late to significant short-term timing jitter. A power delivery scheme is
shown in Figure 11.2. Dedicated power buses with large decoupling capaci­
tors can be applied to the critical timing blocks, e.g., input buffer, clock
trees, etc. These "quiet“ power buses are isolated from other global circuits.
Dedicated power and GND pins are assigned to the DLL along with its own
bussing. The DLL is placed close to the input path to produce shorter routes.
Voltage regulation may be another way to reduce supply noise at the cost of
lower voltage margin at high speed. If the clock is stopped for power sav­
ing, the voltage perturbation caused by supply current changes may result in
longer settling and relock time.
I/O model mismatch is another factor fbr timing skew. Locality is
important fbr better matching in that the input model should be placed in the
proximity of real input buffers with the same buses. Not only logic style, but
gate counts must be modeled, and metal routes must account fbr the feed­
back signal. However, noise injected from a noisy bus (like supplies fbr out­
put data) can make the DLL loop unstable. A relative constant current
profile for the entire timing path is less prone to supply noise. Layout with
proper shielding, isolation, and placement fbr timing critical circuits should
be the first priority. Substrate coupling is also a concern. It is desirable to
properly strap the n-well and substrate. A circuit with better supply noise
rejection, like fully differential current-mode logic (CML), can be very use­
ful and is discussed in Section 11.3.1 of this chapter.
Array interface

Periph logic

DQ output Repeaters DLL core timing (42%) Clock input timing

timing 40% (18%)

Clock trees
(matched/shielded routing)

DQ latches
______ 土.........
DQ pre-drivers & drivers :

DQ pads and supplies

Package

Figure 11.2 Power bussing scheme for a clock synchronization system.

Sec. 11.2 All-Digital Clock Synchronization Design 259

11.2.2 Digital Delay Line

A digital delay line provides an adjustable delay for the ADDLL. The
delay is adjusted discretely by adding or removing gates from the delay line.
A delay line can be characterized with two parameters: unit delay (Zp) and a
number of stages (N). The multiplication of the two parameters defines a
lock range in which a DLL can operate under all PVT variations. The lock
range for a synchronous memory system tends to be wide to cover various
frequencies necessary to support speed grading. A speed grade represents a
set of timing parameters tied to device operation at specific data bus fre­
quencies. For high-speed DRAM, these timing parameters become more
demanding and harder to meet. Fortunately, devices unable to meet all of the
timing specs for a specific speed grade can be down-binned to a slower
speed grade. Different speed grades put a burden on timing circuit design to
cover a wider operating range for a given process.
1122.1 Unit Delay (t]y) Unit delay of the digital delay line deter­
mines the maximum jitter and number of delay stages that must be included
to ensure proper lock range. The value of this unit delay may vary
-40/+60% with process, voltage, and temperature (PVT) variation. Typical
unit delay in current DRAM process lies in the 100-250 ps range, with two
NAND gates connected serially. The NAND gate, instead of an inverter, is
chosen for its simpler control logic and reduced power consumption (due to
unused delay stages). The delay resolution can be improved by increasing
the drive of the basic delay element with the cost of more power dissipation
and larger silicon area. More delay stages are needed to ensure a desirable
lock range with a smaller unit delay. Table 11.3 shows the simulation results
for the NAND-gate-based unit delay under different conditions (which may
vary according to process and layout arrangements). A careful layout to put
the critical delay elements close together can improve the unit delay by
reducing interconnect parasitics. However, the best way to improve the res­
olution of the digital DLL is through delay interpolation, creating finer
delays (as low as 10 ps).
Table 11.3 Unit delay (two NAND gates, in ps) under different conditions.

Schematic Layout
Simulation Corners*
(Fixed Load) (Extraction)

Fast: I.9V,FF,-40℃ 83 81

Typical: 1.5 V,TT,85℃ 142 134

Slow: 1.2V； SS, 110℃ 227 219

*(TT: typical NMOS and PMOS; FF, fast comer; SS, slow comer.)
260 Chap. 11 Timing Circuits

Graphical representation of Table 11.3.

The capacitive load (Q) of one delay element (half of the delay stage)
is the sum of output capacitance, input capacitance of the next delay ele­
ment, and interconnect capacitance. Because the CMOS NAND gate has
four transistors versus two in the CMOS inverter, Cj is comparably larger in
the NAN D-based delay stages. The size of the delay element also deter­
mines the value of Q, which should be a compromise between speed,
power dissipation, and area. For high-speed, low-voltage applications, each
delay stage should provide enough gain (or drive) to prevent the signal tran­
sition slope or slew rate from being too sluggish. The transition time, rise
time (Jplh) + 回 1 time (加⑺，should be less than 40% of the cycle time for
better signal quality and lower attenuation. Unlike analog implementations,
which change delay by varying current for every stage, the all-digital
approach preserves slew rate of the clock signal regardless of the delay
range provided by the DLL.
Making a symmetrical transition "hl = 1plh) is very important for low
duty-cycle distortion. A seemingly small (10 ps mismatch) for the delay
stage can result in significant distortion at the DLL output. Because the
mobility of P devices and N devices are significantly different, it's not easy
to match the rise and fall time over process variation. Stresses seen during
bum-in may result in negative bias temperature instability (NBTI), causing
device mismatch and producing excessive duty-cycle distortion. Duty-cycle
amplification along the critical timing path can make such mismatch even
Sec. 1L2 All-Digital Clock Synchronization Design 261

worse. A two-input, NAND-gate-based delay line has lower overall duty­

cycle distortion [2] than an inverter-based delay line in part because it
doesn't require a MUX array to select an exit point. Matching interconnects
between node A and B in Figure 113 is critical for achieving symmetrical
transitions. The same logic state must be preserved for each delay element
even if it is not used.

Figure 11.3 A symmetrical NAND-gate-based digital delay line.

For an entry-point selection delay line (like a register-controlled delay

line), clock selection and loading is a potential problem. Because the input
clock signal (CLKlri) is fed into every delay stage and only one entry point
exists at a given time, as shown in Figure 11.4, the clock must be gated or
treed for power saving. It is not trivial to control the input clock signal for
reduced jitter and power consumption. Although only a NAND gate is
shown in Figure 11.4 for entry-point selection, based on the shift-register
design, other logic can be used to save power and prevent control disconti­
nuity during shifting.

Figure 11.4 Register-controlled DLL with entry-point selection.

262 Chap. 11 Timing Circuits

To reduce duty-cycle distortion, other configurations can be imple­

mented, such as dual delay lines: one for rising edge clock and one for fall­
ing edge clock. Matching between two delay lines and excessive power
consumption are concerns. A method of phase-locking to both the rising
and falling edges of the external clock can be accomplished by comparing
the falling edges after the rising edges of the reference and feedback signal
are locked. The increased complexity of such an implementation, by using
an extra DLL for duty-cycle correction, is anticipated when a high-perfor­
mance ADDLL is required for a high-speed DRAM.
11.2.2.2 Lock range and input phase selection Lock range is also
referred to as the tuning range of the DLL,立改％ which is the only design
variable in Equation (11.3), When the system clock period («) is changed,
tock should be able to adjust the delay in the worst-case situation. For
instance, the number of total effective delay stages, N, is determined by the
lowest operating frequency. Assume that t(jK is equal to 5ns and delay per
stage (勿)is 100 ps (fast comer: high voltage and low temperature), then N
can be calculated to 50, according to Equation (11.4). Taking "ntri into
account, the number of effective delay stages is still close to 50, depending
on Hntri and process variations. Equation (11.5) can be used to estimate the
N used in the digital delay line.

N 二 ((CK, slow - hNTRlJast^^DJast

(U.5)
While both frequency and process are factored in during initial lock, the
delay line still needs to be able to track voltage and temperature (VT) varia­
tion. As shown in Table 11.2, the intrinsic delay can vary as much as 40% at
lower voltage and higher temperature. The delay line should have sufficient
tuning range to cover such variations. Some other applications may require
a frequency change over a long period of time (called frequency slewing)
for EMI reduction. Depending on the initial lock and lock point, the delay
line needs to accommodate such frequency, voltage, and temperature (FVT)
changes at the cost of extra intrinsic delay, power, and area.
The initial phase relationship between the internal clock and the feed­
back signal is uncertain at power up because of delay variations. After
power up and reset, the delay line should be set to known initial conditions,
which establish a unique clock propagation path. In general, we want n from
Equation (11.4) as small as possible in order to reduce power dissipation
and increase noise immunity. Therefore tock should be limited to one clock
cycle Uck) after lock, even though the delay line could provide a delay
range more than two times the clock period (in some literature, ifs referred
Sec. 11.2 All-Digital Clock Synchronization Design 263

to as false lock when 立农后 is greater than 1酸).Assume N=50,工或=2.5 ns,

and S=200 ps at the slow comer, the tuning range can be upto 10 ns, which
is four times the clock period at higher speed.
For the register-controlled delay line [1, 2], the initial entry point is
toward the right end of the delay line fbr the lowest intrinsic delay. To meet
this design criterion, the phase detector (PD) should be modified to disable
shift-right and, under that condition, initially only add delay to the delay
line. However, setting the initial entry point at the right end of the delay line
is undesirable. One scenario follows: initially tcK ~ tiNTRb and the DLL is
locked with SocL也 During the Read operation and under PVT variations,
Hntri becomes greater than tcK- The DLL wants to compensate for this
change and instead of removing delay, it must add delay until Hntri + ^Lock
equals 2 •
The DLL relocking requires more time, in this case, and creates timing
uncertainty between the internal and external clocks. Several spare delay
stages must be added to allow the DLL to subtract delay after the initial lock
even if the lock point is close to the initial entry point. The spare delay
stages, however, add more intrinsic delay 也而 and noise sensitivity to the
timing path.
A better solution fbr high-speed DRAM is to permit the delay line to
always add delay and to centralize the lock point initially. Depending on the
initial conditions between the reference and feedback signal, the input to the
delay line can be phase shifted to reduce the insertion delay and improve the
efficiency of the delay line. The phase shift can be 1800 (CLK180), as
shown in Figure 11.5, which is merely an inverted version of the CLK sig­
nal. When the feedback signal (CLKFB) lags the CLK within a certain phase
shift (A), e.g., 20°, or CLKFB leads the CLK within (180-A), e.g,, 160°, the
phase-shifted signal (CIAT180) instead of the CLK can be used as the delay
line input and the lock point centralized in the delay line fbr better FVT
tracking. The insertion delay can be reduced from tcK“PEt0 fCK^CKD^PE-
Ignoring intrinsic delay, the worst-case number of delay stages can be given
by

N =，CK, s/ow
hjast' @

(116)
where the @ is the phase-shifting factor. For 180°, 0 equals 2. Using the pre­
vious example, t(jK = 5ns and g = 100 ps, the worst-case number of delay
stages needed is reduced from 50 to 25. The worst-case insertion delay is
264 Chap. 11 Timing Circuits

cut in half by simply adjusting the phase of the input signal. No spare delays
are needed with this arrangement.

Figure 1L5 Input phase selection to improve the efficiency of delay line.

If a clock divider is used for improved clock delivery, multiple 90-

degree, phase-shifted signals are available, as shown in Figure 11.6. The
efficiency of the delay line, including power, layout area, and lock time can
be further improved by selecting one of the clock signals (CKO, CK90,
CK180, and CK270) as the input to the delay line. Clock selection only
occurs once as the system is reset and can be disabled to save power after
the selection is made. Using the prior example with an external clock period
of 2.5 ns and an internal cycle time of 5 ns (based on Equation (11.6)。= 4),
only 13 delay stages are needed, in theory, which is a 74% improvement
over the original design. Duty-cycle correction (for two phases) or phase
generation (for multiple phases) is required to retain all of the edge and tim­
ing information contained in the external clock.
Another benefit of input phase selection is power saving. The average
dynamic power dissipation of the delay line is given by
2
P“g = 2F CdJcLK/DD

(H.7)
where n is the number of delay stages when the DLL is locked, Cj is the
capacitance associated with each delay element, rpp is the power supply
voltage, and fcLK is the operating frequency. Supply voltage is the biggest
knob with regard to power reduction. This reduction may cost some speed
and, due to smaller voltage swings, increase sensitivity to supply noise
(Table 1L2). Reducing the clock frequency with an internal clock divider is
another way to save power. If the clock frequency and supply voltage are
Sec. 11.2 All-Digital Clock Synchronization Design 265

fixed, the only design choices remaining are n and Cj. Since Pavg is propor­
tional to the number of active delay stages (n), power can be reduced by a
factor of(|)with appropriate input phase selection. To further save power
when the DLL is not in use (during active power down, for example), a dis­
able signal can be applied to stop the clock that is fed into the delay line.
The DLL, however, must reacquire lock after the clock is re-enabled.

CLK

CLK* -I： 一
I i
I

CLKO ；
*
CLK90 J
I

CLK180 ； ______
I

CLK270 ； _

Divide by 2

U
O
Q
O £
O
S
x
S,
布
o
o
o
Divide by 2

Feedback signal

Figure 11.6 A clock path with clock dividers and fburphase input signals.

11.2.23 Delay resolution andjitter For ADDLL, jitter is primarily

determined by the resolution the digital delay stages. The peak-to-peak jitter
(Jp.p) and RMS jitter (Jr痴 can be defined as
266 Chap. 11 Timing Circuits

Jp-P =±，Q，JRMS - y

(H.8)
For a traditional delay line made up of logic gates, the minimum gate
delay (an inverter) is approximately 50 ps. To preserve signal polarity, two
inverters are needed, and peak-to-peak jitter will not be any better than 100
ps. This jitter is evident at the device outputs whenever the DLL makes an
adjustment.
To improve resolution for high-speed operations, a finer delay element
can be used, which in turn reduces jitter. Two examples of a fine delay ele­
ment are shown in Figure 11.7: one with load adjustment and the other with
path selection. By adding or removing a small load capacitor, controlled by
a shift register, the delay can be varied dependent upon the selected load
capacitance and the driver. A pair of N and P MOSFETs with their source
and drain tied together, respectively, can serve as load capacitors. Delay res­
olution is determined by the unit load capacitance. The delay will be fairly
constant over PVT variation, and the intrinsic delay is small (basically two
inverters for eight stages in Figure 11.7). For the path selection delay ele­
ment, a delta delay will exist between the slightly different-sized inverter
strings. This topology offers better delay tracking over PVT variations. If
connected in serial, the fine delay line introduces a large intrinsic delay,
which is not good. An interleaved architecture using both interpolation
methods is commonly implemented.
IN — —『上
—>o— ¥ 1~~A- OUT Fast path 广、- kT二

capacitors y y Slow path j

Figure 11.7 Two types of fine delay adjustments.

Another approach is shown in Figure 11.8. This design combines the

outputs of two invert matrix blocks fed by two input signals (A and B) that
differ in time by one unit delay (typically 100 ps) from the traditional delay
line. By mixing the outputs together we create a delay block with fine edge
control and low intrinsic delay. Each matrix block consists of a multitude of
parallel, but differentiated delay paths. A control code determines the over­
all delay of each matrix and, in essence, the slew rate at their outputs. This
approach is called phase blending in some literature [12].
Sec. 11.2 All-Digital Clock Synchronization Design 267

Figure 11.8 Phase mixing by adjusting slew rate and driving strength.

The tuning range of the fine delay should cover the delay variation of
one coarse delay stage in either direction: up or down. The step size (unit
delay) of the fine delay line (茄)is a fraction of the coarse unit delay (力).
A separated loop is required to perform fine delay adjustment in addition to
the conventional DLL loop, referred to as a coarse loop. Dual-loop architec­
ture provides an efficient way to achieve wide lock range (via a coarse loop)
and tight synchronization (via a fine loop). Smooth transition between the
coarse and fine adjustments is a challenge and may cause a large phase jump
or discontinuity of the DLL output.

11.2.3 Phase Detector (PD)

A phase detector (PD) is the decision maker in DLLs. The PD discrimi­
nates phase error between the reference and feedback signals and adjusts the
delay line accordingly. For an all-digital implementation:

1. No integration happens in the PD.

2. Mutual exclusive signals are needed to control the all-digital loop.
3. Such control signals should be well-defined digital signals.

One concern is that the resolution of a DLL is limited by the unit delay
stage, which needs to be factored into the PD design to prevent further
action if the phase error is within a certain boundary (sometimes referred to
as a window detector). Loop dynamics, including lock time, tracking time,
and stability are determined in large part by the phase detector design.
11.23.1 Lock time Lock time, measured in clock cycles, is another
important design aspect. Two situations must be considered:
268 Chap. 11 Timing Circuits

1. After a DLL is reset and enabled, how long will it take to initially
lock the loop before a valid Read command can be issued?
2. After initial lock (if the lock is lost due to voltage and temperature
drift), how long will it take to regain the lock condition?
To answer these questions, we must determine the sampling frequency at
which a digital phase detector (DPD) will make DLL adjustments. For a
high-speed, low-voltage memory system, the intrinsic forward path delay
(hNTRl) could be several times the clock period, as shown in Table 11.1 and
Table 11.2. The loop must be allowed to settle down after a delay adjust­
ment (the input signal propagates through the delay path and feeds back into
the DPD, sometimes referred to as loop settling time). To achieve this, the
sampling frequency 向 must not exceed

fs = r~k
lCK长

(11.9)
where the sampling factor k can be chosen according to

k _ j^j/hNTRI, slow ^Lock,slow\ +女

' ^KJast ‘EVEN

(11.10)
The subscript slow refers to a slow corner condition, while MtcK,fast refers
to the highest operating frequency. The function INT(x)even makes the x to
the closest even integer, for example, x = 0.6 gives 0, and x = 1.4 gives 2,
The k()can be either 1 or 2. The side effect of this sampling factor is an
increase in the lock time by a factor of k.
Making adjustments beyond the desired sampling frequency causes sta­
bility problems for a feedback system. Because the sampling frequency is
determined by the highest operating frequency, it may be too pessimistic for
low-speed operations. Dynamically adapting the sampling frequency based
on clock frequency and intrinsic delay is possible by measuring the loop
delay under varying PVT conditions. As part of the DPD, a digital filter like
an up/down counter or shift register can be used to control the sampling fre­
quency.
With the help of the previous discussion on lock range and input phase
selection, the worst-case lock time can be determined by
Sec. 11.2 All-Digital Clock Synchronization Design 269

=N4-tcK
工 worst

(UJI)
where N is the number of total effective delay stages and is the phase
selection factor. If = 50, A： = 5, and(|)=1, it will take 250 clock cycles to
initially lock the loop, which may be unacceptable (DDR and DDRII
SDRAM usually specs 200 clock cycles for initial lock time). With input
phase selection, the lock time can be reduced to 125 cycles (with 180° phase
shift) or 65 cycles (with 90° phase shift,(|)= 4).
Another way to reduce lock time is to perform segment selection
(grouping several delay stages into a segment) prior to locating the entry
delay stage. After the segment is located, the final entry point can be found
within that segment. The number of delay stages grouped into a segment (S)
can be determined by

S = IN人;CKJast )
2 ' ^D,slow

(11.12)
For example, assume that the fast clock period is 2 ns and the delay per
stage at slow comer is 200ps, then S is equal to 5. The lock time with seg­
ment selection is given by (worst-case):

worst ~ (w + 9 " ‘CK

(11.13)
For the same parameters used in the previous example, the new lock time
calculates to 75 clock cycles, which is four times faster than the lock time
without the segment and phase selections. For a high-speed system, the ratio
of tcK to S is small, somewhat reducing the benefits of segment selection.
Initial lock time can also be reduced by performing a binary search with a
successive approximation register, or SAR. Local feedback (bypassing the
clock tree and I/O model delays) can speed up the initial adjustment by
making continuous changes to a dummy delay line. Also ifs possible to
make loop adjustments proportional to the magnitude of the phase error
without impacting loop stability. This would require an adaptive PD capable
of selecting big (several stages) or small (one-stage) adjustments based
upon phase error.
270 Chap. 11 Timing Circuits

Another method fbr quick initialization is the synchronous mirror delay

(SMD), as shown in Figure 11.9. Two delay lines are used in the SMD: one
for measuring the timing difference between the clock period and the intrin­
sic delay (I/O and clock tree) and another fbr the proper insertion delay
based upon mirroring the timing measurement made in the first delay line.
The SMD can generally acquire lock within two clock cycles at the cost of
an extra delay line. For high-speed operation, the adjustment may require
several cycles to be evident at the output. The delay model is a significant
challenge fbr an open-loop sync circuit such as this.

Figure 11.9 Block diagram of a synchronous mirror delay (SMD).

A closed-loop quick initialization scheme, called measure-controlled

delay (MCD), is illustrated in Figure 11.10. During the initialization, the
internal reference clock (CLKREF) is selected by a MUX as the output sig­
nal. The feedback signal (CLKFB) is sampled inside the measure delay line
by the CLKREF. The timing difference between CLKFB and CLKREF is
converted into digital code by taking a snapshot of each measure delay line
output. The digital code stored in the control logic is mirrored to select the
proper entry point of the DLL delay line. After initial locking and loop set­
tling, the output of the regular delay line is picked as the output, and the
measure delay line is disabled fbr power saving. Normal DLL operation is
resumed by using a phase detector (not shown) to track the VT variations.
The initial lock time is like that of the SMD (just a couple of cycles), which
makes it attractive for quick recovery from power-saving modes.
The lock time for the SMD or MCD is fixed and independent of clock
frequency. With dual-delay line structure, the efficiency of the delay line,
especially the area, can be further improved by using input phase selection.
Any mismatch that exists between the two delay lines can be mitigated by
adding a PD to the MCD, as previously mentioned.
Sec. 11.2 All-Digital Clock Synchronization Design 271

Figure 11.10 Block diagram of a measure-control 1 ed delay (MCD).

After initial lock, a loop can lose lock due to excessive frequency, volt­
age, and temperature (FVT) variations. Generally, a DLL can dynamically
track out these variations and adjust delay to relock the loop. Depending on
the sampling frequency and the magnitude of the phase enor, the tracking
time may vary. After initial lock and during memory Read operations, it's
undesirable to make a big delay change that appears as timing jitter at the
outputs. Rather, small adjustment steps, which improve jitter performance,
tend to extend tracking time. While sync circuits are designed to track long­
term FVT variation, proper filtering is important to ensure loop stability.
1123.2 Circuit implementations For a coarse delay line, the best
resolution is a two-gate delay. The delay line is primarily used to ensure
adequate tuning range and proper lock. A PD block diagram used for course
adjustment is shown in Figure 11.11.

CLKDLL

CLKIn

Unit

Figure 11.11 Block diagram of a PD used for coarse adjustment.

To meet the setup time requirements needed by the input edge-triggered

D flip-flop (DFF), the unit delay in Figure 11.11 can be increased above the
unit delay used in the delay line to prevent jittering (bouncing back and
forth in the delay line) under all conditions. A buffer for the CLKIn, which
272 Chap. 11 Timing Circuits

is a delay approximately one-half of the unit delay, is useful to improve the

jitter performance and helpful to catch correct data for the DFFs when the
DLL is close to lock. The sampling frequency (Equations (11.9) and
(11.10)) for this PD depends on the operating frequency and intrinsic delay
of the loop.
A timing diagram in Figure 11.12 illustrates three different situations:
shift-left, shift-right, and adjusted (locked). An up-down counter or shift
register can be added as a digital loop filter to average out the intermittently
generated QL/QR to further stabilize the loop operation. After initial lock,
the sampling or filtering rate can be adjusted for improved tracking capabil­
ity. This DPD can also be frozen (no shifting) by disabling the sample clock
to prevent roaming caused by power supply noise and jittering due to coarse
shifting. For high-speed phase detection (HSPD), the feedback signal
(CLKDLL) may distort and cause malfunctions to occur. It is a always good
practice to divide the clock down internally for high-speed phase detection.

Three states of the all-digital PD

cwn j_Ln_nj-
CL3
Sample
clock (k=2) n
Less delay, More delay. Adjusted,
Shift-right
Shift-right Shift-left No change

Figure 11.12 Timing diagram of a coarse DPD for an ADDLL.

An arbiter-based fine PD (FPD) in Figure 5.30 can be used to discriminate

small phase error for a fine loop. The arbiter-based FPD can identify phase
error as small as 10 ps with mutual exclusive outputs. A digital filter can
also be used in this FPD to provide stable operation. A window detector can
also be configured using two FPDs, as shown in Figure 11.13. If the feed­
back is within a delay window (defined by the fine delay Z/rp) from the ref­
erence, the decision made by each FPD will be different and no fine
adjustment (shift-right or -left) should be made.
Sec. 11.2 All-Digital Clock Synchronization Design 273

Figure 11.13 Arbiter-based phase detector with a fine window detector.

The FPD is active only when the coarse loop is locked. The coarse PD
can take over the control to correct large phase error while the fine loop is
active. Extra time is needed to lock a fine loop. Similar searching algorithms
to those previously described can be used if a large number of fine delay
stages are used.

11.2.4 Test and Debug

Test and debug for ADDLL circuits is very important, and proper test
functions must be integrated into the design for both coarse and fine PDs.
The most important test issues include loop initialization, loop dynamics
(lock range, lock time, and jitter), and tracking capabilities. Most test func­
tions can be implemented by using test modes and fuses, Test-fbr-speed is
generally a big challenge and requires a compromise between test resources
and functionality. Suggested procedures and techniques follow:

1. Characterize the coarse or fine delay line by disabling the respective

PD and performing manual shifting.
2. Monitor the loop dynamics by outputting shift control signals to des­
ignated output pins for packaged parts. The most important signals
include DLL lock, shift controls (shift-left, -right), etc.
3. Check the timing specs under different PVTF conditions by monitor­
ing a free-running output data strobe.
4. Check different initialization schemes by enabling or disabling differ­
ent functional blocks, for example, MCD initialization, input phase
selection, and segment selection. It is desirable to share the same tim­
ing path with different control logic for evaluation, test, and debug.
5. Build a DLL bypass path for low-speed functional test.
274 Chap. 11 Timing Circuits

6. Check various power-saving modes versus DLL functionality. Lock

time and loop dynamics can be observed in Step 2.
7. Adjust I/O models to meet 碗 timing specs. Static adjustment can be
taken outside of the DLL intrinsic path as illustrated in Figure 11.14.
8. Check DLL tracking capability (track time, jitter, etc.) by varying
voltage and temperature (VT). The VT tracking can be tested in three
configurations: coarse loop only, fine loop only, and dual-loop (both
coarse and fine) at various frequencies. Different sampling frequen­
cies can be analyzed based on the clock speed.

Figure 11.14 Block diagram of a dual-loop ADDLL with t4c trim.

11.2.5 Dual-Loop Architecture

A DLL with dual-loop architecture generally has one loop for wide lock
range and one for better timing resolution. A transition occurs when one
loop (i.e., the fine loop) reaches its limit and interactions between the two
loops are necessary. Smooth transitions are essential fbr better performance.
Other concerns when designing a dual-loop DLL are tuning range, tracking
time, power, and area.
For an albdigital implementation, the dual-loop DLL is made up of a
coarse loop and a fine loop. Each loop has its own delay line and phase
detector (shown in Figure 11.14 and Figure 11.15). The initialization always
starts with the coarse loop. After the coarse loop is locked, the fine loop is
enabled to dial in the timing with its higher resolution capability. It is costly
to rely on only the fine delay fbr VT tracking. When the fine delay line
reaches its end, transitions must be made to activate the coarse adjustment
and reset the fine delay fbr further tuning. Depending on the coarse and fine
delays at different PVT corners, timing discontinuity can occur, which, in
turn, could violate timing specs due to clock glitches and jitter. It is not triv­
ial to monitor the ratio of the coarse and fine delays and dynamically update
and synchronize the transitions. Without resetting the fine delay line, phase
mixing between the two signals (one coarse unit delay apart, as shown in
Figure 11.8) is a preferred method fbr producing seamless transitions, that
is, provided that a proper control signal can be timed to select the clock
entry point in the regular coarse delay line.
Sec. 113 Mixed-Mode Clock Synchronization Design 275

Another preferred method uses dual-delay paths, as illustrated in

Figure 11.15: one path for active fine delay adjustment and the other path
for both coarse and fine adjustments offline. An extra fine PD can be used to
guarantee that the phases of the dual delay line outputs are close to align­
ment when switching. The internal time constant is smaller than the DLL
loop time constant, and the phase error can be adjusted quickly after a
coarse shift is made at the inactive delay line. The coarse adjustment will
only happen in the background (eg, the coarse delay 2 in Figure 11.15) and
won't impact the active timing path (eg, the coarse and fine delay 1 in Fig­
ure 11.15). The transition is seamless and occurs when:

1. The active fine delay hits a certain boundary.

2. The phase error between two delay lines is within one fine unit delay.
3. A switch signal (SEL2) is timed with the falling edge of the output
signal from the fine delay line.

The regular delay line gets adjusted only during initialization and won't
change after initial lock. The coarse delays in the dual-delay lines provide
the proper VT tuning range at the cost of extra intrinsic delay, power, and
area consumption.
Internal coarse shift

Figure 1L15 Dual-path fine and coarse adjustments with smooth transitions.

11.3 MIXED-MODE CLOCK SYNCHRONIZATION DESIGN

A basic delay-locked loop has only two building blocks: delay line and
phase detector (PD). The delay line provides an adjustable delay, and the PD
directs how to vary the delay. To synchronize data outputs with a system
clock in memory devices, an I/O model is also included in the feedback
path. For high-speed DRAM, extra circuits like clock dividers and edge
recovery circuitry (eg, a phase generator), as shown in Figure 1L16, must
be included to improve performance. The core DLL can be left intact, and
the design rules previously discussed for the digital DLL can remain the
same due to similar internal clocking speeds. However, with a clock divider,
276 Chap. 11 Timing Circuits

edge information for the internal clock must be recovered for full-speed I/O
operation. Multiple phases from the phase generator can represent the exter­
nal full-speed clock and are used to time the output data path accordingly.
An analog DLL with a voltage-controlled delay line {VCDL) is generally
suited for such applications. Combining both digital and analog implemen­
tations, a mixed-mode DLL is investigated in this section with a focus on
analog DLL design.

Figure 11.16 Block diagram of a high-speed, mixed-mode DLL.

As the timing window gets smaller and smaller for high-speed interfaces,
the focus shifts somewhat from just output timing to both input and output
timing. The entire data path, from command address capturing to data out­
put at specific CAS latency, must be taken into account for floor planning
and timing analysis. Design changes based on interactions between the data
path and timing circuits are possible. These may include the need for more
than one DLL, depending on the timing budget, flexibility, power, and area
constraints.

11.3.1 Analog Delay Line

Compared to a digital delay line (DDL), an analog delay line (ADL) has
a different configuration. The delay is adjusted by changing the delay per
stage (/£>), not the number of stages (n) (Equation (11.4)). The delay adjust­
ment is continuous for an ADL based upon an analog control signal while
adjustment for a DDL is discrete and based on the minimum delay resolu­
tion. A block diagram of a voltage-controlled delay line (VCDL) is shown in
Figure 11.17 with N delay stages. If the timing difference between CLKIn
and CLKOut is locked to one clock cycle (Jck)，then each stage can provide
a nominal t&N unit delay. For periodic signals, one cycle of delay provides
a 360° phase shift. Various phase shifts can be generated based upon differ­
ent numbers of N. The ADL is well suited for multiple phase generation.
Sec. 11.3 Mixed-Mode Clock Synchronization Design 277

Fewer delay stages make the ADL smaller than a DDL, but with compara­
ble tuning range for clock synchronization. Phase interpolation between two
adjacent taps is generally required to achieve full-range (360°) coverage and
precise de-skewing. For high-speed DRAM, analog implementations
become the only choice in some applications.

Figure 11.17 Block diagram of a voltage-controlled delay line (VCDL),

11.3.LI Unit delay Choosing a proper unit delay is the first step in
building a delay line. An analog control signal (e.g., a bias voltage) is gener­
ally applied to vary a current (Itail) used to charge or discharge internal node
capacitance (Cw). A unit delay (//)) can be defined as the time it takes an
internal node to be charged or discharged by the bias current /幻”.

(D = \dV
hail」

(UH)

A single-ended, current-starved delay buffer is shown in Figure 11.18. The

bias voltages /尸 and control the current source and sink, respectively.
G is the intrinsic capacitance associated with the input MOSFET MP and
MN (current stage and next stage), Cv can be a voltage-controlled capacitor.
A fully differential delay, also referred to as current mode logic (CML), is
also shown in Figure 11.18. A voltage-controlled resistor (/?£)can be real­
ized by a P-type MOSFET pair, which is biased on the triode region. A poly
resistor can also be chosen for better linearity and greater voltage margin at
the cost of larger layout area, matching, and process sensitivity. Within the
linear operating region, there is negligible duty-cycle distortion and better
common-mode noise rejection fbr the fully differential delay as compared to
single-ended. The signal swing doesn't need to be rail-to-rail for the differ­
ential approach. However, level translation is required to bring the signal
levels back to full CMOS.
278 Chap. 11 Timing Circuits

v
Figure 11.18 Single-ended (left) and fully differential (right) analog unit delays.

It is usually difficult to control the bias voltage over process, voltage, and
temperature variation. Sei住bias techniques [3], shown in Figure 11.19, can
be applied to the differential delay buffer for:

1. Better dynamic supply noise rejection.

2. Better linearity for delay versus control voltage.
3. Elimination of bandgap or other bias generators.

The control voltage (Vctrj) is fed from the phase detector and loop filter. The
bias voltage Vgp fbr the loads can be generated from Vctri with the aid of an
op-amp. A replica delay is used to generate Vbn for the current sinks. The
same current is mirrored inside the delay buffer to produce equal rise and
fall times. For a self-bias fully differential buffer, the unit delay can be
given as

f 1 _ a _ g
D n £ J=T^l 3dd/Mtp)

(1L15)

V/p is the threshold voltage and k is the transconductance of the PMOS

device. The unit delay has better supply noise rejection if self-bias tech­
niques produce a constant current.
Sec. 11.3 Mixed-Mode Clock Synchronization Design 279

Figure 11.19 Self-bias generators with a replica delay buffer.

The unit delay (measured from the crossing points of the differential
signals) versus control voltage at different process and temperature comers
is plotted in Figure 11.20 for a clock speed of 625 MHz and a supply voltage
of 1.2 V. A slow process comer (SS) and lower temperature (U0°C) make
the delay tuning range smaller due to limited voltage margin (higher thresh­
old voltage). The signal may get severely attenuated due to sluggish transi­
tions and limited gain. A general rule of thumb is to keep the unit delay less
than 15% of the cycle time (tctd at worst PVT comers. Fast process comer
(FF) and high temperature (110℃) may cause problems for low-speed oper­
ation because the delay is entering a nonlinear operating region.

CO* 篙 4学 寺 等 5gsi 法 盅 饿 9
oocioodooddoooooo

Figure 1L20 Unit delay versus control voltage (Vctrl) at different comers.
280 Chap. 11 Timing Circuits

A comparison between digital and analog unit delays versus supply

variation is shown in Figure 11.2L CML-based analog delays have much
better supply noise rejection than its digital counterpart. Analog delays can
also be biased to achieve the same delay across different process comers,
that is, to achieve consistent latency control.

VCDL delay vs. ”

*DCAP: parasitic load for simulation

a)

*DCAP: parasitic load for simulation

b)
Figure 11.21 Supply noise sensitivity for both a) analog and b) digital delays.
Sec. 11.3 Mixed-Mode Clock Synchronization Design 281

11.3. L2 Gabi and lock range The gain for the self-bias differential
delay line is given as [3]

C n Cn n Cn
(j = 一 ・■，， = . 内 ■- ，■・，，■,- ，・
2" k-{VDD-Vctrl-VTP^

(11.16)

where n is the number of delay stages and Itaii is the bias current. The output
from each delay stage can swing from Vctri to V/)D・ The system may not be
stable when the Vctri is approaching Vdd-Vtp (minimum swing at the out'
put). Furthermore, Vctri must not exceed a specific threshold determined by
the process and operating conditions in order to keep the differential input
transistors in the saturation region.
The lock range of the voltage-controlled delay line (VCDE) is n • g,
where g is the unit delay defined in Equation (11.15). Although the number
of stages (n) for analog delay line is fixed, there is a minimum requirement
to keep the delay per stage 〃£>) within the linear region. If the transition time
is comparable to the bit time (half cycle), the output may not reach full
swing, and the signal could get attenuated so badly that no valid output
would be generated. The rule of thumb is to pick the n greater than eight.
Locking to a different clock phase can be used to eliminate this problem,
just like the input phase selection used in digital delay line. Instead of lock­
ing to 360-degree phase (a full tc/d of the VCDL, the differential feedback
signals can be swapped to achieve 180-degree (0.5 Qk) locking.
With the same clock frequency, the effective delay per stage is 50%
less, and the transition time is always less than the bit time, which provides
better noise margin. Figure 11.22 shows two lock regions under different
configurations: an eight-stage VCDL locking to 360° and a six-stage VCDL
locking to a 180° phase. It is obvious that the 180° lock has better noise
margin and more symmetric transitions (tpiH= tpH》However, duty-cycle
distortion from the input must be addressed to achieve precise phase align­
ment in this case.
There is also an upper limit fbr the VCDL stages. Lock range under dif­
ferent PVT comers generally requires a longer delay line at the cost of large
area and more power. A comparison between eight- and sixteen-stage delay
lines is plotted in Figure 11.23 with regard to slow simulation comers and
linear operating region. The lock range can vary from 425〜980 MHz (eight
stages) to 220〜500 MHz (16 stages). Fewer delay stages are desirable fbr
high-speed operation. To keep the gain constant, an adaptive scheme can be
applied based on the frequency and PVT comers. The minimum VCDL
282 Chap. 11 Timing Circuits

delay is compared to the full-speed clock to determine the internal clock

speed. The VCDL delay can be specified with a scaling factor m, which can
be controlled by a programmable clock divider The phase shift per stage is
adjustable based on the ratio of m/n.

联，D = m 小

(11.17)

Figure 11.22 Lock range and delay versus control voltage for two different VCDLs.

Delay versus Vcu (SS,1.25 V, 11 OX, 500 MHz)

Linear region
700 1200 16 stages]
6 8 stages
00
( 00 —Freq~*16
8 500
) 00
① 00
400 N
H
。

S
)
3 A
Q
A c
e B
2 n
a) b
a
。

i
1
LL

5 o k 5 5
47 52
SO )50
(mV

Figure 11.23 VCDL delay versus Vctri at different conditions.

113.1.3 Noise and jitter The VCDL with current-mode logic (CML)
has better supply noise rejection than the digital delay line with CMOS
logic. With 150 mV lower supply voltage (-10%), the CML delay line (five
stages with loads) has delay variation only +2.1%, while the CMOS inverter
Sec. 11.3 Mixed-Mode Clock Synchronization Design 283

chain (four inverters with loads) has +12.9% variation. Under the same con­
ditions, the level translation from CML (differential) to CMOS (single­
ended, CML2CMOS) has 10% delay variation. Depending on the output
swing, common-mode range and PVT comers, the CML2CMOS may gen­
erate duty-cycle distortion at the output. This is not an issue when only the
rising edges of the multiple phases from the VCDL are used fbr clocking.
For an all-digital implementation, supply noise contributes 76% of total tim­
ing variation (Table 11.2). A timing path with extensive CML delay ele­
ments is a good alternative for high-speed DRAM design.
To improve the jitter performance and noise immunity, several design
and layout techniques for the VCDL are addressed as follows.
• Be careful when selecting the delay element. A device with longer
channel length (L) has better matching and noise performance at the
cost of speed. Current loads and sinks usually use longer L devices.
Devices with lower threshold voltage offer reduced phase noise and
improved voltage margin.
• Remember that large tail currents improve speed and reduce
noise and gain (Equation (11.16)) at the cost of power. Trade-offs must
be made between speed, lock range, gain, power, tracking bandwidth,
and the number of delay stages.
• Match the layout. Differential inputs use an interdigitated or common
centroid layout [8]. Minimize parasitic capacitance and interconnects.
Group all identical circuits into a cell fbr better matching and LVS (lay­
out versus schematic). Keep every delay cell oriented the same, and use
the same layer of interconnects. Identical circuits can be used fbr load
and voltage-level matching at the cost of power.
• Isolate and shield all analog signals, i.e.s Vctri. etc. Shield all high­
speed matched clocks. Put guard rings around critical analog blocks
(VCDL, bias generator, etc.). Use quiet and isolated power buses for
critical analog circuits. Avoid digital routes across sensitive analog
blocks.
• Placement is another key. Keep the clock output buffers (digital, high­
speed) far away from the CML delay line (VCDL) and level translators
(CML2CMOS). Identify and partition the circuits into high-speed
clocks, digital control logic, and sensitive analog circuits. Dedicated
power pins fbr analog circuits are imperative fbr better performance.

11.3. 2 Charge-Pump Phase Detector (CPPD)

To generate an analog bias voltage for the VCDL, a charge-pump phase
detector (CPPD) is generally used along with a loop filter, shown in
284 Chap. 11 Timing Circuits

Figure 1L24. Based on the outputs of the phase detector, a current (Ipump)
can be pumped into or out of the loop filter (LF), which is a simple capaci­
tor for the DLL. When the loop is locked, there is no net change in charge
within the LF, and Vetri remains relatively constant The CPPD shown in
Figure 11.24 is prone to charge-sharing problems. A fully differential
charge-pump with a unit-gain source follower [4] is useful and shown in
Figure 11.25. A current-steering technique [6] can also be used to reduce
ripple at the LF output and improve jitter performance. To avoid a “dead
zone” [3-4] when a phase is close to lock, keep the up-and-down pulses
from the PD the same, with the same duration, depending on PVT varia­
tions. The best resolution available from a CPPD depends on the control
voltage ripple, loop capacitance, matching etc., and is usually 10-20 ps fbr
all practical purposes.
Phase comparisons can be made on both rising and falling edges to
reduce lock time given there is little duty-cycle distortion. The lock time for
the CPPD is usually longer compared to the digital DLL, depending on the
Ipump and Q尸 A large loop capacitor usually improves loop stability (a
first-order system with a pole provided by the LF) but slows down the lock­
ing procedure. Based on initial conditions and clock frequency, a typical
lock time fbr an analog DLL is approximately 50〜100 cycles. If no clock is
running during power-saving modes, there is a “run-away“ problem for the
bias control signal (%〃/), which is generally floated. A digital-to-analog
converter (DAC) can be used to store the bias information and reduce the
relock time after exiting power-saving modes. The extra complexity, power,
and layout area must be traded against performance gains.

Figure 11.24 Block diagram of a charge-pump PD (CPPD).

Sec. 11.3 Mixed-Mode Clock Synchronization Design 285

Figure 11.25 Charge sharing and a potential solution for CPPD.

A potential "fhlse lock” problem can exist with the VCDLy in which the total
VCDL delay equals a multiple of clock cycles. False-lock detection is
needed for the CPPD to provide correct phase generation. Initializing the
VCDL with minimum delay is another way to prevent false lock at the cost
of longer lock times under certain conditions. 6tFast-down^^ mode is useful
to speed up the initial phase acquisition by ignoring the CPPD until a certain
phase relationship is reached. Loop stability must be ensured with regard to
lock time. When a digital DLL (DDLL, for synchronization) and an analog
phase generator (APG, fbr edge recovery) are both used in a mix-mode
clock synchronization system, a lock signal can be generated to start the
DDLL synchronization when the phase difference of the APG is within a
certain timing window. The overall lock time fbr the mixed-mode DLL is
longer compared to each individual DLL.
The loop response time fbr the APG is approximately one cycle because
the feedback is taken from the output of the last delay stage. Except for the
CML delay line, the single-ended to differential converter, CML2CMOS
logic and clock buffers remain subject to supply noise. Voltage and temper­
ature variations are tracked quickly, and tracking time is usually within two
cycles.
For the mixed-mode DLL in Figure 11.16, any clock manipulation in
the DDLL, like input phase selection and MCD locking, impacts the input
signal to the APG, and, therefore, changes the loop dynamics of the APG. To
ensure smooth transitions and stable operations, the CPPD is blocked from
any control voltage changes, and transitions are controlled by the digital
loop. The “block” function is also good for test and debug, when an open­
loop operation is required to manually adjust the control signal and charac­
terize the VCDL.
286 Chap. 11 Timing Circuits

Finally, a layout presentation of a mixed-mode DLL is shown in Figure

1L26. Notice that the APG is approximately one third as large as the DDLL
in this example. The total area is roughly 400 x 320(im2.
Analog phase generator Digital DLL

VCDL

Figure 11.26 Mixed-mode DLL layout example.

11.3. 3 Dual-Loop Analog DLL

If only the analog delay line is used for clock synchronization, some
restrictions may occur due to a limited tuning range provided by the fixed
number of delay stages. From Table 11.2, intrinsic delay variation through
the I/O path and clock tree can still be several nanoseconds depending on
voltage and temperature. A semi-digital, dual-loop DLL is reported in [9],
where the coarse adjustment is performed by phase selection, and fine
adjustment is achieved by phase interpolation. An analog phase generator
(APG) generates multiple phases based on a clock input. Phase selection
and interpolation are controlled digitally by a finite stage machine (FSM).
The VT tracking is performed by walking through different phases gener­
ated by the APG and interpolating between two adjacent phases. A current
can be mirrored equally through different legs and selected by a digital code
(i.e., thermometer code, 11111000). Although there may still be an uneven
phase shift through phase interpolation, especially across the boundary, the
perturbation can be small (in the range of 10 ps, fbr example) compared to
other timing variations (i.e., I/O model mismatch). The tracking time is
comparable to the digital solution, which is proportional to the step size and
loop time constant.
If the internal clock speed is slowed down by a clock divider fbr better
clock distribution, the dual-loop structure can still be used with a slave
delay line, as shown in Figure 11.27. The analog phase generator (APG) is
locked first to provide bias information to the analog DLL. The slave VCDL
Sec. 113 Mixed-Mode Clock Synchronization Design 287

in the analog DLL has the same structure as the main VCDL in the APG
The phase control logic can compare the feedback signal with a reference
signal (CLKIn) and generate control signals fbr phase selection and interpo­
lation. Compared to an all-digital implementation, the analog dual-loop
DLL has less supply sensitivity due to the CML delay elements. If the glo­
bal clock distribution can also be implemented with CML buffers, the sup­
ply noise sensitivity can be further reduced. Intrinsic delay can also be
decreased with CML buffers, which help to reduce loop response time and
latency. A similar scheme has also been chosen fbr high-speed data captur­
ing and per-pin de-skewing fbr gigahertz links (receiver and transmitter
designs) [10].

Phase control logic

UOAnqEwp ipoo
Phase selection &
interpolation

VCDL (slave)

Charge OP
pump PD OE
Local clock tree s

Analog phase generator

lb output path

Figure 11.27 Block diagram of a dual-loop analog DLL.

For input timing in high-speed DRAM, the external system clock, or

data strobes, are usually center-aligned with command, address (CZ4), or
data, respectively, at input pads. It is critical to match the C/A and data paths
with the clock/strobe paths for C/A and data capturing. Because received
clocks must be delivered to different locations for capturing, a clock tree is
inevitable and introduces latency for input timing. To match the timing of
the clock distribution, dummy buffers must be inserted into each command,
address (C/A), and data path to maintain the timing relationship. Extra
power, area, and delay are added with the matching scheme. Another con­
cern is the Write latency, which defines when the data is available in clock
cycles after issuing a Write command.
From the external system and channel, the timing variation《dqss) of
one quarter of the cycle exists between the system clock and the data strobe.
288 Chap. 11 Timing Circuits

Internal mismatch between the clock and strobe distribution paths degrades
the timing even further Known relationships between a decoded Write
command and captured data and a decoded Read command and output data
must be established by clock synchronization circuits. Read-Write latency
control circuits can be timed by the respective sync circuits. Clock domain
crossing, ie, command clock to data strobe or command clock to output
(Read) clock, must be circumvented for various PVTF comers.

Data u
o

板
latch D 03 Q
q ajP o
E O
p
S E l
nq-nduD
- p

<
p

Ae

0 ⑥ iso
0 o
o

C/A
latch

Figure 11.28 Dual-loop DLL for command, address (C/A) and data capture.

An analog DLL can be used to hide clock distribution delay and gener­
ate quadrature clocks fbr data capturing, if needed. In Figure 11.28, internal
clocks are center-aligned with received data by phase selection and interpo­
lation. C/A and data are delivered directly to capture latches after they are
received. A training sequence may be required to properly align the data
with internally generated clocks. Loop dynamics like tracking, filtering, set­
tling time, etc., must be carefully analyzed. For a feedback system, instanta­
neous jitter or skew cannot be corrected by a DLL; therefore, jitter caused
by the supply noise has the biggest short-term timing impact. Voltage regu­
lation may be helpful for a sensitive timing path. Differential signaling is a
plus fbr better supply noise rejection. As the data rate keeps cranking up,
very precise de-skewing fbr both input and output may be inevitable. A
mixed-mode DLL is a good basis fbr high-speed clock synchronization
design.
Sec. 113 Mixed-Mode Clock Synchronization Design 289

11.3.4 Mixed-Mode DLL and Its Applications

A comparison between a dual-loop digital DLL (DDLL) and an analog
DLL (ADLL) is illustrated in Table 11.4. For high-speed DRAM, an ADLL
takes a lead over a DDLL in several categories: lock range, supply noise
rejection, duty-cycle correction, and multi-phase generation. Longer lock
time and added complexity make it unattractive for mainstream DRAM. A
mixed-mode DLL (MDLL) can be configured with a DDLL for clock syn­
chronization and a single-loop analog DLL configured fbr multiple phase
generation. This combination makes multiple phase generation possible
when the internal clock is divided down. An MDLL is a good compromise
between speed and complexity in digital and analog implementations.
One of the most appropriate applications fbr an MDLL is high-speed
clock distribution. In Figure 11.29, a clock divider slows down the internal
clock (CLKIn) fbr propagation through the digital DLL (DDLL) and the
main trunk of clock tree (inverters with loads representing routes) for better
clock and power delivery. A single low-speed clock (DllOut) from the
DDLL is distributed over a long distance with less attenuation, power, and
mismatch. An analog phase generator (APG) is used fbr edge recovery and
placed close to local clock trees, which make matching much easier. Fewer
gain stages (i.e., inverters) are required to drive a low-speed clock over a
long distance. The additional APG is justified by reducing the power and
area fbr the main clock distribution and digital DLL.
290 Chap. 11 Timing Circuits

Table 11.4 A comparison between dual-loop DDLL and ADLL,

DDLL ADLL

Lock range Limited by delay stages Unlimited

Lock time Short Moderate

Dual-loop architecture Yes Yes

Resolution 〜20 ps 〜10 ps

Phase interpolation Option (digital control) Yes (digital control)

Layout area Large Small

Intrinsic delay Large Small

Supply noise rejection Poor Good

Duty cycle correction Extra loop required Inherent

Multiple phase generation No Yes (analog)

Stability Unconditional stable 1 st order, stable

Test and debug Easy with experience Difficult

Portability Good Moderate

Process scalability Good Marginal

Figure 1L29 A mixed-mode DLL fbr clock distribution.

Sec. 11.4 What's Next for Timing 291

Simulation data fbr a DDLL and an MDLL are compared in Table 11.5
with a typical process, 1.5V supply voltage, and 85℃ temperature. The
external clock is at 800 MHz. A full-speed clock runs in the digital DLL
without an APQ and a half-speed clock with two APGs was used in the
mixed-mode DLL. Even with two analog phase generators (APG) in the
MDLL clock trunk, the power is less than that of the DDLL with full-speed
clocks. For the digital DLL itself 40% power saving is achieved. However,
lock time fbr the MDLL increases significantly due to analog phase locking.

Table 11.5 Lock time and power consumption for DDLL and MDLL at 1.5 V TT, 85℃.

DLL Current Clock Trunk Local Clock Lock Time

(mA) (mA) Trees (mA) (ns)

Digital DLL 14 13.3 50 56

Mixed-mode 8.2 13.1 50 146

DLL

If the part is down binned to a lower speed grade, full-speed clocking can
resume and one of the analog phase generators (APG) can be used as an
analog duty-cycle corrector (DCC). Only two phases are generated by the
APG: clock and the 180° phase-shifted clock. Compared to a digital DCC
(another DDLL, 70% area), an APG is small (30%) and has less intrinsic
delay. Better supply noise rejection and quick tracking make an APG a good
DCC fbr clock frequencies anywhere between 400 and 800 MHz.
Finally, a 90° phase-shifted data strobe can be generated by the MDLL.
The strobe can be center aligned with the output data rather than edge
aligned, which makes it easy to capture data at the controller site. The APG
should provide a 45° phase shift fbr a divide-by-two clock or a 90° phase
shift for a full-speed clock. If an MDLL is also used for input timing, an
internal 90° phase-shifted clock can be generated fbr data capture.

11.4 WHAT'S NEXT FOR TIMING

In this chapter, we explored different timing circuits, from a simple digital
DLL to a more complicated mixed-mode DLL. We also investigated various
design issues associated with different architectures and presented examples
of circuit implementations. Based on the application and I/O specs, good
circuit design can reduce timing uncertainty, minimize clock skew, and
improve I/O performance. However, timing circuits also add complexity to
I/O design, increase power, grow layout area, and lengthen debug time.
Moving forward, as DRAM processes scale down into the deep submi­
cron regime, designing high-speed, low-power I/O interfaces becomes a
292 Chap. 11 Timing Circuits

huge challenge. High-performance phase-locked loops (PLLs) may replace

DLLs as alternative timing solutions. While clock data recovery (CDR)
using a PLL provides lower pin counts and strobe-less I/O timing, it puts
more constraints on data patterning. Slower transistors and a limited num­
ber of metal layers make the design even harder. Mixed-signal design might
be a key fbr the future I/O circuitry, in which analog core (i.e., voltage-con­
trolled oscillator, receiver, and transmitter) and digital support circuits (i.e.,
calibration, interpolation, and tuning) co-exist. Circuit partitions, power
bussing, and verification are important fbr achieving better performance.
On the other hand, extensive training and calibration from the memory
controller may be necessary to offload some of the burden of the timing cir­
cuit design from memory devices. At the same time, minimizing on-die tim­
ing variations is critical fbr this scheme to work. Clock distribution, clock
domain crossings, and latency control remain important issues fbr high-per­
formance memory data path design. In Chapter 12 we try to address some of
these issues via clever logic design.
An advanced memory buffer (AMB) used in a fully-buffered, dual in­
line memory module (FB-DIMM) provides another possibility in that tim­
ing circuits and high-speed I/O are moved into the AMB. The AMB, fabri­
cated with logic processes, serves as an intermediate buffer between the
memory controller and DRAM devices. An AMB supports a narrow point-
to-point channel enabling high-speed communication between the host con­
troller and the AMB while also supporting high bandwidth, wide channels
between the AMB and multiple memory devices. Obviously complication is
the price paid fbr higher performance.

REFERENCES

[1] A. Hatakeyama, H. Mochizuki, T. Aikawa, M. Takita, Y. Ishii, H. Tsuboi, S.

Fujioka, S. Yamaguchi, M. Koga, Y. Serizawa, K. Nishimura, K. Kawabata, Y.
Okajima, M. Kawano, H. Kojima, K. Mizutani, T. Anezaki, M. Hasegawa, and
M. Taguchi, UA 256-Mb SDRAM using a register-controlled digital DLL,”
IEEE Journal of Solid-State Circuits, vol. 32, pp. 1728-1732, November
1997.
[2] F. Lin, J. Miller, A. Schoenfeld, M. Ma, and R. J. Baker, <4A register-
controlled symmetrical DLL fbr double-data-rate DRAM,” IEEE Journal of
Solid-State Circuits, vol. 34, pp. 565-568, April 1999.
[3] J. Maneatis, ''Low-jitter process-independent DLL and PLL based on self­
biased techniques/' IEEE Journal of Solid-State Circuits, vol. 31, November
1996.
Sec. 11.4 What's Next for Timing 293

[4] A. Efendovich, Y. Afek, C. Sella, and Z. Bikowsky,. ^Multifrequency zero­

jitter delay-locked loop,“ IEEE Journal of Solid-State Circuits, vol. 29, pp.
67-70, January 1994.
[5] T. Saeki, Y. Nakaoka, M. Fujita, A. Tanaka, K. Nagata, K. Sakakibara, T.
Matano, Y. Hoshino, K. Miyano, S. Isa, S. Nakazawa, E. Kakehashi, J. M.
Drynan, M. Komuro, T. Fukase, H. Iwasaki, M. Takenaka, J. Sekine, M. Igeta,
N. Nakanishi, T. Itani, K. Yoshida, H. Yoshino, S. Hashimoto, T. Yoshii, M.
Ichinose, T. Imura, M. Uziie, S. Kikuchi, K. Koyama, Y. Fukuzo, and T.
Okuda, t(A 2.5-ns clock access, 250-MHz, 256-Mb SDRAM with synchronous
mirror delay,IEEE Journal of Solid-State Circuits, vol. 31, pp. 1656-1665,
November 1996.
[6] K. Wong, E. Fayneh, E. Knoll, R. Law, C. Lim, R. Parker, F. Wang, and C.
Zhao, "Cascaded PLL design for a 90nm CMOS high-performance
microprocessor,v ISSCC Digest of Technical Papers, pp. 422-424, February
2003.
[7] F. Lin, “Research and design of low jitter, wide locking-range all-digital
phase-locked and delay-locked loops/* dissertation, March 2000, University of
Idaho.
[8] R. J. Baker, CMOS Circuit Design, Layout and Simulation, 2nd ed., Hoboken,
NJ: IEEE Press, 2005.
[9] S. Sidiropoulos and M. A. Horowitz, semi-digital dual delay-locked bop,”
IEEE Journal of Solid-State Circuits, vol. 32, pp. 1683-1692, November
1997.
[10] E. Yeung and M. A. Horowitz, UA 2.4 Gb/s/pin simultaneous bidirectional
parallel link with per-pin skew compensation,IEEE Journal of Solid-State
Circuits, vol. 35, November 2000.
[11] B. Razavi, Monolithic Phase-locked Loops and Clock Recovery Circuits:
Theory and Design. Piscataway, NJ: IEEE Press, 1996.
[12] B. Garlepp, K. S. Donnelly, K. Jun, P. S. Chau, J. L. Zerbe, C. Huang, C. V.
 Chapter
12

Control Logic Design

12.1 INTRODUCTION
The control logic design for a DRAM could be considered the most
neglected area of DRAM design in the literature. One obvious reason for
such a lack of attention is that there are probably as many ways to imple­
ment the DRAM control logic as there are logic designers working on the
problem. We define the peripheral logic interface as the set of circuits that
interfaces the array access signaling protocol given in the data sheet with
the peripheral array logic described in Chapter 4. This logic must translate
command and address information coming from the input pins of the
DRAM device and provide control signals to the array peripheral logic. We
consider the peripheral logic interface to be the hub for controlling interac­
tions between the data sheet specifications and three distinct areas of the
DRAM die covered in previous chapters: the input circuit path, the output
circuit path, and the array peripheral logic. Figure 12.1 illustrates the rela­
tionship between each of these three areas of the DRAM highlighting that
the peripheral logic interface acts as the liaison between the peripheral array
logic and the DRAM I/O.
An important note about the peripheral logic interface is that the row
control is minimal out of this area. Wordline accesses or, what is being
referred to here as row accesses, are considered more a part of the periph­
eral array logic since this type of control does not depend on the I/O circuit
path logic for performance and generally does not require the same sort of
treatment that data accesses require fbr a high-performance part. For this
reason, we do not consider the wordline control, also commonly known as
the RAS chain circuitry in this chapter. Please refer to Chapter 2 for more
details about row access control and timing circuitry.

295
296 Chap. 12 Control Logic Design

Array

Peripheral array logic

Data Column Row

Address
valid access control

Figure 12.1 DRAM logic partitions.

With generational changes to the DRAM specifications come changes

in the input and output circuit paths. The I/O circuit paths must be tailored
to higher data rates and sometimes to changes in the command, data, and
address protocols. This is contrasted with the relative consistency in array
access signaling requirements for interfacing with the array peripheral logic
as discussed in Chapter 8. For commodity DRAM, i.e,, DDR, DDR2, etc.,
the column cycle times shown in Table 8.1 (repeated in Table 12.1) have
remained relatively constant through several generations. Any benefits from
process scaling are counter balanced by increases in memory density, which
results in a relatively constant die size. As process scales, voltage scaling
and interconnect delay become the overriding factors limiting array perfor­
mance. For the GDDR, high-performance devices listed at the bottom of the
table, column cycle time is reduced because of array and package bailout
differences, which are outlined in Chapter 8, resulting in higher bandwidth
(8.1). The control logic circuits bridge the gap between the bandwidth
requirements of the synchronous I/O and the asynchronous timing of the
array while, simultaneously, maintaining access delay independent of varia­
tions in clock period. As high-performance DRAM requirements demand
reduced column cycle times, the challenge for the peripheral logic interface
design is to be transparent to frequency changes. The goal is to ensure that
Sec. 12.1 Introduction 297

array access delay characteristics of the DRAM always limit device perfor­
mance.

Table 12.1 DRAM technology and speed relationships.

Min Column
Peak
Technology Burst Length Cycle Time
Bandwidth
flC€D)

SDRAM 1 6 ns 167Mbps/pin

DDR 2 5 ns 400Mbps/pin

DDR2 4 5 ns 800Mbps/pin

DDR3 8 5 ns 1.6Gbps/pin

GDDR3 4 2.5 ns 1.6Gbps/pin

GDDR4 8 3.2 ns 2.5Gbps/pin

Notice in Table 12.1 that burst length increases as bandwidth increases

resulting in a constant column cycle time. With more data delivered on each
column access, we can increase the I/O bandwidth without major changes to
the operation or architecture of the DRAM array. The amount of data trans­
ferred for each column access, expressed relative to each data I/O, is often
referred to as data Prefetch. For example, a device specified to deliver a
burst of eight data bits for each column access is said to have an 8n
Prefetch. If the same device has a data bus width of 32 bits and an 8n
Prefetch, then each column access reads or writes 256 bits of data. This con­
cept is further illustrated by Equation (8.1).
To understand how DRAM logic design differs from more traditional
logic design, we examine logic style choices that may aid performance and
reliability while working within the confines of an inexpensive DRAM pro­
cess. In the remainder of this chapter we discuss performance consider­
ations and design choices for the logical interfaces shown in Figure 12.1.
We first look at command and address capture and processing and how
choices in logic implementation can affect DRAM performance. We then
examine Write data capture and demultiplexing and how Write data proto­
cols for current high-performance graphics memory complicate logic and
timing circuit implementation. Finally, we discuss the Read data path and
how higher bandwidth demands have resulted in increased complexity fbr
the Read data pipeline and serialization circuits.
298 Chap. 12 Control Logic Design

12.2 DRAM LOGIC STYLES

Before we consider details of the control and data path logic design, we
must understand some of the difficulties that the DRAM designer faces
when making decisions regarding logic style and circuit topologies. Making
a choice in logic style primarily becomes choosing between using static
logic gates with clocked flip-flops or adopting a dynamic or domino logic
style employing latches and multi-phase clocks to gain clock skew toler­
ance. The use of dynamic logic and latches implies a departure from a more
common logic design flow involving circuit synthesis tools using hardware
description languages such as Verilog® or VHDL. As we progress through
this section, we see that the choice of logic methodology is driven by three
fundamental limitations:

1 . Low-cost process technology motivated by the commodity status of

most varieties of DRAM products.
2 . Delay variation fbr array column operations across process, voltage
and temperature comers.
3 . Performance and protocol requirements defined by industry stan­
dards and published in product data sheets.

These limitations tend to push us to a more custom logic design flow

rather than a more automated, abstracted approach. The simplicity of
DRAM operation makes gate count manageable fbr a full-custom flow.
However, this simplicity in operation is complicated by the protocol for
interfacing to the DRAM from the outside world and complicates timing
interactions between global circuit blocks on the DRAM die. Even though
synchronous-protocol interfaces dominate current DRAM technology, the
operation of the DRAM array has not changed, and the DRAM remains,
fundamentally, an asynchronous device.

1 2.2.1 Process Limitations

Design process technology used for DRAM has historically served as
an industry-wide development vehicle fbr process feature scaling. Due to
the commodity nature of the DRAM business, the research and develop­
ment costs related to process scaling have paid dividends because of
increased device yields per unit area of processed silicon. The process tech­
nology lead DRAM once enjoyed has eroded in recent years relative to pro­
cess development fbr processors and logic process nodes developed by
contract fabrication companies. The rush to develop finer process geome­
tries is somewhat tempered by the changes to MOS devices and intercon­
nect brought on by process scaling [3], Some of these changes can have
adverse effects on DRAM operation, which can lead to yield loss.
Sec. 12.2 DRAM Logic Styles 299

Logic design for a monolithic DRAM is made difficult because of low­

cost process requirements. Some of the “exotic“ materials and process steps
often used by microprocessor manufacturers are economically out of reach
for the DRAM logic designer. To maintain a reasonable cost profile, the
number of mask steps must be kept to a minimum.
There are two primary barriers to maximum logic circuit performance
as a result of the DRAM process and process scaling. First, high transistor
threshold voltage (尸加 is necessary to minimize DRAM leakage current.
This is an intentional process characteristic for maximizing data retention in
the array. This process characteristic bleeds over to the logic areas of the
device and limits performance of the logic gates. Because of the increased
产冰，gate delay increases, which is illustrated by the alpha power law

model [4] where the gate delay is partially expressed by:

, clvdd
t产------------------ a

(12.1)
where Q is the load capacitance, is the supply voltage, and a is the
velocity saturation index, which is close to one for a short-channel process.
In some cases, the problem of high Kyyys has been overcome by devel­
oping process technology with multiple gate oxide thicknesses and doping
profiles. By providing an option for thinner gate oxide devices in the
DRAM process, performance improvements have been realized for I/O cir­
cuits and peripheral logic interface circuits. Even with process improve­
ments, more aggressive logic styles, such as dynamic logic may be
necessary to overcome the process delay limitations caused by relatively
high Vth devices. Thinning gate oxides and more aggressive logic styles
have helped opened the door to high-performance DRAM devices. But
device scaling has not lead to the performance gains seen with process scal­
ing of logic processes for modem microprocessors [6].
Because of process scaling, changes in interconnect characteristics is a
second major difficulty faced by the DRAM logic designer. Figure 12.2
shows how scaling has resulted in a change in the aspect ratio of the inter­
connect metals on a single layer. Qualitatively, we must consider that the
capacitance is proportional to the dielectric constant, k, wire height, H and
space, S. Using low-k dielectrics to reduce sidewall capacitance is some­
times an option for more expensive logic processes, but this has not been a
cost-effective option for DRAM at this time. Please refer to [7] and [8] for a
quantitative treatment of this topic. Wire scaling results in increased capaci­
tive coupling, which in turn increases the potential for noise. Under worst-
300 Chap. 12 Control Logic Design

case conditions, when neighboring interconnect transitions occur in oppo­

site directions, the capacitance is close to double the statically calculated
capacitance. For example, the transitions on signals Al and A2 in Figure
12.2 "attack" the “victim," K This results in an increase in capacitive cou­
pling current and causes a longer delay on signal V. This scenario must be
considered when determining worst-case delay on a signal bus. The
designer is faced with a trade-off when designing signal busses. One exam­
ple would be the design of the peripheral data bus. Because a high-perfor­
mance DRAM must have larger data Prefetch depths for each data I/O. the
designer is tempted to take advantage of the increased wire density to keep
die size low and still provide the necessary data to maintain I/O throughput.
But because of capacitive coupling and loading, much of the potential inter­
connect density is used for increased space and/or static shields to help mit­
igate the effects scaling has on interconnect [23],

Previous process generation

Advanced process

Figure 12.2 Effects of scaling on interconnect capacitance.

「 H
C 0c

(12.2)
There are other more indirect effects of scaling that impact logic circuit
design. When process scales, voltage is often scaled proportionally. But the
VT^s of the DRAM process have not changed proportionally with scaling
and has therefore lead to a higher ratio of threshold voltage to supply volt­
age. Referring to Equation (12.1), when 々° is reduced there is an increase
in delay,力.Voltage scaling can affect noise immunity fbr dynamic circuits;
but this vulnerability is often offset by the relatively high K777S of the
DRAM process. We see that characteristics of a DRAM process are a hin­
drance to high-performance logic. But we must also consider how DRAM
performance specifications are dictated by the access delays of the DRAM
array. Next, we will consider specifications fbr array operation and how
these specifications affect the circuit topology decisions fbr DRAM logic
design.
Sec. 12.2 DRAM Logic Styles 301

1 2.2.2 Array Operation

Chapter 1 presents some of the generational changes to the command
and address signaling protocol that has occurred over the life of the com­
modity DRAM. Early DRAM protocol called for a broadside address bus,
where both row and column addresses are simultaneously applied for each
access. Second-generation DRAMs introduced the multiplexed address bus,
where the RAS* and CAS* signals latch the row and column address from
the same address signal inputs, respectively. Other developments that sig­
nificantly affect the design of the peripheral logic interface include the idea
of segmenting the memory arrays into banks and developments of alterna­
tive modes of operation. These alternative modes of operation are primarily
focused on data Prefetch modes for column access commands.
The synchronous interface was the most significant development for
third-generation DRAM. This was the beginning of what could be called
“high-perfdrmance" DRAM. The addition of a synchronous interface
increases the complexity of the peripheral logic, but simplifies DRAM pro­
tocol by allowing some of the timing parameters to be expressed with refer­
ence to a global clock. Now, instead of a combination of signals validated
by the CS* signal, we have a continuous clock signal sequencing command
and address inputs.
The point to take away from this discussion is that throughout the evo­
lution of the DRAM interface protocol, the fundamental operation and cir­
cuit topology of the DRAM array has changed very little. Access delay for
opening a wordline or for column data accesses are very dependent on pro­
cess, voltage, and temperature (PVT) conditions. This point is illustrated in
Figure 12.3 where we see Read cycle timing diagrams for both an extended
data out (EDO) DRAM and a SDRAM. Notice the timing parameter for
Icac in the EDO DRAM data sheet relative to the value a = 2*tcK in the
SDRAM data sheet《ck is the clock period). The EDO data sheet expresses
a maximum value for 13ns, which is the data access delay measured
from the falling edge of C4S*. The parameter tcAC is expressed as a maxi­
mum because valid data may be output from the device any time before
tCAG depending on operational PVT. We also see that the CAS* signal looks
similar to a clock: it toggles for each data access that is driven from the
device for the EDO-Page-Mode operation. However, the EDO DRAM is
not a pipelined device. That is why the cycle time of the CAS^ signal,
tpc = 25ns, is greater than the maximum access time, tcAC = 13ns, for each
Read access. This guarantees that successive Column accesses cannot occur
before all data is driven from any previous Read.
302 Chap. 12 Control Logic Design

% Don,t Care
阙 Undefined
b)
Figure 12.3 a) EDO Page mode and b) SDRAM Read cycles.

Early-generation SDRAM had limited pipelining and often operated at

clock cycle times that matched the requirements of the underlying asyn­
chronous DRAM. For instance, lefs look at the SDRAM side of Figure
12,3. If a device is specified to operate at a 133 MHz clock frequency with a
CAS* latency (tci) of 2 and a data access time of 5.4ns, then the total
Sec. 12.2 DRAM Logic Styles 303

maximum access time for the device is 12.9ns. This is expressed as a maxi­
mum because the first valid data out of the device in the SDRAM can still
come any time before the maximum specified access time. The correspond­
ing timing parameter, f°L13ns, for the EDO device reflects approxi­
mately the same access delay as is shown for the SDRAM device.
When we look at a column Read operation for a high-perfbrmance
DRAM, such as the data sheet for a GDDR3 device shown in Figure 12.4,
we see that column access latency has remained relatively constant over the
generational changes of the DRAM protocol. The clock period shown in
Figure 12.4 is 1.25ns; and, with a CAS latency of 10, the maximum column
access time is 12.5 ns. The lack of improved access time is partially because
the transistor speed benefits of process scaling are often tempered by volt­
age scaling and increased interconnect coupling. The sum of the effects of
interconnect scaling with simultaneous increases in area because of increas­
ing array density has caused most DRAM timing parameters to remain con­
stant.
TO T1 T2 T3 T10 T10n T11 T11n T12 T12n T13 T13n

Figure 12.4 GDDR3 Read burst timing diagram.

Another observation from Figure 12.4 is that the short burst length
(BL), which is translated into column cycle time (Jccd) by Equation (12.3)
for a double-data rate (DDR) device,

(12.3)

z = BLt
「CCD 2 ‘CK

allows several Read commands to accumulate before the first data burst is
output from the device. In Figure 12.4, with a CAS latency of 10 and a Read
command occurring every second clock cycle, the result is 5 possible Read
commands pipelined in the device before the first Read command com­
pletes. This is in contrast to early SDRAM designs where each data access
stood alone because the C/A clock frequency was chosen to match the char­
acteristic delays of the memory device. Later in this chapter we discuss how
304 Chap. 12 Control Logic Design

large delay variations fbr array column accesses due to PVT variation can
influence the logic design, especially in the Read data path.

1 2.2.3 Performance Requirements

In the previous section we saw that the characteristic access delays of
the DRAM have remained relatively constant even though data rate require­
ments have changed drastically between Page-Mode DRAM and high-per­
formance DRAM. The data rate requirements for high-performance DRAM
have meant significant changes fbr the peripheral logic interface as well as
the data I/O circuits and analog signaling protocols as compared to earlier
generations of DRAM. To help achieve these higher data rate requirements,
some improvements in peripheral logic transistors have occurred
(Section 12.2.1).
One area where transistor performance significantly improves overall
performance is in the I/O circuitry. This is especially true fbr the receiver
logic and the Read data serialization logic. Also, the interface between the
parallel data bus from the array and the serialization logic greatly benefits
from transistor performance improvements. Later in this chapter, we exam­
ine how Read data is presented to the data serialization logic introduced in
Chapter 10. Transistor performance improvements alone, however, will not
enable high-performance DRAM operation in the peripheral logic interface
circuits. When implementing designs with DRAM process technology, we
must also consider logic circuit style and topology as well as greater paral­
lelism to complete the transition to higher clock frequencies.
Parallelism itself is a natural consequence of clock frequency division.
Clock frequency division, discussed in Chapter 11, also creates multiple
clock phases, which leads to an opportunity for parallel logic paths. For
example, if we are receiving commands in the peripheral logic interface
with a clock running at 1.25 GHz, and we divide the internally distributed
clock frequency by two, we now have a clock frequency of 612.5 MHz. To
maintain the correct command bandwidth established by the external clock
frequency of 1.25 GHz, we must generate a two-phase clock running at
612.5 MHz; with the clock phases shifted a relative 180 degrees. Such a sit­
uation is shown in Figure 12.5. Now we can generate parallel logic paths
based on the two-phase clock signal but at the expense of complications in
the timing circuits and increased gate counts. However, with the process
limitations just outlined and the advantages of clock frequency division dis­
cussed in Chapter 11, the complications in both logic implementation and
silicon area prove to be worthwhile compromises fbr achieving high perfor­
mance.
Sec. 12.2 DRAM Logic Styles 305

800 ps

clk h_h_n_rLrLrLrLrLn_rLrLrLrLn_n_n_n_n_
H~~►：
.1.6 ns ,

ctKor-l— J-1-T-LJ-LJ-LJ_LJ-LJ-LJ-L-
CW ：_j_ i_r_L_j_i_rn_j_lj-1—1~lj-i—rn
800 ps

Figure 12.5 Clock division with resulting multi-phase CLK,

When we look at clock frequency specifications for high-performance

DRAM, we find that there is often a demand for a wide range of clock and
data frequencies. One reason for this is that it is more cost effective fbr a
single device to cover a range of power-performance points. Of course, this
has an impact on the logic design. For instance, Table 12.2 and Table 12.3
show the clock frequency specifications fbr a combination GDDR3/
GDDR4 device.
Table 12.2 GDDR3 frequency and latency range.

Clock Rate
Speed Grade
CL-10 CL = 9 CL = 9 CL = 8

-12 800 MHz

-14 700 Hz

-15 650 MHz

-16 600 MHz

Table 12.3 GDDR4 frequency and latency range.

Clock Rate
Speed
Grade
CL = 14 CL =13 CL = 12 CL-11 CL = 10

-08 1.25 GHz

-09 LI GHz

-1 1.0 GHz

-12 800 MHz

306 Chap. 12 Control Logic Design

To get the best performance over a wide range of operating frequencies,

we must consider how the peripheral logic interface operates relative to the
clock frequency. Along with covering a wide range of operating frequen­
cies, we must also consider that the logic is robust enough that the device
can be speed graded or binned according to the access delays of the array.
What this means is that the logic must not limit the performance of the array
access delays with any dependencies on clock period. In the next two sec­
tions, we briefly examine how logic style can influence the performance of
the DRAM when viewed at the I/O pins of the device. In the remainder of
the chapter, as we discuss more specifically how the peripheral logic inter­
face is implemented, we examine how the chosen logic style must bridge
the gap between the synchronous environment in which the DRAM oper­
ates and the asynchronous DRAM array.

12.2.4 Delay«Chain Logic Style

One style of logic that has been a natural fit for the asynchronous nature
of the DRAM array, and is not widely accepted outside of DRAM design, is
a style we refer to here as delay-chain logic. This logic style is comprised of
a mixture of combinatorial gates, one-shots (monostable multivibrators),
and latches. The reason we define this style of logic as “delay・chain” is
because the logic is usually implemented with a series of delays such that a
sequence of functions or events can be executed without a global timing ref­
erence (clock), In other words, the logic consists of a cascade of events and
functions. We will look at an example of how this style is applied as part of
a column access command to the array.
Figure 12.6 is an example of delay-chain logic used for enabling the
column decoder and helper flip-flop logic (HFF, Section 8.2) circuits for a
column access. At the input to the circuit is a signal named coICyc*. This
signal starts the delay chain and initiates the sequence of events necessary to
enable the column decoders and access data from the array. There are sev­
eral more delay chains cascaded with this circuit that perform other sequen­
tial operations. We are not going into the details about this circuit's
operation as it is only presented here as a simplified example of a delay­
chain logic circuit. However, we will examine the implementation of the
one-shot that follows colCyc^, As a side note, keep in mind that an array
access fbr both Reads and Writes is identical. Think of the array as a Read­
only device. A Write occurs when a Read access is overdriven by Write
driver circuits, through the I/O transistors (Section 1.2) and onto the data-
lines where the sense amplifier restores a full-rail state to the cell.
Sec. 12.2 DRAM Logic Styles 307

Figure 12.6 Delay-chain logic example.

We see in Figure 12.6, that when colCyc^ fires LOW, a one-shot of

1.5 ns duration is triggered. The design of this one-shot circuit is shown in
Figure 12.7. The circuit has an equilibrium state where input A is HIGH
with node C LOW, which forces node B HIGH. Looking at Figure 12.7,
whenever input signal A transitions LOW, node C transitions HIGH, which
causes output Y to fire LOW. The delay cell labeled “In” in Figure 12.7
only delays rising edge inputs. Such a delay element is implemented
through capacitive loading in a string of inverters. After the rising edge
from gate X0 is delayed and inverted through the inverted delay path, a
LOW appears on node B, which forces Y back HIGH completing the
approximately 1.5 ns pulse.

Figure 12.7 One-shot design.

When we consider PVT variation, the pulse width out of the circuit in
Figure 12.7 can vary by 100% over the full operating PVT range. How does
the delay-chain logic style fit in with high-performance DRAMs? The
308 Chap. 12 Control Logic Design

answer is that this logic style is still widely used in the peripheral array
logic region. As we stated at the beginning of this chapter, when synchro­
nous DRAM came into existence, the synchronous interface became a
wrapper around the existing DRAM logic. Thus, we have defined the
peripheral logic interface as the high-performance, synchronous section of
logic that interfaces to the original DRAM peripheral logic. Many DRAM
designs still use some form of delay-chain logic in the peripheral interface
region of the DRAM die. This method can prove effective when directly
interfacing to the array, but as performance demands grow and process con­
tinues to scale, this style of logic has limitations that must be considered.
Some risks when using delay-chain logic in the peripheral logic inter­
face region include questionable delay tracking reliability when operating
over a wide range of clock frequencies as well as increased process variabil­
ity as a result of scaling. The latter results in reliability issues related to sta­
tistical delay variation relative to anticipated delay values. Another
important consideration when using delay-chain logic is the fact that the
heuristic nature of the design approach for this style of logic often leads to
disjointed circuitry with a lack of procedural clarity fbr the design process
flow. What often happens is that delay-chain logic designs are passed
between different designers and the logic is patched to the point that the
original functionality is difficult to discern. When delay-chain logic is used
in the peripheral logic interface area, the circuits must be “tuned“ to operate
fbr a given process, voltage, and frequency. Being forced to tune these cir­
cuits limits the reusability of this style of logic. This is not to say that this
style of logic is not without merit when applied judiciously in the peripheral
array logic region. The greatest strength of using this logic in the array
peripheral region is that the delay chains exhibit some correlation to PVT
variation between the array control signals and the delay characteristics of
the array itself. Another advantage to this style of logic is the ability to
adjust the delay stages using metal mask changes so that post-silicon timing
adjustments can be easily made.
Next, we will consider a dynamic logic methodology. This method
solves some of the problems of delay variation and is well suited to the
asynchronous timing specifications of DRAM array accesses. Two more
important advantages for applying dynamic logic in the peripheral logic
interface is that first, this method provides a rigorous design procedure for
implementing logic functions and second, the logic implementation is only
limited by gate delay and is robust when properly designed.
Sec. 12.2 DRAM Logic Styles 309

12.2.5 Domino Logic

Early DRAMs used NMOS logic for implementing protocol operations,
address decoding, and even output pad drivers. P-channel devices have long
been practical for the implementation of full CMOS gates, yet p-channel
devices still pose a limitation to logic circuit performance, especially fbr a
DRAM process where cost concerns limit process improvements. Another
way to overcome limitations of the p-channel device is by using the domino
logic style [10]. Domino logic is interesting in that each gate requires a
clock signal, yet logic functions can be designed that are not necessarily
synchronous to a global clock. Instead, logic functions can be implemented
that result from a cascade of events, where each gate output has the poten­
tial of generating a clocking function for subsequent gates in a chain of
logic. This description may sound similar to the description of the delay­
chain logic style where logic functions are cascaded and separated by chains
of delays that ensure events follow a known sequence. This concept can be
extended to the use of domino logic fbr implementing many of the neces­
sary logic functions in the peripheral logic interface. In this section, we
examine some of the advantages of domino logic and how the application of
this technique is well suited to the synchronous interface of the DRAM and
the preservation of asynchronous array performance.
Figure 12.8 illustrates the operation of a footed domino gate that imple­
ments an AND function. The term "fboted” refers to the transistor MN2 in
the n-channel stack. This transistor guarantees that the current path through
the n-channel stack is cut off during evaluation. When the clock signal,
CLK, is LOW, the node DYN is precharged to Vcc through the p-channel
device, MPO. This state is referred to as the precharge phase of operation.
When CLK transitions HIGH, the state of A and B is evaluated, which is
referred to as the evaluation phase. If both A and B are HIGH, DYN is dis­
charged to ground through MN0-MN2, forcing the output of the inverter
HIGH. The device MP1 is used as a keeper device to counter device leakage
when the CLK is HIGH and DYN is not discharged. MP1 is sized with a
long channel to reduce crowbar current and performance degradation during
the evaluation phase.
310 Chap. 12 Control Logic Design

Figure 12.8 Domino AND function.

Many designers are unfamiliar with the application of domino logic

because most logic design tools and methodologies focus on synchronous
implementations using hardware description languages, such as VHDL.
This level of abstraction is not as prevalent in DRAM logic design, so the
designer has broader latitude when choosing logic styles. However, this
freedom comes with a caveat. For instance, we must be aware of some of
the risks involved when implementing domino-style logic to achieve a
robust implementation. These risks include interconnect capacitive cou­
pling, leakage, charge sharing, noise tolerance, minority carrier charge
injection, and other issues related to reliability. These topics, though very
important, are not covered in this work. A discussion of these issues can be
found in [11] and [12]. The reader would be well served in researching these
issues before forging ahead with a domino logic design. We will instead
focus on some of the advantages for using domino logic when interfacing
the synchronous DRAM interface with the asynchronous DRAM array.
Perhaps the greatest advantage of using domino-style logic in the
peripheral logic interface is the monotonic signal transitions fbr each term
of a logic function. We define a monotonic signal, with regard to the dom­
ino-style logic, as a signal that can only transition from LOW to HIGH dur­
ing the evaluation phase. The static gate (the inverter in Figure 12.8) that
forms the output stage of the domino gate enforces this regiment by invert­
Sec. 12.2 DRAM Logic Styles 311

ing the output of the dynamic gate and driving a LOW-to-HIGH transition
to the input of the next domino gate. Monotonicity must be enforced
between levels of logic fbr reliable timing and robust operation.
One problem we face in the DRAM interface is that signals coming
from the input circuit path are not monotonic, as we have defined here. Sig­
nals such as CAS* and RAS* transition at the command/address clock (C/A
clock) frequency and can transition both HIGH or LOW between each eval­
uation phase of the clock. If we are to use domino-style logic in the periph­
eral logic interface, we must first convert the control signals, such as CAS*
and RAS*, from a static format referenced to the C/A clock to monotonic
signals compatible with the domino logic style.
Figure 12.9 shows a circuit that converts signals from the static logic
domain to the dynamic logic domain [26], If we look at the input latch cir­
cuit examples from Chapter 9, we see that the output from these circuits is
generally valid for an entire clock cycle. The circuit in Figure 12.9 can be
used to convert from the full-cycle static domain into the monotonic, half­
cycle valid domain. This circuit operates by creating a pulse from the clock
edge through the inverter string driven by the clock signal. When CLK is
LOW, the outputs Q and Qb are precharged to a low value through the p-
channel devices. When CLK transitions HIGH, the data input is evaluated
during the pulse created by the overlap of the HIGH time of the clock signal
and the delayed transition of the clock through the inverter chain. An impor­
tant aspect of the static-to-dynamic converter circuit is that the output is
considered dual-rail. What this means is that if the input data to the con­
verter, D, is LOW, then the output signal Qb will transition HIGH for one
half of a clock cycle. Likewise, if the input D is HIGH during evaluation,
then the output Q will transition HIGH while the output Qb remains in the
Precharge state (LOW). Domino circuits cannot be cascaded to perform
combinatorial logic because the outputs of domino gates are not function­
ally complete. Static hazards can occur where, for instance, a non-mono­
tonic transition from HIGH to LOW could be skewed relative to the
Precharge clock such that the gate evaluates to the incorrect state. Once the
dynamic logic gate incorrectly evaluates, it cannot recover until the next
evaluation cycle. The dual-rail output from the circuit in Figure 12.9 gener­
ates both true and complementary outputs that are both monotonically ris­
ing. This allows us to exploit the skew-tolerance [1] of domino logic and
secure functional completeness for implementing the command decoder
logic.
312 Chap. 12 Control Logic Design

The inset of Figure 12.9 serves as a point of reference for using the
static-to-dynamic circuit in the DRAM input path. Figure 12.9a shows the
connections of the single-ended input receiver circuit used to detect the
incoming signal, followed by Figure 12.9b, the input capture latch, and then
followed by Figure 12.9c, the static-to-dynamic converter circuit. This
example would be common for the command and control input signals such
as RAS* and CAS*.
At this point we should consider the question: Why use dynamic logic
in the command input path of a high-performance DRAM? Part of the
answer to that question is related to the process limitations discussed at the
beginning of this chapter. Remember that DRAM processes implement as
few metal layers as possible, making global clock distribution very difficult.
Not only is it difficult in a DRAM process to distribute a well-matched
clock, but using a global clock results in associated cycle delay dependen­
cies on array access time, thereby limiting the usefulness of a globally dis­
tributed clock. Recall that an important timing parameter for DRAM
performance is column access latency, which we have also referred to as
CAS latency (4)• Because column access latency is constant across the
range of operating frequencies for a device, we must not allow the clock
Sec. 12.2 DRAM Logic Styles 313

period to become part of the array access latency equation. Figure 12.10 is a
diagram showing how internally synchronous operation would detrimen­
tally allow the clock period to become part of the access latency. By syn­
chronizing the command and address logic, we could potentially force a
violation of the specified array access time. In Figure 12.10, the command
and address are captured by the input capture latches (Chapter 9). If the fol­
lowing clock edge is used to align a partially decoded command (assuming
that the total command decode plus synchronization overhead exceeds a
clock cycle), the decode delay will be increased by the difference in clock
period delay between the two operating frequencies shown. In this example,
a device is specified with an operating frequency range of 800 MHz to
1.25 GHz. If we allow a clock cycle to occur between the command capture
clock edge and one stage of the partial command decode, a & delay differ­
ence of 900 ps will be added between the minimum frequency and maxi­
mum frequency operating points. Because of this difference, if devices are
binned with this frequency dependency, many potentially high-performance
devices may not meet the latency specifications.

800 ps 1.25 ns
O

rLrLhLTLFLTL 'ri_rn_h_rT_

Figure 12.10 Example of a synchronous decode path.

One obvious solution for maintaining constant CAS latency across the
operating frequency range is to use a combinatorial circuit fbr the decode
operation and delay the input clock to a second set of retiming latches. The
second set of retiming latches are required fbr realigning the decoder output
to account fbr timing skew and hazards through the input and decode path;
this configuration is illustrated in Figure 12.11. Without realignment of the
decoder output, signal skew can cause false decoder transitions to appear in
314 Chap. 12 Control Logic Design

the command and timing logic paths and possibly lead to data corruption in
the DRAM array. One problem with this solution is that the clock must be
delayed based on the maximum skew and delay through the decode path. If
the delayed clock does not correlate with the delay through the input and
decode paths, then, potentially, a high-performance device may not meet
latency requirements at the highest operating frequency. For example, a
device might meet timing requirements through the decode path but,
because the clock is delayed based on worst-case timing expectations (with
timing margin built in, of course), then any performance gains in the decode
path may be masked by the delayed clock. This would be a greater concern
for those devices operating right at the margin, since the delayed clock may
prevent the device from binning at a higher performance point.
Skew between decoded signals is not the only problem with using com­
binatorial decode circuits for command and timing control. For high-perfor­
mance devices with extremely low clock periods, the decode path could
also have multi-cycle delay paths at higher clock frequencies but still satisfy
single-cycle timing at lower clock frequencies. This problem gets back to
the design and use of the same DRAM device over a wide range of clock
frequencies. In this situation, false decodes could occur if the incorrect
clock edge were to trigger a false decode. And, finally, using the simplistic
method of a combinatorial decoder causes greater difficulty for final timing
verification of the design. This is true because the global or, in some cases,
multiple local clocks that are used for retiming the decoder output must be
comprehensively verified for all combinations of decoder outputs.

Figure 12.11 Example of a gate delayed clock decode path.

12.2.6 Testability
One disadvantage of using domino logic is the issue of testability.
Because domino logic is dynamic, there are possibilities that at extremely
low frequencies, such as the clock rates used during array testing early in
the manufacturing process, charge could leak from the dynamic nodes and
Sec. 12.2 DRAM Logic Styles 315

cause incorrect operation. Because the design of monolithic DRAM largely

follows a fiiil-custom design flow, the implementations of scan chains for
fault checking and automatic test pattern generation are disregarded. The
ratio of the number of transistors used for the logic in the timing and control
of the DRAM compared to the number of array transistors is so small that
logic testing has largely been ignored in the test flow. Although, as package
technology has advanced, the requirement for boundary scan chains, prima­
rily intended for continuity checks during system level manufacturing, has
become necessary. The issue of testability is a very broad issue and is only
briefly treated in this section.
One very simple yet brute-force method fbr implementing logic that
allows low-frequency array testing is to implement a parallel static logic
interface for directly accessing the array peripheral logic. This method
allows the device to be treated as a Page-Mode DRAM similar to the
devices described in the early chapters of this book. This implementation
does not test the high-performance peripheral logic interface, but it does
allow low frequency testing and asynchronously controlled array timing
stress to find weak array bits. When implementing a source-synchronous,
matched-input-path methodology, there are often opportunities to generate
static signals directly from the input logic paths. These static signals are
often branched off of dummy buffer loads, which are used fbr load match­
ing in the matched delay paths (Chapter 9).
Another method fbr low-frequency testing is to place back-end latches
on the dynamic gates to isolate the dynamic node and provide static outputs.
Figure 12.12 shows two examples of how static signals are generated from
the dynamic gates. In case (a), the output of the domino gate is given a full
weak inverter latch [1]. This differs from the usual p-channel keeper (Figure
12.8). The full latch is used on the first gate in a chain of domino logic to
maintain both a HIGH and LOW state on the output of the gate. AH subse­
quent gates sharing the clock phase can be designed with only the p-channel
keeper, thereby maintaining the performance advantages of the domino
logic. The output of this circuit follows the monotonically rising signal pro­
tocol previously discussed.
Figure 12.12b shows what is commonly referred to as an N2Latch [24].
Both the monotonic output and the static output from this gate can be used.
The static portion is routed to test logic in the array and the monotonically
rising outputs are ignored during low frequency testing. The static output of
this circuit remains valid fbr a full cycle, unlike a monotonically rising sig­
nal output from Figure 12.12b. A disadvantage to this method is that the
extra loading on the dynamic gates potentially increases gate delay.
316 Chap. 12 Control Logic Design

Ystatic

Figure 12.12 Static outputs from dynamic gates.

Now that we have examined some of the reasons for utilizing domino
logic for portions of the timing and control circuits in the peripheral logic
interface, we more specifically look at some of the issues related to com­
mand and timing control of array accesses and why we might want to make
particular choices in logic implementations. As stated at the beginning of
this chapter, these solutions are only one method to accomplish the task of
meeting the industry standard specifications for a high-performance
DRAM. The specifications are the thread of commonality that ties together
the operation of DRAM between multiple vendors. The descriptions that
follow are not meant as a design guide, but rather an illustration of the prob­
lems posed by the synchronous interface surrounding what has essentially
remained an unchanged asynchronous DRAM. In the next section, we
examine how domino-style logic can be implemented in the command
decode and address register logic. Following that, we look at the data side of
the DRAM and examine the Read and Write data paths and how the vari­
Sec. 12.3 Command and Address Control 317

able latency of array accesses are made to fit within a deterministic timing
framework established by the global clock.

12.3 COMMAND AND ADDRESS CONTROL

At the center of the peripheral logic interface is the command and address
control logic. This area of logic is the interface between the command and
address input circuit paths discussed in Chapter 9 and the timing and data
I/O interface logic that we will examine in the sections to follow.
Figure 12.13 is a composition of the primary protocol implementation for
the peripheral logic interface with what we consider to be the control and
address logic highlighted for reference. The functional blocks shown in Fig­
ure 12.13 will be the subjects of the remainder of this chapter.

Let us consider the context of each of the highlighted functional blocks

of Figure 12.13. The command decode circuit combines control signals
such as RAS* and CAS* in the context of the system clock and determines
an action to take based on the command protocol. Depending on the com­
mand that is decoded, address inputs may need to be captured from the
address input path. The address information fbr an array access is stored in
registers based on whether a row or column access is decoded. The outputs
318 Chap. 12 Control Logic Design

of the address registers are routed to the peripheral array logic and are
decoded by the row and column decoders. The outputs of the address regis­
ters are not synchronized to a global clock but are instead bundled with sig­
nals that enable the peripheral array logic circuits. The control and timing
signals are bundled and carefully routed in conjunction with the address sig­
nals. This enables correct timing when activating the row and column
decoder circuits in the peripheral array logic. In particular, row address reg­
isters are often an extension of the peripheral array logic, with the style of
logic resembling the delay-chain style described earlier in this chapter. The
outputs of the row address register circuits are often referred to as bank con­
trol logic because each bank is determined by the division of row address
decoders to sections of the DRAM array that can be independently
accessed. We will not fbcus on bank control logic as its implementation
depends on array architecture and is largely unchanged for high-perfor­
mance DRAM, except for the case of reduced latency DRAM described in
Chapter 8.2. Also, see Chapters 2, 3, and 4 for details on how row and col­
umn decode and control logic operate in the peripheral array logic area of
the DRAM die.
The remainder of this section describes our choices for command
decoder logic style and operation for a high-performance DRAM. We also
discuss some architectural decisions that must be made with regard to the
column address registers that are important to achieving low column access
latency and how increased I/O bandwidth requirements affect the way that
column addresses are handled.

12.3.1 Command Decoder

The command decoder is the control center for all DRAM operations.
Control signals originate from this circuitry fbr every array access operation
defined fbr the device. For high-performance synchronous DRAMs, the
command decoder has taken a greater role in sequencing commands in
accordance with the clock cycle time. This implies that the command
decoder can be synchronous fbr some of its operations. The command
decoder operation is becoming more important as the command and address
protocol change due to increased memory density and interest in limiting
the number of pins fbr each DRAM in a system. In some of the most recent
DRAM command-address protocols, the command is not necessarily
aligned with the complete address on the same clock cycle. Pin count can be
reduced by splitting the address between multiple clock cycles to avoid hav­
ing a wide, parallel address bus. Figure 12.14 shows different access modes
fbr command and address protocols.
Sec. 12.3 Command and Address Control 319

Command
X
Address
X
RAS*心AS*

^Lpun
Asynchronous DRAM access Synchronous DRAM access

Command bus
I I

Address bus

RASyCAS*

Future high-performance DRAM access

Figure 12.14 Command and address protocols.

Before synchronous DRAM, column accesses initiated through the

command decoder were timed using a logic style that was the same or simi­
lar to the delay-chain logic style discussed earlier in this chapter. The com­
mand decoder was fully implemented with combinatorial logic, and a signal
or signals that emulate the worst〜case delay through the decode terms was
generated similar to the scheme shown in Figure 12.11. These delayed sig­
nals were implemented as ENABLES on combinatorial decode terms or as
clocks on flip-flops or latches fbr retiming the decode output terms to avoid
false decodes.
Following the command decoder is a block of logic that generates and
sequences the control signals of the DC sense amps (DCSA) or helper flip­
flops (HFF), which are described in Chapter 5. Of course, this logic is still
implemented in the peripheral array logic of high-performance DRAM.
This group of logic circuits is often referred to as the CAS chain. With syn­
chronous DRAM, data accesses are sequenced by the system clock, which
establishes the conditions that allow the command decoder to be sequenced
320 Chap. 12 Control Logic Design

by clock-driven signals. This is because the data sheet specification dictates

the cycle delay between commands.
Our discussion of the command decoder synchronization began in
Section 12.2.5. In that section, we discuss the advantages of domino logic
and how this style of logic can lead to gate-delayed operation. What we
mean by “gate-delayed” in this context is that the logic operations occur
independent of a global clock signal and, therefore, independent of clock
period delay as well as of the synchronous overhead of setup, hold, and
CLK-to-DATA delay associated with synchronous circuits. Remember that
input logic must be independent of clock period because column access
latency is constant regardless of clock frequency.
Figure 12.15 is an example of a gate-delayed logic path implemented
with domino logic. This logic path generates a signal to indicate that the
device has received a Read command. The command inputs RAS*. CAS*,
WE^f and CS* are distributed through the input circuit path and latched in
the input capture latches. In this example, because the delay through the
capture latch is slightly greater than the minimum clock period, the clock
signal is delayed by buffer delay to meet the timing requirements of the
dual-rail static-to-dynamic conversion circuit (Figure 12.9). This delay is
inserted following the critically matched signals of the input capture path
(Chapter 9). The Read command is indicated by C4S*=0,卯E*=l,
and CS*=0. The dual-rail conversion circuit (DE) generates monotonically
rising signals from the clock edge for each of the signal conditions. For
example,印EX would generate a rising output on the Q signal of the con­
version circuit, and the Qb signal would remain LOW as a result of the pre­
ceding Precharge period. The signals resulting from the conversion stage
are routed through successive domino stages that are clocked directly by the
dual-rail signals. The domino gate X5 is clocked by the CASQ* signal with
the CSQ* and RASQ signals combined in the dynamic AND gate. The out­
put signal Column of this gate indicates that a column command is poten­
tially received. The signal Column is then combined with the WEQ signal in
the dynamic gate X6 to generate the monotonic signal Read. Since the clock
signal is only used for the first stage of logic, we can consider the logic path
following the clocked gates to be gate delayed because the logic delay is
based on the gate delay through the path and has no direct timing dependen­
cies on clock period.
Sec. 12.3 Command and Address Control 321

clk njTnTLrLrLrLrLjmnnrL
Command NOP NOP

Internal ------------------------------- /—< /----------------------------------------------------------------------------------------------------------

command N°P NOP___________________________________________________

csq jnjTn—TLrLrLrLrLTLnrLn
CSQ* 口

创 _rLrLrL_rLrLFLrLrLrLrTrm
C4SQ* Fl_______________________________________

.q 丁 LTLrLrLrLrLTLrLrLjrrLrLn
RASQ*

作q ZrLrLnjTrLrLrLTLTLrLrLrLn
WEQ*

READ Fl

Figure 12.15 Gate-delayed domino logic decode path.

Again referring to Figure 12.15, the timing relationship between dom­

ino gates X5 and X6 can take advantage of the footed domino gate so that
the Read signal has a pulse width of one half of a clock period. If we use the
signal Column as the clock for gate X6, then the only timing requirement is
that the 印EQ* meet the hold time requirement for gate X6, [13]. These tim­
ing relationships are shown in Figure 12.16 with t[)Q designating the delay
through the footed domino gate.
322 Chap. 12 Control Logic Design

WEQ* ---------------------

Figure 12.16 Hold time requirement fbr domino gate.

One of the advantages of using domino logic fbr generating the decode
and timing signals is that if a device suffers logic timing issues, they can
usually be overcome, and correct operation is attained, by simply reducing
the clock frequency. In Figure 12.16 we see that reducing the clock fre­
quency increases the pulse widths, which results in more hold time fbr the
final decoder logic stage. Reducing clock frequency has advantages for syn­
chronous static logic as well, but recall that the advantages of domino logic
include reduced reliance on clock distribution networks, skew tolerance,
and a reduction in timing overhead.
The example just described is for simple, single-phase clock implemen­
tation, The pulse width of the signal Read is approximately one half of a
clock period wide. For extremely high frequency, we may be forced to
divide the clock frequency to get wider pulse widths on the decoder output
signals. A wider pulse width on the decoder outputs helps ensure better sig­
nal integrity when distributing the signal over long distances. For this situa­
tion, each gate would require two phases of dynamic logic with the last
static gate in the logic path implementing a merge function so that a single
decode signal would be generated with a fiill-cycle pulse width, clock fre­
quency division is only possible when the array access cycle times are mul­
tiple command clock cycles. For instance, if the burst length (BL)
specification were a minimum of four fbr a DDR device, then, given that the
command clock and data strobe frequencies are equal, the column cycle
time, as determined by Equation (12.3) would be two clock cycles. Clock
frequency division under these conditions would be possible. A frequency
divided decode logic path and timing diagram is shown in Figure 12.17.
Sec. 12.3 Command and Address Control 323

Dec

CMD

[ I-CLKO
CLK^> I-CLK2

clk rrrLrLrLrLrLTLn_rLrLn_rLrL

Command Rpf3" ~NOP NOP

ML ― 而

CLK2

_____________________________________________ |~|______________________
DecQ

Figure 12.17 Frequency divided gate-delayed domino logic decode path.

Depending on the circuits used in the input circuit path, we may be able
to take advantage of the receiver circuits for generating monotonic signals.
The single-ended signaling protocol is the most common signaling protocol
for monolithic DDR DRAM. This signaling protocol applies a reference
voltage on one input of the receiver differential pair, with the target signal
connected to the opposite input (Figure 12J7). An example of this topology
is shown in Figure 12.18 (a). The latch is separated from the receiver, which
allows matched delay interconnect distribution for CLK and DATA. For
high-performance DRAM, a receiver sense-amp latch (SA latch), shown in
Figure 12.18b, can be used as a combination input detection circuit and
latch. The SA latch circuit tightly couples the analog detection circuit with a
static latch. The detection circuit of the SA latch naturally outputs a dual­
rail, monotonic signal to toggle the inputs of a set-reset latch. If these sig­
nals are buffered out of the SA circuit, both the latch stage and single-ended
to double-ended conversion circuits (shown in Figure 12.9) are bypassed.
As a side note, the SA latch can also be used for differential signaling. This
signaling technique has disadvantages when applied to DRAM because of
increased pin count. For differential signaling, channel bandwidth must be
324 Chap. 12 Control Logic Design

increased to achieve equal bandwidth-channel product compared with sin­

gle-ended, wide bus C/A paths.
Once a partial decode is detected, enough information is available to
determine if an address must be loaded into the address register logic. The
address register logic is a secondary register set that reflects the contents of
the input capture latches when a valid column or row command is received.
The input capture latches are updated on every cycle of the clock, and the
secondary registers are loaded only if a valid RAS* or CAS* combined with
a valid chip select is received. In the next section, we examine some issues
related to implementing the secondary address register, focusing on column
address operations.

a)
clk n_n_n_n_n_
-------- SigQ*

”RR
Sig
/ <T Ech
— Sig static
y
SigQ _____ 口______
CLK -------- SigQ SigQ” _n_n_j_Ln_
b) SigStatic | |

Figure 12J8 SA L lal-rail conversion.

12.3.2 Read and Write Data Address Registers

In early synchronous DRAM, the minimum command and address
setup times to the clock were long enough that DRAM designers would
often cheat and use the CAS* setup time to allow address signals to start to
propagate into the device through transparent capture latches. Doing this
allowed the address signals to propagate into the device and achieve faster
column access time for Read operations. This essentially made the synchro­
nous DRAM look internally like an asynchronous DRAM device. For our
high-performance example, we look at an address topology that incorpo­
rates a register that is secondary to the address capture latches. These sec­
ondary latches are controlled and timed from the command decoder circuit.
The functional relationship between the command decoder and the second­
ary address registers is shown in Figure 12.13.
For array Read accesses, we want to retrieve data from the array with
minimum latency. We can determine the delay for array accesses by break­
Sec. 12.3 Command and Address Control 325

ing the delay into two components: array access latency and command
decode latency. Array access latency is the delay measured from the
enabling of the column address decoder to the arrival of valid Read data
from the helper flip-flops (Chapter 5). We designate this delay as tcAA- The
command decode latency is a measure of the sum of the input circuit delay,
the command decoder delay, and the address register logic. We designate
the total command decode delay as tcMD・ Considering these definitions, in
order to minimize array access latency from the perspective of the control
logic, we must distribute the address to the peripheral array logic and enable
the column address decoder with the minimum possible delay.
By partially decoding the command, a monotonic signal that has a pulse
width of one half of the valid address period is generated. This allows the
use of transparent latches for the address register. The latches are transpar­
ent when the CASQ* signal, qualified by the chip-select, is HIGH. The tim­
ing of the latch signal aligns with the address transitions from the input
capture latches so that the address signals start to transition on the address
distribution path following the t[)Q (data to output) delay of the latch circuit.
The timing relationship between the latch signal and the captured addresses
is advantageous since the latch signal is one half of an address cycle wide
while the address signals are one clock cycle wide; thus, the critical timing
parameter is tcp as shown in Figure 12.19.

Read
address Column
Latch
address

Write
address |
厂

Figure 12.19 Read column address register timing for transparent latches.

The column address for Writes must be handled differently from the
Read column address. For Write commands we must wait until the Write
data is available before the column address decoder can be enabled. Figure
12.20 is a timing diagram showing how the arrival of Write data is delayed
326 Chap. 12 Control Logic Design

by the Write latency specification,你,and the burst length, Irl, for data
arrival at the Write drivers in the peripheral array logic. Notice that in this
example, data is not available to be retired to the array before the end of the
first burst of four data bits. While the device waits to receive all of the Write
data, Write commands and addresses continue to be received. This differs
from the column address for Reads where the address is retired as quickly as
possible to achieve minimum array access latency. That is why there is a
mux in the address path shown in Figure 12.19 to illustrate that the Write
address originates from a different register than the Read address.
clk rLrLULrLruTrLrLrLrLrLrLrLTL
Command

Address "

Write
strobe
-
Lhjnirmj-Lru
I I
Write data
—^<XXX>00000<»
—
I

Figure 12.20 Write data timing relationship to Write command and address.

For our high-performance DRAM logic example, the Write addresses

are loaded into a first-in first-out (FIFO) circuit. There are several design
alternatives for FIFO circuits. The first consideration is determining the
maximum Write address FIFO depth required. Knowing the FIFO depth can
help select the type of FIFO circuit best suited to the design specifications.
Again referring to Figure 12.20, we see that the first address must remain
available until the Write data associated with that address has arrived. For
the situation in Figure 12.20, we see that the maximum FIFO depth is deter­
mined by the maximum specified Write data latency and the minimum
allowed burst length. Remember that column cycle time (this is also column
address cycle time) can be expressed as a function of burst length (Equation
(12.3)), The burst length is equal to the Prefetch depth fbr each data I/O and
in our example a is equal to column cycle time. If 机 and 极 are
expressed as a function of discrete command clock periods, then we can
determine the maximum depth of the FIFO. Equation (12.4) is an expres­
sion for Write address FIFO depth where FD is the FIFO depth, WLmax is
the Write latency expressed as an integer number of clock cycles, and is
Write data transfer overhead fbr transferring the Write data and column
address to the peripheral array logic. The FIFO depth must be rounded up to
the nearest integer fbr fractional results.
Sec. 12.3 Command and Address Control 327

BL Mt" i TTTT
2 Max tWT 2 WLMax tWL
-------------- +--- = [ +------- +---
BL Min ，CK BL mm (CK
-2~

(12.4)

Ignoring the transfer overhead,恤,we find that for our example in Fig­
ure 12.20, FD=3. Of course, we can not ignore the transfer overhead, which
is partially dependent on the address FIFO design chosen.
One possible Write address FIFO design is a simple asynchronous FIFO
[14] [15], The advantage of using an asynchronous pipeline design is that
there is very little overhead in the circuit implementation and, if the design
is carefully chosen, the forward latency through the pipeline can be mini­
mized. Another advantage to the asynchronous pipeline design is compati­
bility with the signaling protocol for decoded addresses as examined earlier
in this section. If we generate a decoded Write command similar to the
decoded Read command in Figure 12.15, we can use the generated WRITE
signal to load the Write address FIFO directly from the column address reg­
ister. Figure 12.21 is a top-level block diagram showing this configuration.
The timing parameter, f尸乙 is the forward latency through the asynchronous
address FIFO. This is the sum of delay for each latch controller stage in the
FIFO. This delay, summed with the address distribution delay, must be less
than the minimum Write data arrival time shown in Figure 12.20.
Address

AddLt Fl_______________

Read _______________
address X

Column i _______
address 义 Add

RM*

Figure 12.21 Column address register and MUX control.

There are many details surrounding the design of the Write address
FIFO. If an asynchronous FIFO is used, the throughput must be considered,
especially when the FIFO is at capacity. Also, the asynchronous channel
communication between latch stages must be considered for compatibility
with the logic-signaling environment in which the FIFO is placed. A careful
examination of the references provided should edify the reader on these
328 Chap. 12 Control Logic Design

issues with details of the FIFO design and implementation that are beyond
the scope of this text. Primarily, we must be aware of how the DRAM spec­
ification affects the address distribution logic in the peripheral logic inter­
face area of the DRAM device.
In the next section, we examine one method for enabling the data access
control logic once the command and address are fully received. This
method involves a two-wire interface with the control signals bundled with
the address and data signal busses. We illustrate how the asynchronous
nature of the DRAM array affects the Read and Write I/O data paths.

12.3.3 Column Access Control

Continuing our example of a DRAM logic interface, we now examine a
simple two-wire signaling protocol for controlling timing and cycle time for
array accesses. For data accesses, once the command is decoded and the
address is loaded into the appropriate address register, a signal is generated
and distributed to the DRAM array banks. We call this signal, colCyc, and
its function is to time the arrival of column addresses at the peripheral array
logic. A second signal is generated in the peripheral array logic when Read
data is available for transfer into the peripheral logic interface of the
DRAM. We refer to this signal as, dataValid. It is important that we under­
stand the function of these signals as these are the primary means for com­
municating data accesses between the peripheral array logic and the
peripheral logic interface illustrated at the beginning of the chapter in Fig­
ure 12.13.
Figure 12.22 is a timing diagram showing the timing relationship
between colCyc and the column address for Read and Write addresses as
well as the timing relationship between the dataValid signal and Read data
driven from the HFF circuits in the peripheral array logic. These signals fol­
low what is commonly referred to as a bundled signal protocol. Simply
stated, the bundled signal protocol is the case of a single timing signal that
is driven concurrently with transitions of a signal bus. The bundled signal is
meant as an indicator that data on a bus has changed. This may sound like a
clock signal but, unlike a clock signal, it is not a continuous signal, and it
only transitions when there are transitions on the associated bus.
Sec. 12.3 Command and Address Control 329

clk rLrLrLn_n_rLTLrLrLn_rLrLrLnj-LrL

dataValid
ri______
Figure 12.22 Two-wire bundled signal interface for data accesses.

Notice in Figure 12.22 how colCyc transitions are timed when there is a
Read access versus a Write access. The R/W^ signal is included to indicate
when a Read or Write occur. Remember that the DRAM array really should
be considered a Read-only device when a column address is enabled. Writes
are a result of overdriving the array sense amplifiers through Write driver
circuits. For Reads, the access delay is partially determined by the com­
mand decode delay and address distribution logic. Writes are partially
delayed by the Write data latency and burst length. This is shown in Figure
12.22.
The dataValid signal is sent from the peripheral array logic to the
peripheral logic interface whenever Read data is driven from the HFF cir­
cuits. This signal is very important to the operation of the Read data path in
the peripheral logic interface because tcAA will vary with changes in tem­
perature and voltage. This signal must be considered folly asynchronous
without reference to any clock signal generated near the array interface. On
the other hand, the colCyc signal does have an established cycle time based
on the command clock. This becomes a more important distinction when we
examine the Read data path in Section 12.4.
At this point, we have examined how process can determine which
logic style we might choose for implementing DRAM access functions. We
have also illustrated an application of domino-style logic and how exploit­
ing the gate-delayed potential of domino logic allows us to maintain con­
stant data access latency across a wide range of clock frequencies. Further,
we have examined how the DRAM specification can affect how the column
address is treated for both Read and Write accesses. But there are many
details that have been left out of our discussion of DRAM peripheral logic
implementation. The delay-chain logic style implemented in the peripheral
array logic is more often an art form than an academically rigorous logic
style. There is also the wordline control logic that we have mostly ignored.
330 Chap. 12 Control Logic Design

These sections of logic are important but the details do not contribute
greatly in our discussion of some of the unique problems encountered in
high-performance DRAM logic design. Therefore, the remainder of this
chapter is devoted to examining the data I/O circuit paths between the
peripheral array logic and the data input and output paths. This will include
an examination of the Write data serial-to-parallel conversion circuits
(demultiplexors) and the Read data FIFO circuits, which are necessary to
decouple variations in array access latency versus the synchronous Read
latency timing specifications of the DRAM.

12.4 WRITE DATA LATENCY TIMING AND DATA

DEMULTIPLEXING
In this section, we consider the Write data demultiplexor circuits or what are
more commonly known in DRAM circles as the write pipes. This set of cir­
cuits follows the data capture latches of the input circuit path discussed in
Chapter 9. The purpose of these circuits is to convert the input serial data
stream into parallel data with a bus width and cycle time that is compatible
with the Prefetch data depth of the array. For example, if the per-bit Prefetch
depth of a column access is 8 bits, then each Write pipe will be required to
convert 8-bit data bursts into parallel data for each column access.
There are two distinct processes that occur when a Write command is
received. First, based on the predetermined Write latency, the input data
path is enabled to coincide with the arrival of data and any synchronous ref­
erence, such as a data strobe, at the data inputs. (Programmed Write latency
is sometimes formulaic to Read latency or discretely programmed in a mode
register, as discussed in Chapter 2.) Programmable Write latency is intended
to allow time between the receipt of a Write command and the arrival of
data at the data inputs to the device when operating at high clock frequen­
cies. The reason for this delay is obvious from our discussion of the com­
mand decode path, in that the delay through the command decoder can span
multiple clock cycles when operating at very high frequencies. Write
latency is defined as the delay between the Write command and the arrival
of the first bit of data of a Write data burst. This delay is measured in com­
mand clock cycles. An example of the Write latency specification is shown
in the timing diagram of Figure 12.23. Included in Figure 12.23 is the
parameter Sqss, which is discussed later in this section.
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 331

,PQS.

鲁瑞黑) n_n_n_n_n_tltljti_n_n_n-
Command

WDQS

Data (DQ)

Figure 12.23 Write latency definition.

Following a Write command and data capture, there is a serial-to-paral-

lei data conversion, or what is often referred to as data demultiplexing, in
the data path. The primary purpose for data demultiplexing is to match the
data rate at the input pins with the column cycle time of the array. For exam­
ple, if the array column cycle time is 3.2 ns, and the burst granularity is 4,
the maximum data rate would be 1.25 Gb/s. If we change the burst granular­
ity to 8, then the data rate becomes 2.5 Gb/s. Figure 12.24 is a top-level
view of the data demultiplexing operation for reference.

wData

wDataLat

Figure 12.24 Demultiplexor in DRAM Write data path.

In the remainder of this section, we first consider the necessary steps to

meet Write latency timing. Predicting Write data arrival based on program­
mable Write latency values is fundamental enough to DRAM Write data
capture that any future specifications implemented in high-performance
DRAM will continue to benefit from the current state of the art. However,
332 Chap. 12 Control Logic Design

some have proposed DRAM with interfaces that are based on data packet
protocols [16], which would potentially eliminate the heed to track Write
data latency. At the present time, a packet-based DRAM interface is
intended for niche applications. The packet-based interface is attractive
because of narrow input ports and, hence, low pin count. However, the per­
pin data rates are necessarily higher in order to match the bandwidth avail­
able from alternative devices with wider input ports. There has not been
widespread use of packet-based DRAM, so this protocol is not discussed in
this text. Following the discussions of Write latency timing, we examine, in
greater detail, one method for demultiplexing the data driven from the input
circuit path and how data is assembled and eventually driven to the periph­
eral array logic circuits. The discussion about demultiplexing will also con­
sider issues related to clock frequency division that, for a DRAM, is often
advantageous when dealing with data rates greater than 2 Gb/s.

12.4.1 Write Latency Timing

When SDRAM operating clock frequency specifications were based on
the column cycle time capabilities of the DRAM, random access was possi­
ble and data Prefetch was unnecessary since the specification allowed full
random, cycle-to-cycle Read or Write accesses. For SDRAM, there is not a
programmable Write latency specification; the Write data arrival time is
concurrent with the Write command at the device inputs. Write data can
burst for several cycles, but the LSBs of the column address are automati­
cally changed for each data bit and driven with data to the peripheral array
logic. For DDR SDRAM, however, data Prefetch is a necessity in order to
ensure that the throughput of the data I/O circuits do not become data lim­
ited because of the array cycle time. As was emphasized earlier in this chap­
ter, the DRAM array architecture and performance has changed very little
between early, asynchronous designs and high-performance designs.
Aside from Write latency, another innovation that is part of the DDR
DRAM specification is the data strobe. The data strobe is a burst clock that
is aligned to the center of the data eye for Write data and nominally aligned
to the transition edges of the data eye for Read data (from the point of view
of the DRAM). For Write data, the data strobe serves as a source-synchro­
nous [17] signal that relieves some of the internal clock distribution difficul­
ties that result from the SDRAM specification. SDRAM uses the same
system clock as a timing reference for both data and commands; whereas,
DDR DRAM protocol specifies separate timing references: the data strobe
for data and the C/A clock for commands. Having a source-synchronous
data strobe allows tighter timing for data capture and also allows internally
delay-matched data and clock distribution between the data inputs and the
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 333

data latches, as described in Chapter 9. For commodity-level DDR DRAM,

the data strobe is bi-directional. For high-performance DRAM, such as
graphics DDR DRAM, there are separate, unidirectional data strobes that
are distinguished based on Read and Write commands.
There are two fundamental timing specifications related to the data
strobe. The first timing specification is concerned with alignment of the
strobe to the command clock, while the second specification is the data
setup and hold time relative to the data strobe. The data strobe transitions
are specified to nominally align with transitions on the command clock to
help accommodate deterministic latency relative to Read or Write com­
mands. Because of both data-dependent noise and channel imperfections
(simultaneous switching operations, clock jitter, inter-symbol interference,
and channel cross talk), it is necessary to specify a finite amount of skew
between the data strobe and command clock. This skew specification, illus­
trated in Figure 12.25, is commonly referred to as 力0ss. Skew between the
command clock and the data strobe does not alleviate the data setup and
hold timing specifications between the data strobe and data signals. The
timing parameter tDQSS is a component of Write latency and specifies that
valid data must fall within a deterministic timing slot relative to the com­
mand clock.

歌1-I_ I-I_ I—I r~~l I—|J—|

Data --------------- OOOO------------
woqs L J一|]_~[]
^DQSSnOm 1 1
I I

口曲 ---------------- kt>000----------
।
WDQS ----------------------------------- 节P广产号d _ ----------------------

:^DQSSmaX

加 ------------- OOOO-------------
Q ----------- T_J-----------------------
wd s
^OQSSm'n \ 1

Figure 12.25 Write data strobe and Sqss definition.

Another purpose for variable Write latency is to allow time to enable

and disable the Write data receivers. For high-performance DDR DRAM,
more aggressive receiver designs are often necessary. More sensitive
334 Chap. 12 Control Logic Design

receivers are often not self^biased but instead require a biased front end
(Chapter 9). For some generations of DDR DRAM, the data strobe is bi­
directional and center terminated (the signal is terminated to the middle of
the input signal swing range). Also, if the input data receivers are biased to
an active state, the receivers constantly consume power. Write latency
allows time between the Write command and data so that the receivers can
be enabled only during valid Write data periods. On the other hand, with
supply terminated signaling (high-performance GDDR designs implement
supply terminated signaling), the receivers can be turned on with little
power consumption because if the bus is idle, the termination resistors pull
the input signal to the supply voltage. Also, with a supply-terminated data
strobe, there is less likelihood of false strobe toggles propagating into the
device due to noise occurring around the switching point of the receiver.
With a center-terminated bus, there is a very high probability of false tog­
gles on the data strobe if the receiver is enabled during idle bus states,
which, depending on how the logic is implemented, could lead to data and
logic failures in the capture and demultiplexing logic. Overall, Write
latency is a timing mechanism that increases the timing margin between
decoding the Write command and the arrival of input Write data, and it also
allows time to enable and disable the data strobe and data input receivers fbr
more reliable, lower power operation.
Without getting distracted by the difficulties posed by specification-
related problems such as termination or protocol, the fundamental design
problems related to Write data latency for high-performance DDR DRAM
are associated with the separate clocking domains of the command clock
and Write data strobe (WDQS). There is some relief in that the WDQS is
specified to have a bounded skew (Idqss) relative to the command clock.
Referring back to Figure 12.13, we see that connected to the command
decoder is a block labeled Write latency sequencer. This block essentially
represents a counter that counts command clock cycles in order to align tim­
ing signals generated from the command clock domain to converge with the
arrival of data from the inputs of the device all timed to Write latency. The
problem with this configuration is that the data I/O are often located a long
physical distance from the command decoder, so that sending a signal
across the die to synchronize the completion of the demultiplexing function
requires very careful, customized design.
Figure 12.26 defines signals that we discuss fbr the remainder of this
section. Signals Write[0] and Write[2] are driven from the Write sequencer
logic. These signals are used for completing the sequence between tracking
the deterministic Write latency in the command clock domain and the
arrival of data indicated by valid transitions of the WDQS signal. We con­
sider the Write latency sequencer to be part of the command decoder cir­
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 335

cuitry. The timing diagram in Figure 12.26 shows the relationship between
Write[0,2] and the assemblage of Write data within the demultiplexor block
of Figure 12.24. Also, Figure 12.27 is a diagram showing the physical rela­
tionship between the data input circuit paths and the command decode and
latency timing circuits.

Write[0J or Write⑵
from Write latency
sequencer _ L
Figure 12.26 Write[0,2] definition.

The diagram in Figure 12.27 shows a matched-mode timing method for

enabling the demultiplexing circuits relative to the arrival of data based on
Write latency. We refer to the topology in Figure 12.27 as t4matched-mode,5
because the CMDCLK[0:3] signal distribution is delay matched to the
divided WDQS distribution, which is labeled t2 in Figure 12.27. These two
disparate circuit paths can be nominally delay matched because the CMD・
C£K[0:3] signals can be point-to-point routed to the demultiplexor timing
circuits while the WDQS and C/A CLK signals must be branched and
matched across a physically wide routing area (see Chapter 9).
There are a few interesting features in Figure 12.27 that explain how a
mid-perfbrmance device can time Write data latency using a "brute-fbrce,“
match-mode method. First, notice that as the C/A CLK and WDQS enter the
device, both signals are frequency divided. Frequency division is imple­
mented to gain greater timing margin for logic in the data demultiplexor cir­
cuits and for distribution and capture of the Write[0,2] timing signals into
the demultiplexor sequencing logic. The data capture latches for both com­
mand and Write data are clocked by the full-frequency clocks. It is only fol­
lowing the capture latches that frequency division is utilized to clock the
downstream logic. In the example shown, the input, C/A CLK is divided
into four phases (CMDCLK\Q:3]), This clock division can be accomplished
using a phase generator circuit (Chapter 11), The key to achieving accurate
Write latency timing is to align the arrival of the divided WDQS signal with
timing and sequencing signals Write[0,2]. These signals are synchronized
by CMDCLK[0,2] from the Write latency sequencer at the command
decoder and distributed directly to the frequency divided WDQS domain.
336 Chap. 12 Control Logic Design

The Write[0t2] signals (these signals are labeled 0,2 because they are only
generated from phases 0 and 2 of CA/OCZK[0:3]) are then retimed at the
control logic for the demultiplexor logic circuits. The reason Write[0t2] is
retimed at the demultiplexor circuits with the command domain clock is
because CMDCLK\Q:3] is continuous; whereas, WDQS is a burst strobe that
only toggles when data is driven into the device (Figure 12.25).

---------- Array Read data

Demultiplexor CMDCLK[0:3]
K Write data

WDQS

by 2
C/A clock distribution
matched to frequency
divided input delay (f2) of
divided Write strobe (WDQS)

CMDCLK[0:3] Write[0,2] /
>

Command
Cl" decoder

C/ACLK Divide CMDCLK[0:3] i

C/A CLKb by 2

Figure 12.27 Relationship between data input path and command decode path.

What we are trying to achieve through all of this delay matching is to

predict the arrival time of data that is captured in the source-synchronous
WDQS timing domain relative to the Write command received in the C/A
clock domain based on the predetermined Write latency. One key to the suc­
cess of this method is delay matching between the WDQS capture path, C/A
capture path and the C/A CLK[0:3] distribution path (%).
An alternative method to consider fbr retiming the latency timing sig­
nals is to reduce our reliance on delay matching of the frequency-divided
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 337

distribution paths, indicated by t2 in Figure 12.27. We can do this by includ­

ing a clock alignment circuit, such as a PLL [20], which uses the input cap­
ture path as the reference to the phase detector. This configuration is shown
in Figure 12.28. Here, we have less risk in trying to delay match the long
interconnects routed from the command clock domain out to the Write data
demux circuits with the WDQS distribution to the demultiplexor circuits.
We choose a PLL for this application because of the higher supply noise
rejection ratio when compared to a DLL (Chapter 11). With this configura­
tion, we are still required to delay-match all of the input capture paths but
we are not required to try to match the global clock distribution of the com­
mand clock with the input delay (labeled t2 in Figure 12.27).
Using a branch of the input capture clock tree as a reference to the
phase detector of the PLL forces the branches of the globally distributed
command clock tree to align with the frequency divided WDQS clock at the
demultiplexing circuits. Remember that in the cases we are examining, the
WDQS is a burst strobe. If WDQS were a continuous clock, we could align
the Write clock domain with the command clock domain directly. But the
scheme shown in Figure 12.28 aligns the distributed command clock with
the predicted arrival time of the WDQS based on delay matching between
all of the input capture paths indicated as tin in Figure 12.28.
The problem with the matched-mode methods just described is that as
process scales, device variability becomes a significant problem. Another
consideration is that across large physical distances on the die, there can be
voltage gradients that will cause further mismatch and clock skew across
the die. All of this is exacerbated by higher clock and data frequencies. The
static portion of these sources of timing mismatch can be mitigated through
training sequences for adjusting Write data timing in the data capture paths
and fbr setting the Write data latency relative to the command decoder tim­
ing. These training sequences are mentioned in Chapter 9, but because the
training sequences are specific to a particular DRAM specification, it is
enough to know that the adjustment of the data capture and data latency tim­
ing circuits is primarily based on exchanges of expected data patterns across
the data channels and that these timing relationships may need to be
retrained periodically. Any necessary retraining of timing relationships
causes a great deal of complication in the DRAM specification. However,
many training sequences that have been proposed or implemented are only
set during device power cycles or reset sequences and do not involve peri­
odic re-training. In either case, clock alignment circuits, such as a PLL,
adjust long-term voltage and temperature drifts to help maintain synchro­
nous alignment.
338 Chap. 12 Control Logic Design

PLL clock locked to

divided Write strobe /
(WDQS) across distribution

Figure 12.28 PLL solution for Write latency retiming.

12.4.2 Write Data Demultiplexing

Figure 12.24 (shown in the previous section and repeated in Figure
12.29) is a top-level diagram for a single bit slice showing how the Write
latency timing signals, Write[0t2] converge with the arrival of the serial
data stream at the demultiplexing circuits. The Write[0,2] signals are used
to ensure that data captured in the critical, high-speed capture path is cor­
rectly retired to the DRAM array In this figure, we see that data enters the
device as a serial stream at each data I/O. Because we are looking at a high-
performance DDR data path, the capture latch is implemented as a double­
edge latch consisting of two capture latches: one latch for rising strobe edge
data, Dr, and one latch fbr falling strobe edge data, Df. Following high­
speed data capture, each of the data outputs of the capture latches is now
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 339

one half of the input data rate. In our example data path, there is a fre­
quency-divided, four-phase clock. This clock is distributed to all of the data
bits associated with the WDQS signal and operates at one quarter of the data
rate. Each phase of the clock is used to capture either the rising edge data
from the capture latches (Phases 0 and 2) or the falling edge data from the
capture latches (Phases 1 and 3). At this set of latches, the data rate is down
to one quarter of the of the original data rate at the DRAM inputs. The
demultiplexing circuit then assembles the bits into a parallel set of data that
is driven, along with the wDatLat signal, to the peripheral array logic,
where it is latched and held until the data is driven on the I/O lines of the
DRAM array logic.

wData

wDataLat

Figure 12.29 Input data bit slice.

Figure 12.29 is a simplification of some of the design choices that must

be made in order to implement a robust solution for capturing data across a
physically wide data path. Fortunately, one provision that is part of the spec­
ification for high-performance DRAM is that multiple data strobes are
available for wide data paths. For instance, a 16-bit wide device would have
two data strobes: one for each byte [18]. The data strobes are generally
placed at the center of the data word. Depending on the pad spacing, there is
a relatively long routing distance from the center of a group of data pads out
to the end of the data path pad row. It is this routing distance, combined with
the maximum operational data rate, that can determine the circuit topology
for the control signals used to time the demultiplexing circuits.
340 Chap. 12 Control Logic Design

There are two primary choices when determining the optimum circuit
topology for demultiplexing across a data word. Figure 12.30 illustrates
these two choices. The first example is shown in Figure 12.30a. Here, the
control and timing logic is centered within the data byte with the timing and
control signals globally distributed to each nibble on either side of the con­
trol logic block. If the distribution delay of the timing and data select signals
(DSel) is less than the column cycle time, then this is a viable configuration.
Notice that the latency timing signals (Write[0,2]\ which are inputs to the
control logic, are timed to coincide with the input WDQS, The input data is
clocked into each of the demultiplexor circuits by the distributed WDQS
signal. For each burst, the control logic at the center of the data byte drives
control and timing signals to each of the demultiplexor circuits concur­
rently For this case, the demultiplexor circuits operate independently when
capturing data but the timing of the signals distributed from the centralized
control logic block retimes the data driven from each of the demultiplexor
circuits.
wDatLat

wDatLat

Figure 12.30 Circuit topology choices fbr Write data de-serialization.

Alternatively, in Figure 12.30b, the control logic is localized for each

bit in the byte. There is a central block of logic that times the wDatLat sig­
nal, which is bundled with the Write data and used for latching and retiming
the Write data in the peripheral array logic. The centralized logic in this case
is an additional demultiplexing circuit that is used as a model of the other
demultiplexors in order to time the arrival of data. There is delay margin
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 341

built into the output timing of wDatLat from this circuit to compensate for
any timing skew between each bit within the byte. Part of the compensation
must account for Eqs* Typically, 30ss is +/-0.5 of a bit time (often
referred to as unit interval). If tDQSS violates this specification, then the
latency timing circuits may align to the incorrect starting bit of a burst. The
t[)Qss compensation is handled by the timing domain crossing that is indi­
cated by the convergence of the Write[0t2] signals with the data strobe dis­
tribution generated in the WDQS timing domain. By tracking the status of
the demultiplexing circuits relative to the arrival of the Write[0f 2] signals,
Sqss can be compensated within the demultiplexor tracking circuits. The
primary difference between the circuit topologies in Figure 12.30 is that the
centralized logic in Figure 12.30a generates all of the timing and control
signals for each of the demultiplexor circuits and also generates the bundled
latching signal wDatLat. Whereas, in Figure 12.30 (b), the control and tim­
ing logic is repeated within each of the demultiplexor circuits with only the
latency control logic and 也”logic contained in the centralized circuit
block. The circuit of Figure 12.30b can operate at a higher frequency
because there are no globally distributed control signals, but this circuit suf­
fers from increased power and area relative to the circuit of Figure 12.30a.
Now that we have examined some of the details of Write latency track­
ing and the global data capture topology of the demultiplexor circuits, we
turn our attention to the design of an individual demultiplexor circuit. To
illustrate the demultiplexor operation, a simple flip-flop based circuit is
shown in Figure 12.31. The circuit in Figure 12.31 is an example of only
part of what is required for a full demultiplexor solution. We see from the
timing diagram in the figure that the purpose of the demultiplexor circuit is
to successively expand the data valid window at each pipeline stage until
the data can be safely transferred to the array logic. Because the contents of
the flip-flops are over-written on successive Write cycles, this solution is
only viable if data is transferred to a retiming circuit within three-bit times
(between WDS[3] and WDS[2] in this example) using the wDatLat signal.
Remember from the previous discussion that wDatLat results from the con­
vergence of the Write latency timing signals sent from the command decode
clock domain, with the timing signals generated by the central control cir­
cuit in the Write strobe timing domain.
342 Chap. 12 Control Logic Design

Data ---- 领玲⑥--------

wdqs -LFLTLI
6 ZDOOCZZZZ
二1 X7XT>C
WDS]P] | | |

Figure 1231 Flip-flop-based Write data demultiplexor

One problem that is not obvious from Figure 12.31 is that with a
divided Write strobe, there are two possible relationships between the
receipt of a Write command and the output state of the Write strobe fre­
quency divider. The timing diagram in Figure 12.32 illustrates this case.
This diagram is taken from a GDDR DRAM data sheet and shows the case
where two Write commands are issued with an extra command cycle
inserted between the minimum possible Write cycle time. This Write com-
mand sequence is often referred to as a nonconsecutive Write operation.
First, consider that the alignment between the WDS[0:3] divided strobes is
the same as that shown in Figure 12.31. Then, when the extra command
cycle shown in Figure 1232 is inserted between the Write commands, an
extra toggle occurs on the WDQS signal, and invalid data is loaded into the
data capture latches.
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 343

Figure 12.32 Nonconsecutive Write commands.

Figure 12.33 shows how the divided strobe alignment changes between
the first set of Write data and the second set of Write data. Focusing on the
WDS[0:3] signals, for the first Write, WDS[0] aligns with the first rising
edge of WDQS for the data burst (burst of four, for this example). Because
of the extra cycle inserted between the Write commands, WDS[2] aligns
with the first rising edge of the WDQS fbr the second data burst. Because
the divided strobe signals are operating at one half of the frequency of com­
mand clock, the command to WDQS alignment can assume two different
states. In order to correctly track Write data latency, and drive the correct
data out of the demultiplexor circuit, we need a method to detect the two
possible command-to-data alignment states fbr a divided WDQS. This is the
reason that the Write latency timing signals are paired as Write[0,2].
Following a system reset or power cycle, the WDQS divider and the
Write latency control logic in the command decoder are reset to a known
condition. Depending on the logic or clocking implementation at the Write
decoder and demultiplexor logic, one of the latency timing signals fires fbr
the first Write command and establishes an alignment relationship between
the command decoder logic and the output of the WDQS divider for the first
Write command. For all subsequent Writes, the correct Write[0,2] signal
will align accordingly with the WDS[0,2] signals to correctly time the out­
put of the demultiplexor circuit. We should note that if toggles occur on
WDQS that do not correspond to a Write command, fbr instance, because of
overshoot or undershoot on WDQS due to signal integrity issues, then
untracked toggles may occur on WDS[0:3]. Under these conditions, the
demultiplexor circuits must be designed to track the correct Write[0f2] to
WDS[0,2] alignment, regardless of starting conditions. This is a specific
problem the designer must evaluate, yet solutions to this problem are not
fundamental to this discussion.
344 Chap. 12 Control Logic Design

口而—<XXX—OOOO--------------------------

利 Q^LrLTLTUWJ -----------------------

Dr ><toX2^O<0bX2rX '

Df

WDS⑼ I 一L|—L J― I

IVDSpj I | [ I I
乂 Qg(a) >f~X Q"(b)

X Q 、⑻ X~X Q；1(b)

Q0(a) Q0(b) 乂~~

* Q1 ⑶ Q1(b) X-

>ic_5^"x~x_5^n>c-
I I ____
wDatLat : | | : | I
I I
WritefO] | ] | |
Latency timing

； ；
from command _____________ i ]_________________
clock domain
W 丽 | |
rPls I—I I―I
Figure 12.33 Divided Write strobe alignment fbr nonconsecutive Write commands.

Additional signals are illustrated in Figure 12.33 that help us examine a

simplified solution to demultiplexing fbr the divided WDQS. This discus­
sion is further aided by the circuit diagram of Figure 12.34 [19]. The first
feature to notice is the inset to Figure 12.34. Here, we show a modified data
latch circuit that includes both rising edge latches (rFF) and falling edge
latches (/FF). Of course, the representation as flip-flops is only fbr illustra­
tive purposes; in practice, high-performance latch circuit designs are used
(Chapter 9). The outputs of this latch design are represented as Dr and /yin
the timing diagram of Figure 12.34. The performance of the latch circuit
must allow a data transfer to occur between the rising edge and falling edge
of WDQS at the highest data rate and with the maximum allowed WDQS
duty-cycle distortion. The resulting output of the latch circuit in the inset of
Figure 12.34 is alignment of Dr and Df following each falling edge of
WDQS. Continuing with the example of Figure 12.34, the operation of the
demultiplexor circuit is similar to that of the circuit in Figure 12.31 except,
in this example, the Write[0,2] signals are used to track the alignment of
Sec. 12.4 Write Data Latency Timing and Data Demultiplexing 345

Write commands with the WDS[0,2] signals to mux the correct data into the
demultiplexor circuit path. The block labeled "Pls" represents centralized
logic to control the timing of the outputs QO-3 of the circuit to align the data
for transfer to the peripheral array logic. The wDatLat signal is used to
accommodate the data transfer to the peripheral array logic. By examining
the timing diagram in Figure 12.33, we see that the second set of Write data
aligns to the complementary WDS signals relative to the first Write com­
mand because of the nonconsecutive Write condition.

Figure 12.34 Circuit diagram for divided WDQS demultiplexor.

In this section, we have examined some of the fundamental problems

for tracking latency and demultiplexing Write data. Two fundamental prob­
lems emerged in this discussion. First, that Write latency is specified rela­
tive to the command clock domain, and, as such, requires designers to
employ a synchronization method to transfer associated timing information
between the command clock and WDQS timing domains. And, second, that
the serial Write data on the wide I/O data bus must be demultiplexed and
retimed for each individual bit slice before it is transferred to the peripheral
array logic.
Keep in mind that much of the preceding discussion may be oversimpli­
fied in some cases. This is necessary so as not to obfuscate the fundamental
346 Chap. 12 Control Logic Design

problems encountered by the DRAM logic designer when executing a full

Write data capture design solution. As was stated at the beginning of this
chapter, DRAM peripheral logic design is as much an art form as it is a sci­
entific exercise. In the next section, we visit the Read data logic path. The
problems encountered in the Read data path have many similarities to those
of the Write data path. Similar to Write data, Read data access delay is spec­
ified with a deterministic latency relative to the command clock, and there
is also a data format conversion (parallel-to-serial, in this case) necessary
fbr compatibility with the array data path and output data rates.

12.5 READ DATA LATENCY TIMING AND DATA MULTIPLEXING

Read data accesses, similar to Write data accesses, are composed of two
major logical operations: latency timing and data formatting. We saw that
Write latency timing is controlled by transferring timing and synchroniza­
tion signals from the command clock domain to the WDQS timing domain.
Write data is also reformatted from a serial data stream to parallel data
through the demultiplexor circuits. Write data reformatting is necessary so
that the I/O data rate is bandwidth matched to accommodate the array col­
umn cycle time. Stated another way, a 2 Gb/s serial data stream matched to
a 4 ns column cycle time would require an array data bus width of 8 bits per
data I/O. This same bandwidth matching occurs fbr Read data where Read
data on the array bus cycles at 250 MHz with an 8n Prefetch providing
enough data bandwidth to feed I/O circuits operating at 2 Gb/s.
Read data is managed much the same way as Write data except that
with Read data the ordering of the array access and I/O transfer are
reversed. For Read commands, the array is immediately accessed and data
is transferred from the HFFs, in the peripheral array logic, to the Read data
path of the peripheral logic interface. On the other hand, for Write com­
mands, Write data is captured in the input circuit path, then assembled in the
demultiplexer circuits. Write data is then retired to the array with an array
access. In the Read data path, there is a multiplexor circuit that is the con­
verse of the demultiplexor circuit in the Write data path. Just as the Write
demultiplexor is often referred to as the Write pipes, the Read multiplexor
circuits are often referred to as Read pipes. Refer to Figure 12.35 for a sim­
ple block diagram showing the relationship of the peripheral logic interface
sandwiched between the array logic and the input and output circuit paths.
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 347

Peripheral array logic

(HFF/DCSA, CAS timing chain, etc.)

Penpheral logic
wDatLat DataValid
interface region

Write data 250 MHz Read data 250 MHz

x8/DQ x8/DQ

Command
decode
Column address
Write pipes Read pipes
registers wDatLat

Command/
,address latches

Figure 12.35 Peripheral logic interface Read and Write data paths.

Figure 12.35 illustrates the relationship between the Read and Write
pipes and how the data rate of the I/O circuit paths translate to array column
cycle time. Another feature shown in this figure is separated Read and Write
busses in the peripheral logic interface region. The separated Read/Write
busses allow concurrency for Reads and Writes, which may be necessary
for tight Read-to-Write or Write-to-Read data timing. Also consider that the
Read data array accesses are controlled by the timing of the command clock
domain, while Write data array accesses are timed to the WDQS timing
domain. If we consider SQSS combined with variations in address access
delay for Reads 面)，there could be overlap on a shared Read data and
Write data bus in the peripheral logic interface region. There will be a tim­
ing diagram later in this discussion that will better exemplify this possibil­
ity. Another notable feature of Figure 12.35 is the address bus with the
bundled signal, colCyc*. Whenever a Read command is received, the
address and colCyc* signal are immediately driven to the peripheral anay
logic to get the lowest possible array access latency. For Write data, the col-
Cyc* signal is fired based on the wDatLat signal that comes at the end of the
data demultiplexor operation (Section 12.4), Recall that Read data accesses
348 Chap. 12 Control Logic Design

are initiated as quickly as possible to get the lowest possible array access
latency, while Write data accesses require that the column address is held
until all of the data of a burst is fully captured in the demultiplexor circuits.
Let us look at an example of a Write-to-Read access to see how the
array column cycle time is maintained for data accesses. Figure 12.36 is a
timing diagram showing a series of column accesses. In this example, there
is a Write command followed by a Read command, which results in a mini­
mum array column address cycle time at the peripheral array logic. We need
to be aware that a colCyc* initiated by the wDatLat signal must have timing
similar to a colCyc* initiated for a Read command originating from the
command decoder. By having similar timing relative to the launching event,
we can maintain constant column cycle time between Read and Write data
accesses. In the timing diagram, Read latency (CL) is four command clock
periods, while Write latency (WL) is CL- 1=3. At the end of the Write data
burst, the wDatLat signal transitions. As outlined in Section 12.4, the wDat-
Lat serves two functions. The first is to retime data in a set of latches in the
peripheral array logic. The second is to enable the firing of the colCyc* sig­
nal to cycle data accesses from the array. Write data is then overdriven
through the sense amps and, ultimately, to the Mbit.
Again refening to Figure 12.36, a Read command is issued at the end of
Write data. For this example, the Read command occurs two command
clock periods, which is one column cycle period, from the end of the Write
data. We see that the tight timing of the Read command relative to the end
of Write data results in close to a minimum column cycle period transition
on the colCyc^ signal between the Write and Read accesses (labeled t盯r in
the figure). The chip architect must decide between a shared data bus or two
unidirectional busses for interconnect between the peripheral logic interface
region and the peripheral array logic fbr Read and Write data. Variables
such as bus loading, interconnect routing, column cycle time, data through­
put, Read and Write access concurrency, die area, and any other perfor­
mance specification-related issues must be considered when making this
decision.
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 349

b rLrLrLruT_rLn_rLpLrLrLrLn_n_rLn_
一二二;二一"x^>c —:
wDatLat

cdCyc*
I_r
Address
bus x Write address ；* Read addr 嬴
Write data bus
X
Read data bus (late access)
^>1
-X
Z)1CZ
Read data bus (early access)
WL=3 ] !
CL=4

DQ
oooo
Write data
^>00<>
Read data
DQS
LTLTLJ
Figure 12.36 Timing diagram for internal data bus utilization.

Another important timing parameter for the peripheral logic interface

design is address access latency, labeled 桃 in Figure 12.36. This delay is
measured from the firing of colCyc^ to the transition of data at the output of
the peripheral array logic. There are two sets of Read data shown in Figure
12.36. The first set of data, labeled uRead data bus (late access)n shows
address access timing where data arrives late in the access cycle but early
enough to satisfy the programmed value for CL. The second set of data,
labeled “Read data bus (early access)“ shows address access timing that
occurs early in the access cycle, well ahead of the programmed CL delay.
Variations in Iaa are fundamental to the operation of the DRAM array.
Array access delay, measured from the initiation of a change in column
address, depends solely on timing models formed by delay chains that are
used to reflect variations in on-die process, voltage, and temperature condi­
tions. For this reason, array access delay varies by a wide range and,
because it can vary by multiple clock cycles, Read data cannot be synchro­
nized in the peripheral array logic but is instead synchronized in the output
circuit path. This means that the Read data path design must take into
account the minimum possible access delay as well as the maximum possi­
ble access delay in order to maintain continuous data throughput. The impli­
cation is that for minimum access delay, any data driven into the data path
must be held until it is driven through the synchronous portion of the output
circuit path, based on the programmed CL value. At the same time, for max­
350 Chap. 12 Control Logic Design

imum array access delay, the data must arrive at the synchronization circuits
ahead of the expiration of the programmed CL value.
For the remainder of this section, we will examine the two main compo­
nents of Read data timing managed within the peripheral logic interface.
First, we look at the purpose of the Read pipes as well as the performance
characteristics necessary for the Read pipe circuit block. After looking at
the Read pipe circuitry, we examine synchronous Read latency tracking fbr
a high-performance DRAM and look at an example circuit used fbr tracking
Read latency over various operating conditions.

12.5.1 Read Data FIFO

The Read pipe is functionally a FIFO. Its purpose is to absorb array tim­
ing variation （Z/力）and maintain data throughput between the array data path
and the synchronous portion of the output circuit path. For the remainder of
this chapter, we will refer to the Read pipe as the Read data FIFO or Read
FIFO. Looking back at Figure 12.36, we see how variations in 爆，relative
to CL, cause the Read data to have an indeterminable timing relationship to
the synchronous Read timing control. There are several possible design
variations for the Read data FIFO. For mid-to-low frequency DRAM, there
is very little depth required for the Read data FIFO. Often, the CL value is
low and the burst granularity is high to ensure that there is sufficient array
timing margin to achieve high yield for commodity level devices. However,
for high-performance DRAM, the Read data FIFO requires much more
attention. With lower column cycle times and higher data rates, the storage
and timing performance of the Read data FIFO is more important. In this
section, we look at a design example fbr a Read data FIFO. This examina­
tion considers aspects of the DRAM specification to determine the perfor­
mance parameters necessary fbr a successful design. We do not study a
specific circuit design fbr the FIFO but instead outline the performance
issues associated with the FIFO design and suggest some alternatives fbr the
FIFO architecture.
Figure 12.37 is a block diagram showing the position of the Read data
FIFO between the peripheral array logic and the output data serializer. In
this example, a four-phase clock drives the output serializer. This clock is
frequency-divided and quarter-aligned and is discussed in Chapter 11. Each
rising edge of the four-phase clock is separated by a single bit time so that
output data transitions correspond to each rising edge of the output data
clock. We will see that using a divided, four-phase clock in the output cir­
cuit path is an important consideration for the Read data FIFO design.
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 351

Figure 12.37 Read data FIFO reference.

In our example, the array data path delivers an eight-bit data Prefetch
for each DQ. The eight-bit Prefetch is delivered to the Read data FIFO bun­
dled with the data Valid signal, which is discussed in Section 12.3.3 and is
shown in Figure 12.22. The data is then reduced to a fbur-bit width through
the Read data FIFO and delivered to the data serializer. The output data seri­
alizer reduces the data from four bits down to singular output data as well as
retimes the data to the output data clock. There are a few functional details
about the FIFO that need to be clarified. One is that the Read data FIFO
must be capable of multiplexing and sequencing the correct four bits of data
through the output data path. For some DRAM specifications, there is what
is called critical word first. This feature uses a lower column address bit to
indicate which nibble is output first in an 8n burst (a choice between 0-3 or
4-7). This feature is a part of the DDR SDRAM specification and is some­
times referred to as burst ordering.
There is another multiplexing function necessary in the Read data
FIFO. Most device specifications call for commands to be referenced to the
positive transition of the command clock. When the clock is frequency
divided, there are two phases of the output data clock referenced by the pos­
itive edge of the command clock. For our example, phO and ph2 are consid­
ered proxies for a positive transition on the command clock. When data is
routed from the peripheral array logic to the Read data FIFO, there is no
output clock phase information included. A combination of CL and Read
352 Chap. 12 Control Logic Design

command alignment to the four-phase output data clock results in two pos­
sible data-to-clock alignments. Figure 12.38 shows an example of a noncon-
secutive Read access where the first Read forces the first data bit of an
access to align to output data clock, ph2 and the second Read forces the first
data bit to align to output data clock, phO. The remaining bits are also multi­
plexed to the conect clock phase so that the phase order fbr the first access
is 2,3,0J and the second output phase order is 0,1,2,3.
^mand ULrm_Rwi_rLnjTrLrUmjnjuU'L

8mmand j ]

Cl =10

Output -------------------------------------------------------------------- ----------------------------------------- ----------------------------

Read data X_ 2.3,01 X 0/23

Figure 12.38 Nonconsecutive Read access and multi-phase Read clock.

The starting phase alignment information originates from the Read

command latency circuits. These circuits are an extension of the command
decoder circuits and are discussed later in this section. The starting phase
information is determined by the programmed CL value and the relationship
between the command clock and the multi-phase output data clock. Because
the Read data FIFO only carries array data, another FIFO is necessary to
feed the starting phase information to the multiplexing circuits in the Read
data FIFO. Figure 12.39 shows the secondary FIFO labeled “orthogonal”
FIFO. It is so named to aid in visualizing this circuit as providing a perpen­
dicular control term to the Read data FIFO.
Another notable feature in Figure 12.38 is the timing parameter, tjg.
Remember from previous discussions in this text that high-performance
DRAM commonly employ clock alignment circuits that are used to align
the output data and strobe to the external system clock. To accomplish this
feat, an I/O delay model is included in the feedback of the clock alignment
circuit to advance the output data clock by the delay of the input and output
circuit paths (Chapter 5 and Chapter 11). Because the data output clock is
advanced relative to the command clock by the asynchronous I/O delay,
once data is retimed and driven through the data output path, the data
appears nominally aligned to the command clock at the device I/O pins.
This configuration is necessary to simplify system design so that there is a
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 353

common timing reference for all DRAMs in a system. The side effects of
advancing the timing of the output clock for on-die Read latency tracking
will become more apparent later in the discussion about Read latency (CL)
tracking.

Peripheral array
interface

Read data dataValid

Figure 1239 Data alignment to multi-phase output clock through the Read data FIFO.

To get a better idea of how the FIFO must perform in the DRAM Read
data path, lefs examine the diagram in Figure 12.40. Here, we see a contin­
uous stream of consecutive Read commands. The boxes represent relative
percentages of the overall CL time for each operation of a Read command.
Notice that for the “Fast" operating condition, the time allotted for the FIFO
is the majority of the CL delay. Because the CL specification is based on
worst-case timing conditions, if the device happens to perform better than
the data sheet specification indicated for CL, there is no way fbr the system
to take advantage of the improved array performance. The reason for this is
that there may be several DRAMs in the system that perform worse or bet­
ter. Therefore, the FIFO must match the throughput and latency of the array
accesses with the expected data arrival based on the programmed CL value
for all of the devices. The device data sheet specifies the minimum and
maximum value for CL. The value of CL is based on the acceptable worst­
case performance of any one DRAM in the system. For the "Slow” operat­
ing condition, the FIFO period is much less than。力 and, fbr this case, we
are concerned with the length of delay through the FIFO circuitry (Jfi)，
354 Chap. 12 Control Logic Design

M______________________ CL^IO______________________
Command1r 1 ~~ 1 7__ ___ ___ __ __ __
c,ock rLpLrLrLrLn_rLrLTLrLrLti_rLrLrLn_rLn_

--------------------------- kxxx>oo<xxx>oooo
IUCI 一 I FIFO 「 「"；
I F'F。 | 正1_____
闻 Gl - I 一 垂. I ci
~ im1工 i : 石fo Tei
i Last bit out of FIFO from
Fast operating Y-j first Read command
conditions i /
Slow operating /
conditions ] 匕 >। J
I 露 fc,| . j FIFO | |

I■L I d J I Fljo 1 L |
R/l Qe I - 」 「丽 1工 I

Figure 12.40 Consecutive Read command timing.

Figure 12.40 is interesting in the way that it illustrates the asynchronous

nature of DRAM operation. Another point, which will be made again in our
discussion about Read latency, is the way in which the clock alignment cir­
cuit operates in the DRAM and its effect on the FIFO period. Notice that in
the example for the fast operating comer how the short I/O delay forces the
output clock closer to the CL expiration time. At the same time that the out­
put clock pushes out the internal data timing, the input delay and access
delay is pulled in due to the faster operating conditions. This forces data
from the array to be output closer to the Read command, making the overall
FIFO period much longer. The longer FIFO period forces a deeper FIFO
circuit design. Conversely, in the slow operating case, the clock alignment
circuit pulls in the output clock timing relative to the CL expiration time
due to the longer I/O delay. At the same time, the slower operating condi­
tions pushes the timing of the data access out from the Read command mak­
ing the FIFO period much shorter and constraining the FIFO design to a
lower forward latency （牡）requirement
When designing the Read data FIFO, the first step is to determine the
required depth of the FIFO. The depth of the FIFO determines the through­
put capability of the outputs to ensure that the output circuits are neither
overflowing with data nor starved for data. There are two critical bound­
aries that must be considered. The first is to make sure that when data is
read from the array, and the internal Read latency count has not expired, that
the Read data can be stored until it is driven through the output circuit path.
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 355

The peripheral array logic does not have storage available for data and, fiir-
thermore, if array accesses are purposely delayed, then a violation of mini­
mum column cycle time will occur for subsequent Reads during
consecutive burst Read operations. The second condition is that the delay
through the Read data FIFO is not large enough to significantly impact
access delay and, therefore, prevent the device from meeting the minimum
Read latency specified by the data sheet. For this case, the delay through the
FIFO needs to be minimized.
With these two criteria in mind, we consider an example for a high-per-
fonnance device. This example is taken from the specifications fbr a
GDDR3 high-performance graphics DRAM. The maximum data rate is 1.6
GT/s, CL=10 and BL=4. The example is illustrated through a simple, step-
by-step calculation procedure that shows some of the necessary characteris­
tics of the Read data FIFO when evaluating circuit design options for the
FIFO implementation:

1. First, because we need to deliver four bits of data fbr each column
cycle, we know that the column cycle time will be 2.5ns. Next, the
serializer circuit requires 4 bits of data at its input-one bit to align
with each phase of the Read data output clock. Therefore, the cycle
time for each bit at the output of the FIFO is:

收=总2.5 ns
T

(12.5)

where DR=1.6GT/s.
2. Next, we need to determine the maximum input-to-output delay QFL.
max), otherwise known as forward latency, allowed through the FIFO
in order to meet the minimum programmed CL value:
6ns > > 3.5ns array access delay (max., min.)
tCMD = 1 ns Command to colCyc delay
tsus = 0 5 ns Data setup time requirement at serializer
input relative to the Read data output clock
2.8ns > tjo > 1.5 ns asynchronous I/O delay (max., min.)

lFLmax = CL」CK -tCMD -AA -SUS -IO

tFLmax = 22ns

(12.6)
356 Chap. 12 Control Logic Design

The values for all of the timing parameters (except 山 the tpi
calculation are fbr the worst-case operating comer. The value kK is
the minimum period for the command clock frequency. The defini­
tion fbr and tjQ show the wide range of delay over which these
timing variables can vary.
3. Depending on the chosen circuit architecture, the preceding FIFO
design parameters are dependent on the FIFO depth. Given a series
of consecutive Read accesses, the FIFO depth is determined by the
number of column accesses that can be loaded into the FIFO before
the first access in the series is consumed. Read FIFO depth must be
calculated using the longest possible latency and shortest 爆 and %
delay. This timing difference will determine the maximum required
FIFO depth (FD) to maintain constant data throughput. We derive
the equation for determining the FIFO depth by looking at the Stfast
operating,, conditions in Figure 12.40. The FIFO depth is based on
the peak number of column accesses that can be simultaneously
loaded into the FIFO (for DDR I/O):

2CL tl ： + tCMD + tAA - + tO f

r^r\ / max . i、 min min min
FD - ( ------------
'BL + 1 ))--------------------------------------------------
tCCDmin

FD = 3.6

(12.7)

For fractional values of FIFO depth, FD must be rounded up to the

next integer value. If the CL value is increased relative to the value
defined in this example, the FIFO depth would have to be increased
accordingly. Likewise, any improvement in delay through the output
circuit path can result in increased FIFO depth.
From the preceding calculations, we find that our FIFO design must
have a maximum 2.2 ns forward latency, a minimum 2.5 ns/bit cycle time at
the output of the FIFO and a depth of 4. However, fbr increased data rates
and increased range of programmable CL values, the FIFO circuit can
become much more complicated. This is primarily because conventional
circuit designs, particularly at the FIFO output to serializer interface, can
have too much overhead and limit performance. For example, if the data
rate were to increase to 2.5 GT/s, then, with a four-phase output clock, the
per-bit cycle time at the output of the FIFO would be reduced to 1.6 ns.
Circuit implementations for the Read data FIFO can take many forms.
For high-performance DRAMs, the FIFO depth can increase greatly
depending on the minimum column cycle time. If we consider that mid-per-
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 357

fbrmance DRAMs typically have latencies of two or three cycles, with col­
umn cycle times of two clock periods, the latency-to-column cycle time
ratio is close to 1. However, for high-performance DRAM, such as a
GDDR4 device, latencies grow to 10 or 12 cycles and column cycle times
are four clock periods leading to latency-to-column cycle time ratios of 2-3.
Worse yet, for GDDR3 devices, the column cycle times are 2 and latency is
10 or higher leading to latency-to-column cycle time ratios of 5 or greater.
The result is that for high-performance devices there is a requirement for
deeper data FIFOs. A traditional pointer-based FIFO can be used for deep
FIFOs. The problem with these circuits is that output data muxing require­
ments can increase the output delay of the FIFO so that data cannot cycle at
the output clock cycle time.
Another option is to implement the FIFO as a set of series connected
latches controlled by asynchronous protocol controllers [14, 25]. This type
of FIFO implementation reduces the FIFO design complexity and can
increase performance at the output of the FIFO; but because the FIFO is a
series connection of latches, this type of design suffers from higher tpi and,
hence, longer latency. This text does not discuss specific FIFO implementa­
tions. The designer must choose the implementation based on the calculated
requirements.
For high-performance DRAM, synchronization at the output of the
FIFO can be complicated by the data rate. There are two primary alterna­
tives: either the circuit design meets the cycle time requirements through
specific circuit implementation techniques, or the width of the data path at
the FIFO output is increased at the cost of design complexity. Figure 12.41
shows an example of both approaches. The difference between the two
approaches is that for the low-cycle time approach, an aggressive circuit
implementation using dynamic logic techniques must be considered. For the
increased data path width approach, a method for toggling between the data
sets must be implemented and the data must be muxed correctly through the
FIFO circuits. This muxing would be in addition to, or a modification
the clock phase based muxing shown in Figure 12.39. Another consider­
ation when increasing the data width is the increased clock loading and
complexity of the serialization circuit.
358 Chap. 12 Control Logic Design

Ph10J II―|—I____ |一I_

DatalOJ^X Data[OJ /飞而
I
[0j

DataQ]) ^ata(0]

I
DonefO] | |
OE _____ H-

Select I

Figure 12.41 FIFO output synchronization solutions for multi-phase output clock.

In the diagrams of Figures 12.39 and 12.41, the signal OE is shown as

part of the output synchronization logic fbr the Read data FIFO. This signal
is the data output enable signal (OE) and its purpose in the synchronization
logic is to indicate that the output circuit path has consumed data from the
FIFO. For unidirectional data paths, the OE signal is used to indicate the
period fbr which valid output data is driven through the output circuit path.
The same can be said about OE for bi-directional data channels except the
OE signal also synchronously enables the data serializer (Chapter 10).
Implementation of the synchronous output logic depends on the FIFO cir­
cuit implementation. The OE signal is typically shifted through synchroni­
zation logic and timed signal transitions (labeled Done in Figure 12.41) are
sent back to the Read data FIFO for each bit consumed by the output circuit
path. The Read data FIFO responds by switching the data bits on the outputs
of the FIFO circuit.
The timing of the OE signal is generated by the latency control logic.
For high-performance DRAM, CL is deterministic. What this means is that
when the value of CL is programmed into the DRAM, the device will
deliver Read data exactly CL command clock cycles following a Read com­
mand. For memory subsystems, deterministic latency simplifies the data
interface so that the memory controller can time the arrival of data from the
memory device with a known relationship to the system clock. Tracking
Read latency on the DRAM side of the channel sounds relatively simple
until we consider that the asynchronous delays through the I/O paths can
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 359

span multiple clock cycles. Add to this fact that delay variation through the
I/O paths is directly impacted by process, voltage and temperature, which
can cause the delay to vary across clock period boundaries. What we find is
that achieving deterministic Read data latency is not a simple matter of
counting clock cycles. In the next section, we investigate difficulties related
to deterministic Read latency tracking. We also investigate how the inclu­
sion of clock synchronization logic (Chapter 11) on the DRAM simulta­
neously complicates CL tracking and accommodates synchronization and
clock domain crossing between the command clock and the output data
clock domains.

12.5.2 Read Latency (CL) Tracking

Thus far, in our discussion of the peripheral logic interface, we have
defined Read latency (CL) and how this specification affects the Read data
FIFO design. In Chapter 5 and Chapter 11, the output clock synchronization
circuit methodologies using either a DLL or PLL circuit were discussed.
Part of the discussion of clock alignment touched on the specification.
This timing specification dictates the maximum variation that data and any
source-synchronous strobes can have relative to the command clock of the
DRAM. Ideally, this timing specification is 0 ps. The clock alignment cir­
cuit is used to advance the output data clock by the delay through the output
circuit path (tout) in order to minimize At the same time, the reference
clock at the input of the phase detector of the clock alignment circuit is
delayed by the input distribution delay (tin) of the command and address
capture clock. To counteract the asynchronous I/O delay, a delay model of
the I/O circuits is added to the feedback path of the alignment circuit forcing
the reference and feedback inputs to a 0-degree phase relationship. Figure
12.42 is a simplified block diagram showing a DLL clock synchronization
path fbr a DDR DRAM.
When the command clock period begins to approach the asynchronous
I/O delay of the input and output circuit paths, tracking Read latency
becomes much more difficult. In early SDRAM, the clock period was much
greater than the I/O and command decoder delays and the command clock
served as both the input clock and output clock. In this situation, the
delay was essentially the delay through the output circuit path. The input
delay could be ignored since it was masked by the clock period, which was
long enough to encompass the input distribution and command decoder
delay. For these early devices, data was actually driven from the device at
CL-1 and tAc delay allowed enough setup time relative to the clock edge
occurring at CL that the system timing budget allowed data to be captured
with the command clock at the corresponding CL clock edge. If we fast-fbr-
360 Chap. 12 Control Logic Design

ward to DDR DRAM devices and beyond, we find a clock alignment circuit
on the DRAM generating an output clock domain separate from the com­
mand clock domain. The output clock domain is advanced relative to the
command clock by the output circuit path delay, similar to the diagram in
Figure 12.42. The clock alignment circuit is used to align the output data to
the clock edge corresponding to CL and, at the same time, minimize t^Ct0
predominately short-term components of output clock jitter.

Figure 12.42 Effect of synchronization circuit on output data timing (DLL example).

Considering that for high-performance DRAM, CL is still defined in

terms of the command clock, we must consider the implications of transfer­
ring timing information, established by the command clock domain, to cor­
rectly align with data output timing generated in the output clock domain.
Notice in Figure 12.42, we see that CL is defined relative to C/A CLK and
that the OE signal is synchronous with the RdCLK.
For high-performance DRAM, the clock period is often less than the
input and output circuit path delays. Additionally, as voltage and tempera­
ture vary, the output of the clock alignment circuit can vary by a relatively
significant amount of time. The preceding statements imply that the phase
relationship between the C/A CLK and RdCLK is indeterminate. Therefore,
passing timing signals from the C/A CLK to the RdCLK, while simulta­
neously tracking CL relative to the C/A CLK, is somewhat complicated. For
the remainder of this section, we will examine one method [20] employed
on several high-performance DRAM designs that both establishes a deter­
minate relationship between the C/A CLK and RdCLK domains as well as
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 361

tracks the relative delay variations between the two clock domains with
long-term changes in voltage and temperature.
The method for CL tracking that we are about to examine can be diffi­
cult to explain and, depending on how well it is explained, difficult to
understand. To facilitate our understanding of this method, we will break
the function of the circuits into two parts. The first part will explain how the
relative phase relationship between C/A CLK and RdCLK is comprehended.
By determining the phase relationship between the two clock domains, we
can establish a basis for tracking delay variations (resulting in phase align­
ment changes) that occur between the two clock domains. The second part
will examine how variations in the phase relationship between the two clock
domains are tracked and, ultimately, how we determine the precise internal
RdCLK clock cycle from which to fire the OE signal so that data is correctly
aligned CL cycles from the receipt of a Read command.
To summarize, for high-performance DRAM, precise CL control
depends on the inclusion of a clock alignment circuit. Without the clock
alignment circuit, simply counting the correct number of command clocks
as a basis for firing the data transmitter circuits will result in a large time
variation between the anticipated CL and actual CL. This variation is a
result of the asynchronous input and output circuit path distribution delay.
Therefore, if cycle-accurate CL is part of the protocol specification, then a
clock alignment circuit is a necessity.
If we refer to Figure 12.42, we see that there is already a circuit in the
clock alignment topology that tracks the phase difference between the C/A
CLK and the RdCLK domains. This circuit is the I/O feedback model used
in the clock alignment circuit. The absolute phase difference between the
C/A CLK and the RdCLK is given by:

t - 4~ / .
9 =叫_竺2冗rad
，CK

(12.8)

If tin + tout > tQ^ then the absolute phase difference is greater than a sin­
gle clock period. Given this situation, the method used to establish the phase
difference between the clock domains must account for this absolute differ­
ence and not just the fractional phase difference that might be present. Fig­
ure 12.43 is a circuit configuration that can synchronize a signal between
the two clock domains.
362 Chap. 12 Control Logic Design

Initialization
logic output.

StCA

C/ACLK

Figure 12.43 Synchronization between output data clock (RdCLK) domain

and command clock (C/A CLK) domain (DLL-based example).

The circuit diagram in Figure 12.43 shows a signal being transmitted

from the RdCLK domain back to the C/A CLK domain through an I/O delay
model identical to that used for the clock alignment circuit. Because the
RdCLK signal is advanced relative to the C/A CLK signal by the clock
alignment circuit, and the phase difference between the feedback clock and
C/A CLK is nominally 0 rad, then we can assume that the transmitted signal,
StCA, is synchronized to the C/A CLK domain. It is a simple case where
StCA is delayed by the same delay for which the RdCLK is advanced. Given
that StCA originates from the RdCLK clock domain, it is safe to assume that
StCA is synchronized to the C/A CLK domain following the I/O delay
model. Of course, there will be imperfections in delay matching of the I/O
model circuits, and there will be phase differences between the reference
and feedback clocks at the phase detector of the clock alignment circuit. For
these cases, margin within the transmitting clock cycle should be consid­
ered and, if the clock period is very short, frequency division of the clock
should also be considered.
Now that we have established a method to synchronize a signal between
the C/A CLK and RdCLK domains and also accounted for the absolute phase
difiference between the two clock domains, we now require a method to
continuously track phase differences between the two clock domains given
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 363

variations in voltage and temperature. Figure 12.44 is a modification of Fig­

ure 12.43 showing how the phase difference can be referenced and phase
variations tracked between the C/A CLK and RdCLK domains. What is
shown here is a method whereby the counter referenced to the RdCLK
domain is started at the same time that the StCA signal is launched through
the I/O delay model. When the StCA signal emerges from the I/O delay
model, it is used to synchronously start the counter referenced to the C/A
CLK domain. The result, shown in the timing diagram of Figure 12.44, is
two free running counters with corresponding values separated by the phase
difference between the C/A CLK and RdCLK clock domains. Now, as long­
term changes in voltage and temperature occur, the corresponding counter
values will adjust accordingly and always maintain the correct phase differ­
ence because each counter is running independently in each of the clock
domains!

rklk rLrLrLrLrLjn rm rLTLnjTrLrLrL

Initialization ——i__________________________________________________
I
logic output____________________ 1
counter .
I
StCa 1 |

c/a clk LrLrLWLAj-L^j-LrLrLrLnLTLrLr

Figure 12.44 Counter-based phase tracking initialization.

Up to this point in the description of our CL tracking scheme, we have

established a continuous reference between the C/A CLK and RdCLK clocks
that reflects the phase difference between the clock domains, even consider­
ing changes in voltage and temperature. The next step is to account for the
programmed CL value and any synchronization points that are in the Read
command decode path or OE timing and logic path. Synchronization points
are flip-flops or retiming logic that is in the logic path affecting signals
associated with Read command processing or OE generation. An example
would be a flip-flop used to retime the Read command to the command
364 Chap. 12 Control Logic Design

clock following the command decoder. Keep in mind, of course, from the
previous discussion, that the Read command sent to the array is not retimed
to any clock; but the Read command used fbr CL tracking is synchronized
because CL tracking is, inherently, a synchronous operation.
If synchronization points are present in the CL tracking logic, and these
retiming circuits logically delay signals or logic used for CL timing, then
we must account for these cycles when tracking CL. Continuing with our
CL tracking methodology, Figure 12.45 is again a modification of earlier
diagrams to show a calculation that is used to determine a load value for the
C/A CLK clock domain counter. Because the synchronous initialization of
the counters accounts for the I/O delay induced phase difference between
the C/A CLK and RdCLK domains, by preloading one of the counters with
an offset based on the compensated programmed CL value, the delay
between corresponding values of the two counters will reflect the internal
CL timing necessary to ensure correct external CL timing. The internal CL
timing is given by:

= (CL-SP)tCK-(tin + tout)

(12.9)

where SP is the number of synchronization points in the CL tracking logic

path. There are two parts to Equation (12.9). The first part is the synchro­
nous portion that accounts fbr the programmed CL and the synchronization
points (SP) in the tracking logic. The second part accounts fbr the asynchro­
nous phase difference between the RdCLK and C/A CLK clock domains
caused by the clock alignment circuit.
Before continuing, the following discussion will summarize the condi­
tions shown in the timing diagram of Figure 12.45. The offset value that is
loaded into the C/A CLK domain counter prior to initialization is based on
the synchronous portion of Equation (12.9). Let us consider the hypothetical
situation where SP = 0 and (tin + 10t4=0. For this situation, the value of
tCLini would be equal to the external value of a &CL = CL • Zck) and corre­
sponding count values for the C/A CLK counter and RdCLK counter would
occur simultaneously. If we now assign a non-zero value to (//„ + tou^ and
leave SP = 0, then any count value out of the C/A CLK counter would corre­
spond to the same RdCLK counter value but be advanced by (f加 + tout)
before & expires. At this point, the relationship between the counters is
accounting for the input and output circuit path delays. Next, if we were to
start the C/A CLK counter with an offset value equal to CL-SP, while start­
ing the RdCLK counter at the value 0, we would observe that any count
value occurring at the output of the C/A CLK counter would occur tCLint
Sec. 12.5 Read Data Latency Timing and Data Multiplexing 365

later at the output of the RdCLK counter. With the counters configured in
this fashion, we are now internally tracking CL, which is defined relative to
the external command clock domain.

Rd* rLrLrLrLrLtTrLrLrLTLrLrmLrLrL
Initialization
logic output
RdCLK
counter
StCa
C/ACLK
counter

C/ACLK "LrLrLTLrLrLnL^urLrLrLrLrLrrrLr

Figure 12.45 Load value for C/A CLK clock domain counten

Now that we have established a method for tracking the internal CL

timing, we need a means to store count values following a Read command.
The method described so far requires careful consideration of the synchro­
nous relationship between when a Read command is captured at the inputs
of the DRAM and when transitions occur on the C/A CLK counter. For a
matched-mode command and address capture, where the clock and com­
mands are transported together to the capture latches, we must ensure that
the command decoder generates a signal, labeled “Read” in Figure 12.46,
which has a known correspondence to the clock edge that is aligned to the
Read command at the inputs of the device. This signal must be used to cap­
ture the transition out of the C/A CLK counter that corresponds to the cycle
on which the Read command is decoded.
If we refer to Figure 12.40 in Section 12.5.1, we see that in this example
five Read commands are received before the expiration of the internal CL
time. The implication of this is that if we are to capture the value of the C/A
CLK counter for consecutive Read commands, then we must store multiple
count values before the captured value of the first Read command is con­
sumed. Once again, we use a FIFO to solve this problem. Referring to Fig­
366 Chap. 12 Control Logic Design

ure 12.46, each time a Read command is decoded, the signal Read is
generated and used to load the C/A CLK counter value into the FIFO. Once
the RdCLK counter attains the current output value of the FIFO, a digital
comparator indicates a match and signals back to the FIFO that the value
has been consumed. At the same time, the comparator indicates to the OE
timing circuit that the transmit circuits must be enabled on the next cycle.
The OE timing circuit drives the OE signal to the data serializers in the out­
put circuit paths for the length of the Read burst. The sequence of events
required to enable the output data path is shown in Figure 12.47. Note that a
synchronous transfer of the OE signal from the digital comparator to the
serializer would cost one SP. Likewise, the synchronous transfer from the
OE timer circuit to the serializer would cost another SP, Gray-code counters
are used for the clock domain counters [22]. By using gray-code counters,
false compares are eliminated from the digital comparator and FIFO loads
are more reliable because only one bit transitions on each clock cycle. This
configuration is illustrated in Figure 12.46.

Figure 12.46 Functional block diagram of CL tracking circuit

Sec. 12.6 Comments on Future Direction fbr DRAM Logic Design 367

c，aclk T_TL?LrLrLrLrLn_nrLrLrLrLrLn_r
I I--- 1
Read ^J |______________________________________________________________
I
FIFO output DC 2..........

RddK run|rLrLrLn_^LrLn_rLrLrLTLn_rL
然霜 QXI"><Q<®XD@X£<IXIXIXD<1XIX^>
「Compare 1 心. 、
valid I I _

OE | |
CL=10，SP = 3,tJL=175%

fCL 祝= (cl—sp)*-(，+ L)
Figure 12.47 Timing diagram of output data path enable.

The preceding discussion of the Read latency tracking circuit is but one
example of how we might track Read latency. Another proposal [21] is very
similar but uses tokens in connected shift registers rather than counters to
track the difference between clock domains. This method has merit except
that there is higher clock loading and potentially longer output delay due to
a large number of mux terms if a wide range of latency values is required.
For high-performance DRAM, the absolute value of CL increases greatly
relative to SDRAM and early DDR DRAM and, because the device must
often be designed to cover a wide range of data rates, the number of CL val­
ues that are valid is greatly increased over less demanding DRAM perfor­
mance levels. Another advantage of the method described in this text is that
the C/A CLK and RdCLK are not required to be physically routed to the
same area of the die. The use of a FIFO allows the clock routing for the sep­
arate clock domains to remain in their respective areas of the die.

12.6 COMMENTS ON FUTURE DIRECTION FOR DRAM LOGIC

DESIGN
Processor performance continues to advance as more cores and higher bus
frequencies demand higher data rates from the DRAM. These demands
directly impact the DRAM interface protocol and, hence, the DRAM con­
trol logic. This chapter is for from a comprehensive overview of DRAM
control logic. There are many details not included in this discussion but
hopefully the reader will come away with a new appreciation of some of the
368 Chap. 12 Control Logic Design

difficulties and, perhaps, a few ideas for solutions to problems associated

with DRAM interface logic design.
One area not covered in this chapter, but becoming of greater impor­
tance in the area of DRAM interface logic design, is channel initialization
algorithms fbr the timing of data capture. At the time of this writing, several
proposals are under consideration for standardized channel initialization
algorithms. To date, memory controller designs have used the published
DRAM protocols fbr initializing the data channels. Whether this trend con­
tinues and channel initialization remains in the province of the memory con­
troller or becomes a large part of the DRAM logic design remains to be
seen.
There are other areas of consideration fbr high-performance DRAM
that will profoundly impact DRAM logic design. In some niche applica­
tions, packetized DRAM [16] has been employed to reduce pin count at the
expense of increased channel bandwidth requirements. In future DRAM,
the packetized protocol may become more prevalent and standardized. As a
result, DRAM logic designers will be faced with a host of new and interest­
ing problems. Differential signaling is another possible trend. With differen­
tial signaling, channel bandwidth can and must be greatly increased to
justify the cost of higher pin count. Unfortunately, the implications of these
possible trends fbr DRAM logic design are not presented here but are men­
tioned so that the reader might pursue these areas fbr forward-looking
research.

REFERENCES

[1] D. Harris, Skew Tolerant Circuit Design. San Francisco, CA: Morgan
Kaufman Publishers, 2001.
[2] John P. Uyemura, CMOS Logic Circuit Design. Norwell, MA: Kluwer
Academic Publishers, 1999.
[3] Y. Taur, “CMOS scaling and issues in sub-0.25um systems'* in Design of
High-performance Microprocessor Circuits ed. by A. Chandrakasan, W. J.
Bowhill, and F. Fox. Piscataway, NJ: IEEE Press, 2001.
[4] T. Sakurai and A. R. Newton, "Alpha-power law MOSFET model and its
application to CMOS inverter delay and other formulas,n IEEE Journal on
Solid-State Circuits, vol. 25, pp. 584-593, April 1990.
[5] K. Chen and C. Hu, ''Performance and scaling in deep submicron CMOS,”
IEEE Journal on Solid-State Circuits, vol. 33, pp. 1586-1589, Oct. 1998.
[6] R. K. Krishnamurthy, A. Alvandpour, S. Mathew, M. Anders, V. De, B.and
Borkar, <€High-perfbrmance, low-power, and leakage-tolerance challenges fbr
sub-70nm microprocessor circuits,in Proc. 28th European ESSCIRC 2002,
pp. 315-321.
References 369

[7] Y. Takao, H. Kudo, J. Mitani, Y. Kotani, S. Yamaguchi, K. Yoshie, K., and M.

Kawano, UA 0.11 |im CMOS technology with copper and very-low-k
interconnects for high-performance system-on-a chip cores,in IEDM
Technical Digest International2000, pp. 77-80.
[8] S. Wong, G. Lee, and D. Ma, “Modeling of interconnect capacitance, delay
and crosstalk in VLSI,“ IEEE Transactions on Semiconductor Manufacturing,
vol. 13, pp. 108-111, Feb. 2000.
[9] T. Sakurai, ^Closed-form expressions for interconnection delay, coupling and
crosstalk in VLSIs/' IEEE Transactions on Electron Devices, vol. 40, pp.
118-124, Jan. 1993.
[10] R. H. Krambeck, C. M. Lee, and H. S. Law, “High-speed compact circuits
with CMOS," IEEE Journal on Solid-State Circuits, vol. SC-17, pp. 614-619,
June 1982.
[11] P. Gronowski, “Issues in dynamic logic design“ in Design of High-
performance Microprocessor Circuits, ed. by A. Chandrakasan, W. J. Bowhill,
and F. Fox. Piscataway, NJ: IEEE Press, 2001.
[12] K. Bernstein, K. M. Carrig, C. M. Durham, P： R. Hanson, D. Hogenmiller,
E. J. Nowak, and N. J. Rohrer, High-speed CMOS Design Styles. Norwell,
MA: Kluwer Academic Publishers, 1999.
[13] D. V. Campenhout, T. Mudge, and K. A. Sakallah, ^Timing verification of
sequential domino circuits,IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems, vol. 18, pp. 645-658, May 1999.
[14] I. E. Sutherland, ^Micropipelines,'' Communications of the ACM, vol. 32, pp.
720-738, June 1989.
[15] S. B. Furber and P. Day, uFour-phase micropipeline latch control circuits/*
IEEE Transactions on Very Large Scale Integration, vol. 4, pp. 247-253. June
1996.
[16] J. H. Choi, Y.-S. Sohn, C.-K. Kim, W.-K. Park, J.-H. Lee, U. Kang, G-S.
Byun, I.-S. Park, B.-C. Kim, H.-S. Hwang, C.-H. Kim, and S.-L Cho, “A 5.0
Gbps/pin packet-based DRAM with low latency receiver and process
insensitive PLLJ in Symp. on VLSI Circuits Digest of Tech. Papers 2005,
2005, pp. 50-51.
[17] H. A. Collins and R. E. Nikel, 4tDDR-SDRAM, high-speed, source-
synchronous interfaces create design challenges,M EDN, Sept. 2,1999.
[18] 1Gb DDR3 SDRAM datasheet, Micron Technology, Inc.
[19] S. H. Lee and J. B. Johnson. 2006. Method and apparatus for timing domain
crossing. US Patent #7,084,680, August, 2006.
[20] J. B. Johnson, B. Keeth, F. Lin, and H. Zheng, *Thase intolerant latency
control for a combination 512Mb 2.0Gb/s/pin GDDR3 and 2.5Gb/s/pin
GDDR4 SDRAM,“ in Digest ISSCC 2007, 2007, pp. 49195.
[21] S.-B. Lee, S.-J. Jang, J.-S. Kwak, S.-J. Hwang, Y.-H. Jun, S.-I. Cho, and C.-G
Lee, “A 1.6-Gb/s/pin double data rate SDRAM with wave-pipelined CAS
370 Chap. 12 Control Logic Design

latency control/' IEEE Journal on Solid-State Circuits, vol. 40, pp. 223-232,
Jan. 2005.
[22] D. D. Gajski, Principles of Digital Design. Saddle River, N.J.: Prentice-Hall,
Inc., 1997.
[23] T. Zhang and S. Sapatnekar, “Simultaneous shield and buffer insertion fbr
crosstalk noise reduction in global routing,M IEEE Transactions on Very Large
Scale Integration, vol. 15, pp. 624-635, June 2007.
[24] C. Svenssen and J. Yuan, “High-speed CMOS circuit technique,IEEE
Journal on Solid-State Circuits, vol. 24, pp. 62-70, Feb. 1989.
[25] J. Cortadella, M. Kishinevsky, A. Kondratyev, L. Lavagno, and A. Yakovlev,
Logic Synthesis ofAsynchronous Controllers and Interfaces. Berlin: Springer­
Verlag, 2002.
[26] H. Partovi, “Clocked storage elements/' in Design of High-Petformance
Microprocessor Circuits ed. by A. Chandrakasan, W. J. Bowhill, and F. Fox.
Piscataway, NJ: IEEE Press, 2001.
 Chapter
13

Power Delivery
An interesting thing happens on the way to high performance. Tolerance fbr
supply noise and voltage gradients drops off a cliff This increased sensitiv­
ity to supply voltage arises from two different factors. The first is that the
timing margins for high-frequency designs are very thin, leaving less mar­
gin on the table fbr supply variation. The second is that the inevitable scal­
ing of supply voltages creates a need fbr better power delivery networks just
to maintain the status quo. After all, 100 mV of supply noise on a 1-volt
supply is a much greater problem than 100 mV of supply noise on a 2.5-volt
supply. In this chapter, we explore the power delivery problem in greater
detail and ways to mediate it.
As an obvious outcome of moving toward higher clock frequencies and
faster data rates, on-die timing gets squeezed. This is especially true in and
around high-speed I/O, pipeline, and control circuits. As timing budgets
shrink, the margins available for supply variation also shrink. This leads to
the inevitable conclusion that to reduce these variations circuit designs must
either tolerate more power supply variation or improve the power delivery
network. Typically the most timing-sensitive circuits are involved in I/O
operations, such as capturing input data and driving output data. Data cap­
ture circuits often involve clock receivers, DLL or PLL blocks, clock distri­
bution networks, and input buffers and capture latches. The actual path from
the clock and data pins to the capture latches can be very long-on the order
of multiple clock cycles in electrical length. A long electrical path, of
course, leads to a greater sensitivity to power supply noise and variation.
The delay sensitivity of typical CMOS logic to supply voltage ranges
between two to ten times greater than that of fully differential circuits, such
as current mode logic (CML) [1]. As a result, highly sensitive timing cir­
cuits should use as much differential logic as possible. Furthermore, CML

371
372 Chap. 13 Power Delivery

logic is significantly faster due to its current-steering mode of operation,

and also from smaller signal swings. Propagation delays for CML logic are
shown to be 60% faster than those of standard CMOS logic. However, the
static power of CML circuits is higher than that of CMOS logic because
they gobble current regardless of whether they are switching or not [2].
Dynamic power remains comparable between the two logic styles.
So, given the fact that we use supply-tolerant logic such as CML
throughout our high-speed critical timing circuits, do we still need to worry
about power delivery? Unfortunately, yes, power delivery remains critical
for high-performance designs. This is true for critical timing circuits, but
also for power-hungry circuits in which the on-die supply network must be
designed to meet both peak and sustained current demands. A good exam­
ple of the latter is the helper flip flop (HFF) block in a GDDR3 memory
device. Supply current to the HFF block is high because it services a wide
data path over very short (2.5ns) column cycle times. We fully investigate
this example later in the chapter.
Historically, power delivery networks (PDN) have not received the
level of attention that they deserve. Given the performance demands of
modem high-performance parts, a lax attitude towards PDN design is no
longer an option. As shown in Figure 13.1, supply voltages are on a perpet­
ual downward trend. While reductions in supply voltage help reduce power
dissipation, they also place significantly greater demands on PDN designs,
as previously discussed. High-performance designs can generally tolerate
only 3-5% of AC noise on the supplies. This translates to 75-125 mV for a
2.5-volt supply, while we're looking at only 36-60 mV for a 1.2-volt sup­
ply. Ultimately, the PDN needs ongoing improvements at a rate that tracks
reductions in the supply voltage. This is necessary merely to maintain the
status quo as it will take even greater efforts to get ahead of the curve.

Figure 13.1 DRAM supply voltage trend.

Sec. 13.1 Power Delivery Network Design 373

13.1 POWER DELIVERY NETWORK DESIGN

There are a variety of methods available to ensure that the power delivery
network (PDN) is capable of meeting design requirements. The methods all
involve some type of PDN model and either static or dynamic simulations.
One popular method, widely supported by modem IC design tools, is static
supply analysis. This requires a PDN model that is either built up by hand or
extracted from the actual layout database using specialized software. This
type of analysis uses static current loads connected throughout the PDN
model to evaluate the capabilities of the design. Each current load model
idealizes the load currents of circuits that operate in the vicinity of each con­
nection point. This method is a reasonable approach to solving PDN prob­
lems while producing relatively quick simulation turn times.
Another approach to PDN design involves using full dynamic simula­
tions. Again, this requires a PDN model but with actual circuits in lieu of
current sources to stimulate the model. Using full dynamic simulations pro­
vides a notable advantage over static analysis because it supports PDN anal­
ysis with a variety of vectors, and it also exposes the effects of PDN noise
on circuit performance and timing. Thus, it shows both cause and effect.
Furthermore, if the PDN model incorporates a full package netlist, then this
flow both enables and supports device and package co-design. Co-design is
a process in which the device and package are designed and tuned in concert
with one another to maximize their combined performance. This is a power­
ful and evolving piece of high-perfbrmance design flows. Co-design is an
absolute necessity in the multi-gigahertz realm.
Lefs take a look at an example of PDN design using the now-familiar
GDDR3 device. Our exemplary GDDR3 device has three levels of on-die
metallization plus one redistribution layer (RDL) that permits connection to
a four-layer flip chip in package (FCIP) substrate. That's a total of eight
metal layers for this high-performance DRAM. It may seem a bit excessive,
but the results are truly impressive.
Figure 13.2 shows both the power bussing and signal routes of the
meta!3 layer from the GDDR3 design. There is actually dense metaI3 over
the white areas in the array. For clarity, this is not shown in the image.
Essentially, meta!3 is the upper-level metal used extensively for on-die
power distribution. It works in concert with the meta!2 layer, that runs
orthogonal to meta!3 to deliver power throughout the entire die. The pri­
mary bond pads of this GDDR3 device are separated into two rows, as
shown in Figure 13.2. The power pads within these pad rows tie directly
into the metal3/metal2 power bus network.
374 Chap. 13 Power Delivery

Figure 13.2 GDDR3 meta13 layout.

Obviously, what is shown in Figure 13.2 represents an actual layout

from the exemplary GDDR3 device. During the early and intermediate
phases of a DRAM design, this level of layout does not actually exist. Ide­
ally, layout of the die would not even start without a preliminary floor plan
and PDN plan in place. While it is possible to scribble out a floor plan and
PDN plan on the back of a napkin, it is more prudent and reliable to build
and tune these representations together using simulation models. The dia­
gram shown in Figure 13.3 is a schematic diagram of an actual PDN model
used fbr tuning the GDDR3 device. This model is built using parameterized
wire elements that represent metal bus segments in a proposed PDN. The
elements are parameterized to permit easy changes to bus widths and
lengths as required by the tuning exercise. The model also includes decou­
pling capacitors scattered throughout its structure. These capacitors are
important to PDN design and analysis because they serve as local sources of
charge for all of the circuit blocks.
 Figure 13.3 GDDR3 on —die P D N modeL
一 contained in the A l r c h i p simulation
B a s i c a l l y y e v e r y circui 一 muse

n
e
s
s
be m a p p e d the P D N model. This e H m i n a s s -he ideal p o w e r a n d
o
a
o

g r o u n d c o n n e c - i o n s that w e g e n e r a H y d e a l w i t h in s i m u F - i o n s p a c p r e 一
p ac
，
i n g一 h e m w i _ h c o n n e c t i o n s o the n o n i d c a H z e d p o w e r g r i d ，E l i m i n a t i o n o f
-

—
ideal p o w e r c o n n e x i o n s s i g n s c a 二 n y increases simulafion H m e b e c a u s e
e v e n small a m o u n t s o f curren- m o v i n g a r o u n d h e n e t w o r k can impact large

-
s e g m e n t s o f一 h e vo-fage g Localized disHrbances o f the supply grid
r
F
i m p a c - s一 h e circ 1 in a variety o f ways- N o ab 二 y ransis 一 or a n d
E,
s
i
m
a
a
i
o
n

-
-
ga_e behavior d e p e n d o n o c a l conditions- A s a r e s u F s o m e circuits speed
up w h i o 一 o hers d o w p Signaling e v e F within a circuit and
s
o
w

b
s
h

b
o
c
k
b e t w e e n b o c k s 》are affec-ed b y p o w e r supp-y noise a n d gradients ，C h a n g e s
in一 h e c o m m o n m o d e v o H a g e f r o m o n e a r e a o f the die to a n o - h e r i m p a c t
b o f h s i g n一
a and dis bution ， n essence 》y o u r s i m一 u a H o n w一 i一 exhibit

5.
c
o
c
k
m o r e rea丁 w o r l d behavior than ever b e f b r p A s a y o u w三 b e to

E-
r
e
s
L

a
z
c
observe p o w e r s u p p y v o y a g e s throughoutthe P D N u n d e r a variety o f o p ?

-
a i n g c o n d i t i o n s a n d simii 二a o n vectors- m o a i m p o r t a n t ^ y o u c a n see

b
&
-
 ・♦・・»••3••・・•••・,・»««»««•<•・•♦・B

驷
I
aiiiuniiiimuiuinfliiNniii

卦
i_
?l:
r:;
热

i:;.
j
r-:
sl»

ssn
£!t
z?:
su

a
slla
A
5:3;
g:.
8?!&
2:
i:=
;:
376 Chap. 13 Power Delivery

how the resulting supply noise impacts circuit operation-a tremendous

one-two punch in observeability.

13.2 DEVICE/PACKAGE CO-DESIGN

With the addition of a package model to the PDN simulation just described,
it becomes possible to perform device/package co-design. As stated earlier,
the package used for the exemplary GDDR3 device is designed with FCIP
technology. Figure 13.4 illustrates what an FCIP package looks like in cross
section. Figure 13.4a is a conceptual drawing, while Figure 13.4b-d are
cross-sectional photographs of an actual GDDR3 package at various magni­
fications. Essentially the die is mounted with its circuit side facing the pack­
age substrate. An RDL layer is commonly used to move die connection
points to locations that are compatible with the underlying substrate tech­
nology. The substrate itself is typically constructed of conventional PCB
materials, although at a very fine pitch. The number of conductor layers
within the substrate is based on a balance between design needs and cost.
All four conductor layers are clearly visible in the GDDR3 package shown
in Figure 13.4b-d.
The package design used on a first-generation GDDR3 device is
depicted through its various metal layers in Figure 13.5. All five layers are
shown in Figure 13.5a-e, starting with the RDL layer and proceeding down
through the four substrate layers from top to bottom. Figure 13.5f is an
exploded view showing all five package layers in their proper stack-up
order. This package is the product of a chip and package co-design flow.
The flow optimized each element in the overall power delivery network,
including the package and on-die PDN, for maximum device performance.
Sec. 13.2 Device/Package Co-design 377

c)

d)

Figure 13.4 a) GDDR3 FC1P conceptual package diagram, b) GDDR3 FCIP

package cross section photograph, c) GDDR3 FCIP package cross section detail,
and d) GDDR3 FCIP package cross section, magnified detail
378 Chap. 13 Power Delivery

a)

b)

O O O O IO o 。o o
C)

Figure 13.5 a) GDDR3 RDL layer, b) GDDR3 top substrate layer,

and c) GDDR3 second substrate layer.
Sec. 13.2 Device/Package Co-design 379

d)

e)

f)
Figure 13.5 d) GDDR3 third substrate layer, e) GDDR3 bottom substrate layer,
and f) GDDR3 package layer stack.
380 Chap. 13 Power Delivery

13.3 FULL-CHIP SIMULATIONS

Let's take a look at how full-chip simulations using PDN and package mod­
els can be used to improve power delivery and device performance using a
GDDR3 example. A full-chip simulation fbr the GDDR3 device was con­
structed using a device model, a PDN model, a package model, and a bus
model, as shown conceptually in Figure 136 The bus model essentially
connects the package model external ball terminals to power supplies and
bus wire models to ensure that the input signals entering the package model
meet specification and also to ensure that output signals are properly
loaded. All stimulus and supply voltages reach the device only by traversing
through the bus, package, and PDN model.

Figure 13.6 Conceptual simulation diagram.

To identify problem areas, a variety of full-chip simulations were run on

the design in which both timing and power supply voltages were analyzed.
One particular area of concern that emerged from the simulations was the
Read clock trees. The clock trees provide critical timing information to the
output data path and contribute significantly to several measurable output
timing specifications. Initial simulations showed that the clock tree power
busses produced approximately 100 mV of noise and sag whenever the
clock trees were enabled. This result, labeled as "Before,“ can be seen in the
simulation waveforms shown in Figure 13.7. Using the package/PDN/
device simulation, we analyzed a variety of design changes to correct this
problem. We ultimately chose to implement a design change that added
Sec. 13.3 Full-Chip Simulations 381

three pairs of Vdd and 飞 bond pads within each of the Read clock trees.
These new bond pads were readily supported by the FCIP packaging tech­
nology of this design. Simulation of the revised design yielded the improve­
ment shown in Figure 13.7 and labeled, “After." As shown, the power
supply noise and sag were reduced to only 50 mV, which met our internal
design goal for this voltage supply.

Figure 13.7 Vdd simulation waveforms for Read clock trees.

Figure 13.8 once again shows the metal3 layout for the GDDR3 device.
In this view, however, the 3/G power pads, which were added to sup­
port the Read clock trees, are circled to show their relative locations.

Figure 13.8 GDDR3 metaB layout with clock tree power pads.
382 Chap. 13 Power Delivery

Another example from recent work on the GDDR3 design is depicted in

Figure 13.9 and Figure 13.10. In these befbre-and-after images, it is easy to
see the gains achieved through the PDN/package simulation and co-design
efforts fbr the HFF block. The waveforms show the Vdd and Vss nets in the
HFF block fbr a back-to-back Write and back-to-back Read vector. This
type of pattern places a tremendous amount of stress on column cycle time
and HFF block power delivery. In Figure 13.9, the supply busses are clearly
inadequate to provide stable voltage levels to the HFF circuits. They are vir­
tually collapsing. In Figure 13.10, the HFF supply busses show a significant
improvement after much-needed PDN tuning. The DC voltage levels
remain stable, and the amount of AC noise lies within design targets. With­
out PDN/package simulations, the problem illustrated in Figure 13.9 would
not have been discovered and corrected prior to tapeout. A very costly and
perhaps catastrophic problem was averted.
Owing to the demands of high-performance specifications and shrink­
ing supply voltages, power delivery is fast becoming the bane of semicon­
ductor designers. This is especially true for leading-edge projects where the
timing and supply-induced problems grow significantly more complex with
each passing generation. The reality of ever-tighter design margins and
shrinking supply voltages must foment solutions that move device simula­
tions closer and closer to actual silicon behavior. Power delivery problems
must be dealt with early in the design flow and addressed through an
aggressive combination of simulation and co-design. Nothing, absolutely
nothing can be overlooked in this regard, since everything matters.

Figure 13.9 Vdd and P5j waveforms fbr HFF block before tuning.
Sec. 13.3 Full-Chip Simulations 383

10ns/dMsion

w --------出MM2*温—

Figure 13.10 %)and Vgg waveforms for HFF block after tuning.

REFERENCES

[1] J. Chapman, J. Currin, and S. Payne, “A low-cost high-performance CMOS

timing vernier for ATE,“ in Proceedings International Test Conference J995,
October 1995, pp. 459-468.
[2] I. Li, S. Raghavendran, and D. Comer, “CMOS current mode logic gates for
high-speed applications/ in 12th NASA Symposium on VLSI Design, October
4-5,2005.
 Chapter
14

Future Work
in High-Performance Memory
The semiconductor industry will continue its inevitable march for­
ward-deriving ever greater benefits through additional scaling and higher
levels of integration. Barring the unforeseen, device density, power, and
speed will continue to follow their historical trends.
We gain insight into what the future may hold fbr DRAM by looking at
the past and at various trends. For instance, Figure 14.1 is a graph showing
a trend line fbr DRAM data rate per pin by year. This data shows steady
gains from 1995 through 2010 with the data rate doubling approximately
every three years. While any data beyond present day is somewhat specula­
tive, it is based on recent trending and extrapolated historical data. Figure
14.1 shows that device performance, as it relates to I/O bandwidth, will
continue to advance beyond the end of the decade. This is both exciting and
foreboding fbr designers. It is exciting because engineers relish a challeng­
ing design project, yet foreboding because the trend presented in Figure
14.1 represents a never-ending series of design problems that become
increasingly more difficult.
While higher bandwidths are a perceived benefit, unfortunately this
benefit comes at a cost. The power dissipated by these high-strung designs
also appears to rise with each new generation. Figure 14.2 shows a trend
line of power dissipation fbr a variety of recent DRAM interface technolo­
gies. The data shown is based on the IDD4 specification fbr each technol­
ogy. IDD4 measures supply current for back-to-back page Reads and/or
back-to-back page Writes at minimum column cycle time. This specifica­
tion examines power at the maximum sustainable device bandwidth.
Despite the fact that supply voltages drop with each new technology, power

385
386 Chap. 14 Future Work in High-Perfbrmance Memory

is nevertheless going up. The fact that the trend line rises is not surprising
since it is tightly correlated to I/O bandwidth, internal clock frequencies,
and the data pre-fetch size-all of which are on the increase.

1995 2000 2005 2010

Year

Figure 14.1 Trend line fbr DRAM data rate per pin.

900.0E-03

8OO.OE-O3

700.03-03
」
8M0d

600.0E-03

500.0E-03

400.0E-03
256 Mbit SDRAM 512 Mbit DDR 1 Gbit DDR2 1Gbit DDR3
167MHz 200MHz 333 MHz 333MHz
Technology

Figure 14.2 1DD4 power dissipation trend line versus technology.

Most every computer system in use today and those being designed for
future use are highly power constrained. This is at odds with the device
power dissipation depicted in Figure 14.2. As a result, one should expect
that future technologies will adopt more sophisticated power management
schemes to keep device and system power within budget. These schemes
Chap. 14 Future Work in High-Performance Memory 387

will strive to reduce both active and standby power to absolute minimum
levels through a variety of mechanisms. The most obvious being to shut
down circuits not needed for ongoing operations. The degree to which cir­
cuits can be shut down depends on our tolerance fbr the delays we encoun­
ter when turning these circuits back on. Nothing comes for free.
Another issue pertaining to power is leakage current. All transistors
leak when they are turned off-some more, some less. Historically, DRAM
transistors are designed to have low leakage, especially the array Mbit tran­
sistors, since leakage impacts data retention and Refresh. However, the
quest for higher perfbnnance and high-bandwidth I/O is creating pressure to
build better transistors. Better, in this case, meaning higher drive and lower
Unfortunately, what leads to higher performance also leads to higher
leakage current. So, in order to keep the array leakage current low while
increasing peripheral transistor performance, DRAM manufacturers will be
forced to add process complexity. This will translate to more masks, more
process steps, and higher cost, which is just the opposite of what is needed
in a highly competitive commodity market. Nevertheless, if the perfor­
mance goals are to be realized, then added process complexity is the price to
be paid.
Without a doubt, the most notable and predictable feature of DRAM
technology is memory density and its associated growth over the years.
Advances in memory density are fueled by a perpetual scaling of the under­
lying CMOS process technology. Figure 14.3 shows a ten-year timeline of
historical DRAM shipments by density. Although the cadence in recent
years appears to have tapered off a bit, memory density has essentially dou­
bled every two to three years. Accordingly, Figure 14.4 shows a forecast of
shipments by density through the end of 2010. As expected, the lower den­
sities continue to taper off while the higher densities, starting at 512MB,
grow in share until they peak and tail off. There's nothing surprising about
this data since it merely follows the historical trend.
Increased memory density does create some problems for high-perfor­
mance DRAM devices. There are obvious architectural ramifications, as
discussed earlier in Chapter 8. Large blocks of memory can lead to long
wire routes, cycle time problems, and longer latency. Not exactly what you
expect in a high-performance part. Typical first-generation parts of a spe­
cific memory density will not achieve the same level of performance as
their lower density counterparts. It usually takes one or two process shrinks
before performance levels rise to targeted levels.
388 Chap. 14 Future Work in High-Performance Memory

00 00 8s s

寸 0
OS QO 0® 0s 00
Z6

86

6 。6
9 。6

96

coo

90
96

96

Z6

86

66

8
ounr

a) Q Q 0 O©
Q
C
。

Figure 14.3 Historical DRAM shipments by density.

4Gb
100%

80%

60%

40%

20%

§ 6 0
卜

0 0
60DZ

0
CODE
60g

-
800 L
90ON

90g

03Z
058
goq
90g
ZOOL
ZGQZ
zoom

9 0 O g
L

Calendar quarter

Figure 14.4 DRAM shipments by density forecast.

So, what will DRAM devices look like in the future? Given the histori­
cal trends, they should continue to grow in density and speed, bum more
power, and operate at lower and lower supply voltages. But will that neces­
sarily be the case? It's reasonable to expect these trends to hold for the fore­
seeable future, but they are not sustainable indefinitely. Recent
transformations in the microprocessor world should serve to illustrate this
point For years, the personal computer world looked like a horse race
Chap. 14 Future Work in High-Performance Memory 389

involving system benchmark performance and clock frequencies. This con­

tinued year after year until processor frequencies began to reach beyond
3GHz. At this point, the design challenges associated with such high-fre­
quency operation coupled with dizzying power dissipation problems forced
a change in direction. Basically, the major players started to walk away
from single core processors toward multi-core designs. This resulted in
seemingly higher overall performance at more reasonable power levels-a
winning strategy for everyone involved. While the benchmark wars con­
tinue unabated, focus has shifted from maximizing clock frequency to
increasing the number of processor cores available. The race is on, but in a
new direction.
Given the course change in the microprocessor world, it seems reason­
able to expect similar changes in the DRAM world. After all, I/O frequen­
cies cannot rise indefinitely, especially when cost and power remain critical.
I am unable to predict what those changes will be or how they will impact
DRAM design, but they must produce improved system performance and
lower overall power dissipation to succeed.
Research is underway on the next generation of high-performance
DRAM. In fact, this type of research never ends. However, the convergence
of lower supply voltages, tighter power constraints, and ever-higher clock
frequencies are indeed creating a difficult set of problems to overcome. In
addition, process and device variability is rising to prominence as a fore­
boding new challenge for the DRAM world. Variability becomes more
prevalent with each new process node. This phenomenon is the by-product
of transistor scaling. The most common sources of variability are line edge
roughness and nonunifbrm dopant distribution. Design engineering teams
must come to grips with variability and learn to design circuits and architec­
tures with demonstrably lower sensitivity. This will not be an easy task. It
hinges, in part, upon an ability to model variability and simulate its effects.
The future for DRAM technology looks both exciting and challenging.
A combination of new architectures, circuits, and process technologies pro­
vide the needed capabilities to forge ahead. Tomorrow's high-performance
devices will no doubt make today's best devices seem quaint by compari­
son. Each new generation presents an opportunity to advance the state of the
art. Even while we strive to use our best ideas on today's design project, we
nevertheless discover a better idea for use on the next. This is why DRAM
design is so exciting. We are forever adding new knowledge to old. We raise
device performance and customer expectations with each new project-all
while standing on the shoulders of giants.
 Appendix

Supplemental Reading
In this second edition of DRAM Circuit Design, we introduce the reader to a
variety of topics—from introductory to advanced. However, we may not
have covered specific topics to the reader's satisfaction. For this reason, we
have compiled a list of supplemental readings from major conferences,
journals, and books categorized by subject. It is our hope that unanswered
questions are addressed by the authors of these readings, who are experts in
the field of CMOS design.

General DRAM Design and Operation

[1] S. Fuji, K. Natori, T. Furuyama, S. Saito, H. Toda, T. Tanaka, and O. Ozawa,
UA low-power sub 100 ns 256K bit dynamic RAMJ IEEE Journal ofSolid­
State Circuits, vol. 18, pp. 441-446, October 1983.
[2] A. Mohsen, R. I. Kung, C. J. Simonsen, J. Schutz, P. D. Madland,
E. Z. Hamdy, and M. T. Bohr, "The design and performance of CMOS
256K bit DRAM devices,IEEE Journal of Solid-State Circuits, vol. 19,
pp. 610-618, October 1984.
[3] M. Aoki, Y Nakagome, M. Horiguchi, H. Tanaka, S. Ikenaga, J. Etoh,
Y. Kawamoto, S. Kimura, E. Takeda, H. Sunami, and K. Itoh, “A 60-ns 16-
Mbit CMOS DRAM with a transposed data-line structure,IEEE Journal of
Solid-State Circuits, vol. 23, pp. 1113-1119, October 1988.
[4] M. Inoue, T. Yamada, H. Kotani, H. Yamauchi, A. Fujiwara, J. Matsushima,
H. Akamatsu, M. Fukumoto, M. Kubota, I. Nakao, N. Aoi, G. Fuse,
S. Ogawa, S. Odanaka, A. Ueno, and H. Yamamoto, 16-Mbit DRAM
with a relaxed sense-amplifier-pitch open-bit-line architecture,IEEE
Journal ofSolid-State Circuits, vol. 23, pp. 1104-1112, October 1988.

391
392 Appendix

[5] T. Watanabe, G. Kitsukawa, Y. Kawajiri, K. Itoh, R. Hori, Y. Ouchi,

T. Kawahara, and R. Matsumoto, ^Comparison of CMOS and BiCMOS 1-
Mbit DRAM performance,v IEEE Journal of Solid-State Circuits^ vol. 24,
pp. 771-778, June 1989.
[6] K. Itoh, ^Trends in megabit DRAM circuit design/' IEEE Journal of Solid­
State Circuits, vol. 25, pp. 778-789, June 1990.
[7] Y. Nakagome, H. Tanaka, K. Takeuchi, E. Kume, Y. Watanabe, T. Kaga,
Y. Kawamoto, F. Murai, R. Izawa, D. Hisamoto, T. Kisu, T. Nishida,
E. Takeda, and K. Itoh, “An experimental 1.5-V 64-Mb DRAM,“ IEEE
Journal of Solid-State Circuits, vol. 26, pp. 465-472, April 1991.
[8] R Gillingham, R. C. Foss, V. Lines, G. Shimokura, and T. Wojcicki, "High­
Speed, High-reliability circuit design fbr megabit DRAM,“ IEEE Journal of
Solid-State Circuits, vol. 26, pp. 1171-1175, August 1991.
[9] K. Kimura, T. Sakata, K. Itoh, T. Kaga, T. Nishida, and Y. Kawamoto,
block-oriented RAM with half-sized DRAM cell and quasi-fblded data-line
architecture,M IEEE Journal of Solid-State Circuits, vol. 26, pp. 1511-1518,
November 1991.
[10] Y. Oowaki, K. Tsuchida, Y. Watanabe, D. Takashima, M. Ohta, H. Nakano,
S. Watanabe, A.Nitayama, F. Horiguchi, K. Ohuchi, and F. Masuoka, “A 33­
ns 64-Mb DRAM," IEEE Journal of Solid-State Circuits, vol. 26,
pp. 1498-1505, November 1991.
[11] T. Kirihata, S. H. Dhong, K. Kitamura, T. Sunaga, Y. Katayama,
R. E. Scheuerlein, A. Satoh, Y. Sakaue, K. Tobimatus, K. Hosokawa,
T. Saitoh, T. Yoshikawa, H. Hashimoto, and M. Kazusawa, UA 14-ns 4-Mb
CMOS DRAM w汕 300-mW active power/' IEEE Journal of Solid-State
Circuitsy vol. 27, pp. 1222-1228, September 1992.
[12] K. Shimohigashi and K. Seki, uLow-voltage ULS1 design,IEEE Journal of
Solid-State Circuits, vol. 28, pp. 408-413, April 1993.
[13] G. Kitsukawa, M. Horiguchi, Y. Kawajiri, T. Kawahara, T. Akiba,
Y. Kawase, T. Tachibana, T. Sakai, M. Aoki, S. Shukuri, K. Sagara,
R. Nagai, Y. Ohji, N. Hasegawa, N. Yokoyama, T. Kisu, H. Yamashita,
T. Kure, and T. Nishida, Ct256-Mb DRAM circuit technologies for file
applications,IEEE Journal of Solid-State Circuits, vol. 28, pp. 1105-1113,
November 1993.
[14] T. Kawahara, Y. Kawajiri, M. Horiguchi, T. Akiba, G. Kitsukawa, T. Kure,
and M. Aoki, “A charge recycle refresh for Gb-Scale DRAMs in file
applications/' IEEE Journal of Solid-State Circuits, vol. 29, pp. 715-722,
June 1994.
[15] S. Shiratake, D. Takashima, T. Hasegawa, H. Nakano, Y. Oowaki,
S. Watanabe, K. Ohuchi, and F. Masuoka, 4tA staggered NAND DRAM array
Appendix 393

architecture for a Gbit scale integration,in 1994 Symposium on VLSI

Circuits, June 1994, p. 75.
[16] T. Ooishi, K. Hamade, M. Asakura, K. Yasuda, H. Hidaka, H. Miyamoto,
and H. Ozaki, “An Automatic Temperature compensation of internal sense
ground for sub-quarter micron DRAMs/* in 1994 Symposium on VLSI
Circuits^ June 1994, p. 77.
[1 刀 A. Fujiwara, H. Kikukawa, K. Matsuyama, M. Agata, S. Iwanari,
M. Fukumoto, T. Yamada, S. Okada, and T. Fujita, “A 200MHz 16Mbit
synchronous DRAM with block access mode,“ in 1994 Symposium on VLSI
Circuits, June 1994, p. 79.
[18] Y. Kodama, M. Yanagisawa, K. Shigenobu, T. Suzuki, H. Mochizuki, and
T. Ema, “A 150-MHz 4-bank 64M-bit SDRAM with address incrementing
pipeline scheme/' in 1994 Symposium on VLSI Circuits, June 1994, p. 81.
[19] D. Choi, Y. Kim, G. Cha, J. Lee, S. Lee, K. Kim, E. Haq, D. Jun, K. Lee,
S. Cho, J. Park, and H. Lim, “Battery operated 16M DRAM with post
package programmable and variable self refresh,n in 1994 Symposium on
VLSI Circuits^ June 1994, p. 83.
[20] S. Yoo, J. Han, E. Haq, S. Yoon, S. Jeong, B. Kim, J. Lee, T. Jang, H. Kim,
C. Park, D. Seo, C. Choi, S. Cho, and C. Hwang, WA 256M DRAM with
simplified register control for low power self refresh and rapid burn-in,*' in
1994 Symposium on VLSI Circuits, June 1994, p. 85.
[21] M. Tsukude, M. Hirose, S. Tomishima, T. Tsuruda, T. Yamagata,
K. Arimoto, and K. Fiyishima, ^Automatic voltage-swing reduction (AVR)
scheme fbr ultra low power DRAMs J in 1994 Symposium on VLSI Circuits,
June 1994, p. 87.
[22] D. Stark, H. Watanabe, and T. Furuyama, <6An experimental cascade cell
dynamic memory,in 1994 Symposium on VLSI Circuits, June 1994, p. 89.
[23] T. Inaba, D. Takashima, Y. Oowaki, T. Ozaki, S. Watanabe, and K. Ohuchi,
“A 250mV bit-line swing scheme for a IV 4Gb DRAM,“ in 1995
Symposium on VLSI Circuits9 June 1995, p. 99.
[24] I. Naritake, T. Sugibayashi, S. Utsugi, and T. Murotani, UA crossing charge
recycle refresh scheme with a separated driver sense-amplifier fbr Gb
DRAMs/5 in 1995 Symposium on VLSI Circuits, June 1995, p. 101.
[25] S. Kuge, T. Tsuruda, S. Tomishima, M. Tsukude, T. Yamagata, and
K. Arimoto, “SOLDRAM circuitt technologies for low power high speed
multi-giga scale memories,M in 1995 Symposium on VLSI Circuits^ June
1995, p. 103.
[26] Y. Watanabe, H. Wong, T. Kirihata, D. Kato, J. DeBrosse, T. Hara,
M, Yoshida, H. Mukai, K. Quader, T. Nagai, R Poechmueller, K. Pfefferl,
M. Wordeman, and S. Fiyii, (<A 286mm2 256Mb DRAM with X32 both-
ends DOJ in 1995 Symposium on VLSI Circuits, June 1995, p. 105.
394 Appendix

[27] T. Kirihata, Y. Watanabe, H. Wong, J. DeBrosse, M. Yoshida, D. Katoh,

S. Fujii, M. Wordeman, P. Poechmueller, S. Parke, and Y. Asao, ^Fault-
tolerant designs for 256 Mb DRAM J in 1995 Symposium on VLSI Circuits,
June 1995, p. 107.
[28] D. Takashima, Y. Oowaki, S. Watanabe, and K. Ohuchi, “A novel power-ofF
mode for a battery-backup DRAM,“ in 1995 Symposium on VLSI Circuits,
June 1995, p. 109.
[29] T. Ooishi, Y. Komiya, K. Hamada, M. Asakura, K. Yasuda, K. Furutani,
T. Kato, H. Hidaka, and H. Ozaki, UA mixed-mode voltage-down converter
with impedance adjustment circuitry for low-voltage wide-frequency
DRAMsJ in 1995 Symposium on VLSI Circuits^ June 1995, p. 111.
[30] S.-J. Lee, K.-W. Park, C.-H. Chung, J.-S. Son, K.-H. Park, S.-H. Shin,
S. T. Kim, J.-D. Han, H.-J. Yoo, W.-S. Min, and K.-H. Oh, “A low noise
32bit-wide 256M synchronous DRAM with column・decoded I/O line,^^ in
1995 Symposium on VLSI Circuits^ June 1995, p. 113.
[31] T. Sugibayashi, I. Naritake, S. Utsugi, K. Shibahara, R. Oikawa, H. Mori,
S. Iwao, T. Murotani, K. Koyama, S. Fukuzawa, T. Itani, K. Kasama,
T. Okuda, S. Ohya, and M. Ogawa, “A 1-Gb DRAM for file applications,
IEEE Journal of Solid-State Circuits, vol. 30, pp. 1277-1280, November
1995.
[32] T. Yamagata, S. Tomishima, M. Tsukude, T. Tsuruda, Y. Hashizume, and
K. Arimoto, “Low voltage circuit design techniques for battery-operated
and/or giga-scale DRAMs/' IEEE Journal of Solid-State Circuits, vol. 30,
pp. 1183-1188, November 1995.
[33] H. Nakano, D. Takashima, K. Tsuchida, S. Shiratake, T. Inaba, M. Ohta,
Y. Oowaki, S. Watanabe, K. Ohuchi, and J. Matsunaga, UA dual layer bitline
DRAM array with Vcc^ss hybrid precharge fbr multi-gigabit DRAMs/* in
1996 Symposium on VLSI Circuits, June 1996, p. 190.
[34] J. Han, J. Lee, S. Yoon, S. Jeong, C. Park, I. Cho, S. Lee, and D. Seo, “Skew
minimization techniques fbr 256M-bit synchronous DRAM and beyond,” in
1996 Symposium on VLSI Circuits, June 1996, p. 192.
[35] H. Wong, T. Kirihata, J. DeBrosse, Y. Watanabe, T. Hara, M. Yoshida,
M. Wordeman, S. Fujii, Y. Asao, and B. Krsnik, “Flexible test mode design
for DRAM characterization,in 1996 Symposium on VLSI Circuits, June
1996, p. 194.
[36] D. Takashima, Y. Oowaki, S. Watanabe, K. Ohuchi, and J. Matsunaga,
“Noise suppression scheme for giga-scale DRAM with hundreds of I/Os,in
1996 Symposium on VLSI Circuits, June 1996, p. 196.
[37] S. Tomishima, F. Morishita, M. Tsukude, T. Yamagata, and K. Arimoto,
“A long data retention SOI-DRAM with the body refresh function,in 1996
Symposium on VLSI Circuits, June 1996, p. 198.
Appendix 395

[38] M. Nakamura, T. Takahashi, T. Akiba, G. Kitsukawa, M. Morino,

T. Sekiguchi, I. Asano, K. Komatsuzaki, Y. Tadaki, Songsu Cho,
K. Kajigaya, T. Tachibana, and K. Sato, “A 29-ns 64-Mb DRAM with
hierarchical array architecture,IEEE Journal of Solid-State Circuits,
vol. 31, pp. 1302-1307, September 1996.
[39] K. Itoh, Y. Nakagome, S. Kimura, and T. Watanabe, ''Limitations and
challenges of multigigabit DRAM chip design,“ IEEE Journal ofSolid-State
Circuits, vol. 32, pp. 624-634, May 1997.
[40] Y. Idei, K. Shimohigashi, M. Aoki, H. Noda, H. Iwai, K. Sato, and
T. Tachibana, "Dual-period self-refresh scheme for low-power DRAMs with
on-chip PROM mode register/5 IEEE Journal of Solid-State Circuits,
vol. 33, pp. 253-259, February 1998.
[41] K. Kim, C.-G Hwang, and J. G Lee, “DRAM technology perspective for
gigabit era," IEEE Transactions Electron Devices, vol. 45, pp. 598-608,
March 1998.
[42] H. Tanaka, M. Aoki, T. Sakata, S. Kimura, N. Sakashita, H. Hidaka,
T. Tachibana, and K. Kimura, t4A precise on-chip voltage generator for a
giga-scale DRAM with a negative word-line scheme,in 1998 Symposium
on VLSI Circuits. June 1998, p. 94.
[43] T. Fujino and K. Arimoto, uMulti-Gbit-scale partially frozen (PF) NAND
DRAM with SDRAM compatible interface/, in 1998 Symposium on VLSI
Circuits^ June 1998, p. 96.
[44] A. Yamazaki, T. Yamagata, M. Hatakenaka, A. Miyanishi, I. Hayashi,
S. Tomishima, A. Mangyo, Y. Yiikinari, T. Tatsumi, M. Matsumura,
K. Arimoto, and M. Yamada, UA 5.3Gb/s 32Mb Embedded SDRAM core
with slightly boosting scheme,in 1998 Symposium on VLSI Circuits, June
1998, p. 100.
[45] C. Kim, K. H. Kyung, W. P. Jeong, J. S. Kim, B. S. Moon, S. M. Yim,
J. W. Chai, J. H, Choi, C. K. Lee, K. H. Han, C. J, Park, H. Choi, and
S. I. Cho, "A 2.5V, 2.0GByte/s packet-based SDRAM with a l.OGbps/pin
interface,in 1998 Symposium on VLSI Circuits, June 1998, p. 104.
[46] M. Saito, J. Ogawa, H. Tamura, S. Wakayama, H. Araki, Tsz-Shing Cheung,
K. Gotoh, T. Aikawa, T. Suzuki, M. Taguchi, and T. Imamura, u500-Mb/s
nonprecharged data bus fbr high-speed DRAMs J IEEE Journal of Solid­
State Circuits, vol. 33, pp. 1720-1730, November 1998.

DRAM Cells
[47] C. G Sodini and T. I. Kamins, ""Enhanced capacitor fbr one-transistor
memory cell,,i IEEE Transactions Electron Devices, vol. ED-23,
pp. 1185-1187, October 1976.
396 Appendix

[48] J. E. Leiss, P. K. Chatteijee, and T. C. Holloway, “DRAM design using the

taper-isolated dynamic RAM cell J IEEE Journal ofSolid-State Circuits, vol.
17, pp. 337-344, April 1982.
[49] K. Yamaguchi, R. Nishimura, T. Hagiwara, and H. Sunami, uTwo-
dimensional numerical model of memory devices with a corrugated
capacitor cell structure/ IEEE Journal of Solid-State Circuits, vol. 20,
pp. 202-209, February 1985.
[50] N. C. Lu, P. E. Cottrell, W. J. Craig, S. Dash, D. L. Critchlow, R. L. Mohler,
B. J. Machesney, T. H. Ning, W. P. Noble, R. M. Parent, R. E. Scheuerlein,
E. J. Sprogis, and L. M. Terman, 4tA substrate-plate trench-capacitor (SPT)
memory cell for dynamic RAMs J IEEE Journal of Solid-State Circuitsy vol.
21, pp. 627-634, October 1986.
[51] Y. Nakagome, M. Aoki, S. Ikenaga, M. Horiguchi, S. Kimura, Y. Kawamoto,
and K. Itoh, “The impact of data-Iine interference noise on DRAM scaling,v
IEEE Journal ofSolid-State Circuitsy vol. 23, pp. 1120-1127, October 1988.
[52] K. W. Kwon. I. S. Park, D. H. Han, E. S. Kim, S. T. Ahn, and M. Y. Lee,
“Ta2()5 capacitors for 1 Gbit DRAM and beyond,'' in 1994IEDM Technical
Digest, pp. 835—838.
[53] D. Takashima, S. Watanabe, H. Nakano, Y. Oowaki, and K. Ohuchi,
uOpen/folded bit-line arrangement fbr ultra-high-density DRAMs/' IEEE
Journal of Solid-State Circuits, vol. 29, pp. 539-542, April 1994.
[54] Wonchan Kim, Joongsik Kih, Gyudong Kim, Sanghun Jung, and Gijung
Ahn, “An experimental high-density DRAM cell with a built-in gain stage,
IEEE Journal ofSolid-State Circuits, vol. 29, pp. 978-981, August 1994.
[55] Y. Kohyama, T. Ozaki, S. Yoshida, Y. Ishibashi, H. Nitta, S. Inoue,
K. Nakamura, T. Aoyama, K. Imai, and N. Hayasaka, t4A fully printable,
se序aligned and planarized stacked capacitor DRAM access cell technology
for 1 Gbit DRAM and beyond/' in IEEE Symposium on VLSI Technology
Digest of Technical Papers, June 1997, pp. 17-18.
[56] B. El-Kareh, G. B. Bronner, and S. E. Schuster, “The evolution of DRAM
cell technology, Solid State Technology, vol. 40, pp. 89-101, May 1997.
[57] S. Takehiro, S. Yamauchi, M. Yoshimaru, and H. Onoda, “The simplest
stacked BST capacitor fbr future DRAMs using a novel low temperature
growth enhanced crystallization/' in IEEE Symposium on VLSI Technology
Digest of Technical Papers, June 1997, pp. 153-154.
[58] T. Okuda and T. Murotani, *'A four-level storage 4-Gb DRAM/5 IEEE
Journal ofSolid State Circuits, vol. 32, pp. 1743-1747, November 1997.
[59] A. Nitayama, Y. Kohyama, and K. Hieda, “Future directions fbr DRAM
memory cell technology,“ in 1998 IEDM Technical Digest, pp. 355-358.
[60] K. Ono, T. Horikawa, T. Shibano, N. Mikami, T. Kuroiwa, T. Kawahara,
S. Matsuno, F. Uchikawa, S. Satoh, and H. Abe, "(Ba, Sr)TiC>3 capacitor
Appendix 397

technology for Gbit-scale DRAMs/' in 1998 IEDM Technical Digest,

pp. 803-806.
[61] H. J. Levy, E. S. Daniel, and T. C. McGill, UA transistorless-current-mode
static RAM architecture/' IEEE Journal of Solid-State Circuits, vol. 33,
pp. 669-672, April 1998.

DRAM Sensing
[62] N. C.-C. Lu and H. H. Chao, ^Half-Ppp/bit-line sensing scheme in CMOS
DRAMs," IEEE Journal of Solid-State Circuits, vol. 19, pp. 451-454,
August 1984.
[63] P. A. Layman and S. G. Chamberlain, <4A compact thermal noise model for
the investigation of soft error rates in MOS VLSI digital circuits/' IEEE
Journal of Solid-State Circuits, vol. 24, pp. 79-89, February 1989.
[64] R. Kraus, ^Analysis and reduction of sense-amplifier offset,n IEEE Journal
ofSolid-State Circuits, vol. 24, pp. 1028-1033, August 1989.
[65] R. Kraus and K. Hoffmann, ^Optimized sensing scheme of DRAMs,n IEEE
Journal of Solid-State Circuits, vol. 24, pp. 895-899, August 1989.
[66] H. Hidaka, Y. Matsuda, and K. Fujishima, **A divided/shared bit-line sensing
scheme for ULSI DRAM cores,M IEEE Journal of Solid-State Circuitsy vol.
26, pp. 473Y78, April 1991.
[67] T. Nagai, K. Numata, M. Ogihara, M. Shimizu, K. Imai, T. Hara, M.
Yoshida, Y. Saito, Y. Asao, S. Sawada, and S. Fuji, “A 17-ns 4-Mb CMOS
DRAM,“ IEEE Journal of Solid-State Circuits^ vol 26, pp. 1538-1543,
November 1991.
[68] T. N. Blalock and R. C. Jaeger, <4A high-speed sensing scheme for IT
dynamic RAMs utilizing the clamped bit-line sense amplifier,'' IEEE
Journal of Solid-State Circuits^ vol. 27, pp. 618-625, April 1992.
[69] M. Asakura, T. Ooishi, M. Tsukude, S. Tomishima, T. Eimori, H. Hidaka,
Y. Ohno, K. Arimoto, K. Fujishima, T. Nishimura, and T. Yoshihara, **An
experimental 256-Mb DRAM with boosted sense-ground scheme,n IEEE
Journal of Solid-State Circuits^ vol. 29, pp. 1303-1309, November 1994.
[70] T. Eirihata, S. H. Dhong, L. M. Terman, T. Sunaga, and Y. Taira, “A variable
precharge voltage sensing/* IEEE Journal of Solid-State Circuits, vol. 30,
pp. 25-28, January 1995.
[71] T. Hamamoto, Y. Morooka, M. Asakura, and H. Ozaki, uCell-plate-line/bit-
line complementary sensing (CBCS) architecture for ultra low-power
DRAMs/5 IEEE Journal of Solid-State Circuits, vol, 31, pp. 592-601, April
1996.
[72] T. Sunaga, t4A full bit prefetch DRAM sensing circuit,IEEE Journal of
Solid-State Circuits, vol. 31, pp. 767-772, June 1996.
398 Appendix

DRAM On-Chip Voltage Generation

[73] M. Horiguchi, M. Aoki, J. Etoh, H. Tanaka, S. Ikenaga, K. Itoh, K. Kajigaya,
H. Kotani, K. Ohshima, and T. Matsumoto, UA tunable CMOS-DRAM
voltage limiter with stabilized feedback amplifier/' IEEE Journal of Solid­
State Circuits, vol. 25, pp. 1129-1135, October 1990.
[74] D. Takashima, S. Watanabe, T. Fuse, K. Sunouchi, and T. Hara, "Low-power
on-chip supply voltage conversion scheme for ultrahigh-density DRAMs,“
IEEE Journal ofSolid-State Circuits, vol, 28, pp. 504-509, April 1993.
[75] T. Kuroda, K. Suzuki, S. Mita, T. Fujita, F. Yamane, F. Sano, A. Chiba,
Y. Watanabe, K. Matsuda, T. Maeda, T. Sakurai, and T. Furuyama, <4Variable
supply-voltage scheme for low-power high-speed CMOS digital design/'
IEEE Journal of Solid-State Circuits, vol. 33, pp. 454-462, March 1998.

DRAM SOI
[76] S. Kuge, F. Morishita, T. Tsuruda, S. Tomishima, M. Tsukude, T. Yamagata,
and K. Arimoto, “SOLDRAM circuit technologies for tow power high speed
multigiga scale memories,'' IEEE Journal ofSolid-State Circuits, vol. 31, pp.
586-591, April 1996.
[77] K. Shimomura, H. Shimano, N. Sakashita, F. Okuda, T. Oashi,
Y. Yamaguchi, T. Eimori, M. Inuishi, K. Arimoto, S. Maegawa, Y. Inoue,
S. Komori, and K. Kyuma, 4tA 1-V 46-ns 16-Mb SOI-DRAM with body
control technique,IEEE Journal of Solid-State Circuits, vol. 32,
pp. 1712-1720, November 1997.

Embedded DRAM
[78] T. Sunaga, H. Miyatake, K. Kitamura, K. Kasuya, T. Saitoh, M. Tanaka,
N. Tanigaki, Y. Mori, and N. Yamasaki, “DRAM macros for ASIC chips,“
IEEE Journal of Solid-State Circuits, vol. 30, pp. 1006-1014, September
1995.

Redundancy Techniques
[79] H. L. Kalter, C. H. Stapper, J. E. Barth, Jr., J. Di Lorenzo, C. E. Drake,
J. A. Fifield, G. A. Kelley, Jr., S. C. Lewis, W. B. van der Hoeven, and
J. A. Yankosky, <4A 50-ns 16-Mb DRAM with a 10-ns data rate and on-chip
ECC," IEEE Journal ofSolid-State Circuits, vol. 25, pp. 1118-1128, October
1990.
[80] M. Horiguchi, J. Etoh, M. Aoki, K. Itoh, and T. Matsumoto, UA flexible
redundancy technique for high-density DRAMs,“ IEEE Journal of Solid­
State Circuits, vol. 26, pp. 12-17, January 1991.
Appendix 399

[81] S. Kikuda, H. Miyamoto, S. Mori, M. Niiro, and M. Yamada, “Optimized

redundance selection based on failure-related yield model for 64-Mb DRAM
and beyond,M IEEE Journal of Solid-State Circuits, vol. 26, pp. 1550-1555,
November 1991.
[82] T. Kirihata, Y. Watanabe, Hing Wong, J. K. DeBrosse, M. Yoshida, D. Kato,
S. Fujii, M. R. Wordeman, P. Poechmueller, S. A. Parke, and 丫 Asao,
wFault-tolerant designs for 256 Mb DRAM/5 IEEE Journal of Solid-State
Circuits, vol. 31, pp. 558-566, April 1996.

DRAM Testing
[83] T. Ohsawa, T. Furuyama, Y. Watanabe, H. Tanaka, N. Kushiyama,
K. Tsuchida, Y. Nagahama, S. Yamano, T. Tanaka, S. Shinozaki, and
K. Natori, **A 60-ns 4-Mbit CMOS DRAM with built-in selftest function,M
IEEE Journal ofSolid-State Circuits, vol. 22, pp. 663-668, October 1987.
[84] P. Mazumder, ^Parallel testing of parametric faults in a three-dimensional
dynamic random-access memory/, IEEE Journal of Solid-State Circuits,
vol. 23, pp. 933-941, August 1988.
[85] K. Arimoto, Y. Matsuda, K. Furutani, M. Tsukude, T. Ooishi, K. Mashiko,
and K. Fujishima, “A speed-enhanced DRAM array architecture with
embedded ECC,“ IEEE Journal of Solid-State Circuits, vol. 25, pp. 11-17,
February 1990.
[86] T. Takeshima, M. Takada, H. Koike, H. Watanabe, S. Koshimaru, K. Mitake,
W. Kikuchi, T. Tanigawa, T. Murotani, K. Noda, K. Tasaka, K. Yamanaka,
and K. Koyama, “A 55-ns 16-Mb DRAM with built-in self-test function
using microprogram ROM,“ IEEE Journal of Solid-State Circuits, vol. 25,
pp. 903-911, August 1990.
[87] T. Kirihata, Hing Wong, J. K. DeBrosse, Y. Watanabe, T. Hara, M. Yoshida,
M. R. Wordeman, S. F可ii, Y Asao, and B. Krsnik, ^Flexible test mode
approach for 256-Mb DRAM," IEEE Journal of Solid-State Circuits,
vol. 32, pp. 1525-1534, October 1997.
[88] S. Tanoi, Y. Tokunaga, T. Tanabe, K. Takahashi, A. Okada, M. Itoh,
Y. Nagatomo, Y. Ohtsuki, and M. Uesugi, <4On-wafer BIST of a 200-Gb/s
failed-bit search fbr 1-Gb DRAM,“ IEEE Journal of Solid-State Circuits,
vol. 32, pp. 1735-1742, November 1997.

Synchronous DRAM
[89] T. Sunaga, K. Hosokawa, Y. Nakamura, M. Ichinose, A. Moriwaki,
S. Kakimi, and N. Kato, “A full bit prefetch architecture for synchronous
DRAMs,M IEEE Journal of Solid-State Circuits^ vol. 30, pp. 998-1005,
September 1995.
400 Appendix

[90] T. Kirihata, M. Gall, K. Hosokawa, J.-M. Dortu, Hing Wong, P. Pfefferi,

B. L. Ji, O. Weinfiirtner, J. K. DeBrosse, H. Terletzki, M. Selz, W. Ellis,
M. R. Wordeman, and O. Kiehl, “A 220-mm/sup2/» fbur-and eight-bank,
256-Mb SDRAM with single-sided stitched WL architecture,*' IEEE Journal
ofSolid-State Circuits, vol. 33, pp. 1711-1719, November 1998.

Low-Voltage DRAM
[91] K. Lee, C. Kim, D. Yoo, J. Sim, S. Lee, B. Moon, K. Kim, N. Kim, S. Yoo,
J. Yoo, and S. Cho, “Low voltage high speed circuit designs for Giga-bit
DRAMs,n in 1996 Symposium on VLSI Circuits^ June 1996, p. 104.
[92] M. Saito, J. Ogawa, K. Gotoh, S. Kawashima, and H. Tamura, ^Technique
for controlling effective in multi-Gbit DRAM sense amplifier,in 1996
Symposium on VLSI Circuits, June 1996, p. 106.
[93] K. Gotoh, J. Ogawa, M. Saito, H. Tamura, and M. Taguchi, 4*A 0.9 V sense­
amplifier driver for high-speed Gb-scale DRAMs/' in 1996 Symposium on
VLSI Circuits, June 1996, p. 108.
[94] T. Hamamoto, Y. Morooka, T. Amano, and H. Ozaki, “An efficient charge
recycle and transfer pump circuit for low operating voltage DRAMs,“ in
1996 Symposium on VLSI Circuits. June 1996, p. 110.
[95] T. Yamada, T. Suzuki, M. Agata, A. Fujiwara, and T. Fujita, ''Capacitance
coupled bus with negative delay circuit for high speed and low power
(lOGB/s < 500mW) synchronous DRAMs,“ in 1996 Symposium on VLSI
Circuits, June 1996, p. 112.

High-Speed DRAM
[96] S. Wakayama, K. Gotoh, M. Saito, H. Araki, T. S. Cheung, J. Ogawa, and
H. Tamura, u10-ns row cycle DRAM using temporal data storage buffer
architecture,in 1998 Symposium on VLSI Circuits, June 1998, p. 12.
[97] Y. Kato, N. Nakaya, T. Maeda, M. Higashiho, T. Yokoyama, Y. Sugo,
F. Baba, Y. Takemae, T. Miyabo, and S. Saito, ^Non-precharged bit-line
sensing scheme for high-speed low-power DRAMs,“ in 1998 Symposium on
VLSI Circuits, June 1998, p. 16.
[98] S. Utsugi, M. Hanyu, Y. Muramatsu, and T. Sugibayashi, “Non-
complimentary rewriting and serial-data coding scheme fbr shared-sense­
amplifier open-bit-line DRAMs,“ in 1998 Symposium on VLSI Circuits, June
1998, p. 18.
[99] Y. Sato, T. Suzuki, T. Aikawa, S. Fujioka, W. Fujieda, H. Kobayashi,
H. Ikeda, T. Nagasawa, A. Funyu, Y. Fujii, K. I. Kawasaki, M. Yamazaki,
and M. Taguchi, “Fast cycle RAM (FCRAM); a 20-ns random row access,
Appendix 401

pipe-lined operating DRAM,” in 1998 Symposium on VLSI Circuits, June

1998, p. 22.

High-Speed Memory Interface Control

[100] S.-J. Jang, S.-H. Han, C.-S. Kim, Y.-H. Jun, and H.-J. Yoo, UA compact ring
delay line for high speed synchronous DRAM,” in 1998 Symposium on VLSI
Circuits, June 1998, p. 60.
[101] H. Noda, M. Aoki, H. Tanaka, O. Nagashima, and H. Aoki, <(An on-chip
timing adjuster with sub-100-ps resolution for a high-speed DRAM
interface,M in 1998 Symposium on VLSI Circuits. June 1998, p. 62.
[102] T. Sato, Y. Nishio, T. Sugano, and Y. Nakagome, 4<5GByte/s data transfer
scheme with bit-to-bit skew control for synchronous DRAM," in 1998
Symposium on VLSI Circuits, June 1998, p. 64.
[103] T. Yoshimura, Y. Nakase, N. Watanabe, Y. Morooka, Y. Matsuda,
M. Kumanoya, and H. Hamano, UA delay-locked loop and 90-degree phase
shifter for 800Mbps double data rate memories,M in 1998 Symposium on
VLSI Circuitsy June 1998, p. 66.

High-Performance DRAM
[104] T. Kono, T. Hamamoto, K. Mitsui, and Y. Konishi, “A precharged-capacitor-
assisted sensing (PCAS) scheme with novel level controlled for low power
DRAMs,M in 1999 Symposium on VLSI Circuits^ June 1999, p. 123.
[105] H. Hoenigschmid, A. Frey, J. DeBrosse, T. Kirihata, G. Mueller, G Daniel,
G Frankowsky, K. Guay, D. Hanson, L. Hsu, B. Ji, D. Netis, S. Panaroni,
C. Radens, A. Reith, D. Storaska, H. Terletzki, O. Weinfiirtner, J. Alsmeier,
W. Weber, and M. Wordeman, 7F2 cell and bitline architecture featuring
tilted array devices and penalty-free vertical BL twists fbr 4Gb DRAMs” in
1999 Symposium on VLSI Circuits, June 1999, p. 125.
[106] S. Shiratake, K. Tsuchida, H. Toda, H. Kuyama, M. Wada, F. Kouno,
T. Inaba, H. Akita, and K. Isobe, UA pseudo multi-bank DRAM with
categorized access sequence/' in 1999 Symposium on VLSI Circuits, June
1999, p. 127.
[107] Y. Kanno, H. Mizuno, and T. Watanabe, WA DRAM system for consistently
reducing CPU wait cycles/* in 1999 Symposium on VLSI Circuits^ June
1999, p. 131.
[108] S. Perissakis, Y. Joo, J. Ahn, A. DeHon, and J. Wawrzynek, ^Embedded
DRAM for a reconfigurable array,“ in 1999 Symposium on VLSI Circuits,
June 1999, p. 145.
402 Appendix

[109] T. Namekawa, S. Miyano, R. Fukuda, R. Haga, O. Wada, H. Banba,

S. Takeda, K. Suda, K. Mimoto, S. Yamaguchi, T. Ohkubo, H. Takato, and
K. Numata, ^Dynamically shift-switched dataline redundancy suitable for
DRAM macro with wide data bus,” in 1999 Symposium on VLSI Circuits,
June 1999, p. 149.
[110] C. Portmann, A. Chu, N. Hays, S. Sidiropoulos, D. Stark, P. Chau,
K. Donnelly, and B. Garlepp, 4tA multiple vendor 2.5-V DLL fbr 1.6-GB/s
RDRAMs,M in 1999 Symposium on VLSI Circuits. June 1999. p. 153.

High-Performance Logic
[111] D. Harris, Skew Tolerant Circuit Design. San Francisco, CA: Morgan
Kaufman Publishers, 2001.
[112] W. J. Dally and J. W. Poulton, Digital Systems Engineering. Cambridge, UK:
Cambridge University Press, 1998.
[113] L E. Sutherland, uMicropipelines,,, Communications of the ACM, vol. 32, pp.
720-738, June, 1989.
[114] C. Svenssen and J. Yuan, ^High-speed CMOS circuit technique,IEEE
JSSC, vol. 24, pp 62-70, Feb. 1989.
[115] A. Chandrakasan, W. J. Bowhill, and F. Fox, High-performance
Microprocessor Circuits. Piscataway, NJ: IEEE Press, 2001.
[116] I. Sutherland and S. Fairbanks, t4GasP: a minimal FIFO control,Async
2001, pp. 46-53, March 2001.
[117] C. J. Myers, Asynchronous Circuit Design. Hoboken, New Jersey: John
Wiley & Sons, Inc, 2001.

I/O and Timing

[118] Y. Moon, J. Choi, K. Lee, D. K. Jeong, and M. K. Kim, “An all-analog
multiphase delay-locked loop using a replica delay line for wide-range
operation and low-jitter performance,IEEE Journal on Solid-State Circuits,
vol. 35, pp. 377-384, Mar. 2000.
[119] H. H. Chang, J. W. Lin, C. Y. Yang, and S. L Liu, “A wide-range delay-
locked loop with a fixed latency of one clock cycleJ IEEE Journal on Solid­
State Circuits, vol. 37, pp. 1021-1027, Aug. 2002.
[120] Yasuhiko Sasaki,Naoki Kato, and Hiroaki Nakaya, ^Constant-Ratio-Coupled
Multi-Grain Digital Synchronizer With Flexible Input-Output Delay
Selection fbr Versatility in Low-Power Applications/5 IEEE Journal on
Solid-State Circuits, vol. 41, May 2006.
[121] K. Minami, M. Mizuno, H, Yamaguchi, T. Nakano, Y. Matsushima, Y. Sumi,
T. Sato, H. Yamashida, and M. Yamashina, UA 1 GHz portable digital delay-
Appendix 403

locked loop with infinite phase capture ranges,M in IEEE Int. Solid-State
Circuits Conf. (ISSCC) Dig. Tech. Papers, Feb. 2000, vol. 469, pp. 350-351.
[122] J.-B. Lee, K.-H. Kim, C. Yoo, S. Lee, O.-G. Na, C.-Y. Lee, H.-Y. Song, J.-S.
Lee, Z.-H. Lee, K.-W. Yeom, H.-J. Chung, L-W. Seo, M.-S. Chae, Y.-H.
Choi, and S,-I. Cho, ^Digitally-controlled DLL and I/O circuits for 500
Mb/s/pin xl6 DDR SDRAM,“ in IEEE Int, Solid-State Circuits Conf.
(ISSCC) Dig. Tech. Papers, Feb. 2001, vol. 431, pp. 68-69.
[123] T. Matano, Y. Takai, T. Takahashi, Y. Sakito, I. Fujii, Y. Takaishi,
H. Fujisawa, S. Kubouchi, S. Narui, K. Arai, M. Morino, M. Nakamura,
S. Miyatake, T. Sekiguchi, and K. Koyama, “A 1-Gb/s/pin 512-Mb DDRII
SDRAM using a digital DLL and a slew-rate-controlled output buffer,“
IEEE Journal on Solid-State Circuits, vol. 38, pp. 762-768, May 2003.
[124] J.-T. Kwak, C.-K. Kwon, K.-W. Kim, S.-H. Lee, and J.-S. Kih, “Low cost
high performance register-controlled digital DLL fbr IGbps x32 DDR
SDRAM," in Symposium VLSI Circuits Dig. Tech. Papers, 2003, pp.
283-284.
[125] H. H. Chang and S. I. Liu, “A wide-range and fast-locking all-digital cycle-
controlled delay-locked loop,“ IEEE Journal on Solid-State Circuits, vol.
40, pp. 661-670, Mar. 2005.
[126] J. S. Wang, Y. W. Wang, C. H. Chen, and Y. C. Liu, “An ultra-low power
fast-lock-in small-jitter all-digital DLL," in IEEE Int, Solid-State Circuits
Conf. (ISSCC) Dig. Tech. Papers, Feb. 2005, pp. 422-433.
[127]1S. R. Han and S. I. Liu, “A 500-MHz-1.25-GHz fast-locking pulse-width
control loop with presettable duty cycle,v IEEE Journal on Solid-State
Circuits, vol. 39, pp. 463-468, March 2004.
[128] M. Mota and J. Christiansen, four-channel self-calibrating high-
resolution time to digital converter,M in Proc. IEEE Int. Conf. Electronics,
Circuits, and Systems, 1998, pp. 409-412.
[129] Y. J. Wang, S. K. Kao, and S. I. Liu, ^All-digital delay-locked loop/pulse-
width-control loop with adjustable duty cycles,'' IEEE Journal on Solid­
State Circuits, vol. 41, June 2006.
[130] T. Hsu, B. Shieh, and C. Lee, “An all-digital phase-locked loop (ADPLL)-
based clock recovery circuit,,5 IEEE Journal on Solid-State Circuits, vol. 34,
pp. 1063-1073, Aug. 1999.
[131] J. Dunning, G. Garcia, J. Lundberg, and E. Nuckolls, uAn all-digital phase-
locked loop with 50-cycle lock time suitable fbr high-performance
microprocessors,v IEEE Journal on Solid-State Circuits, vol, 30, pp.
412T22, Apr. 1995.
[132] T. Saeli, K. Minami, H. Yoshida, and H. Suzuki, UA direct-skew-detect
synchronous mirror delay fbr application-specific integrated circuits, IEEE
Journal on Solid-State Circuits, vol. 34, pp. 372-379, Mar. 1999.
404 Appendix

[133] K. Sung, B. D. Yang, and L.S Kim, “Low power clock generator based on
an area-reduced interleaved synchronous mirror delay scheme,in Proa
IEEE Int. Symposium Circuits and Systems. 2002, pp. 671-674.
[134] Y.-H. Koh and O.-K. Kwon, “A fast-locking digital delay line with duty
conservation/' in Proc. IEEE Asia-Pacific Conf. Circuits and Systems, 1998,
pp. 287-290.
[135] T. Seaeki, H. Nakamura, and J. Shimizu, “A 10 ps jitter 2 clock cycle lock
time CMOS digital clock generator based on an interleaved synchronous
mirror delay scheme,in Symposium VLSI Circuits Dig. Tech. Papers, 1997,
pp. 109-110.
[136] D. Shim, D.-Y. Lee, S. Jung, C.-H. Kim, and W. Kim, “An analog
synchronous mirror delay for high-speed DRAM application/' IEEE Journal
on Solid-State Circuits, vol. 34, pp. 484-493, April 1999.
[137] Y. M. Wang and J. S. Wang, “A low-power half-delay-line fast skew­
compensation circuit/' IEEE Journal on Solid-State Circuits, vol. 39, June
2004.
[138] Y. Jung, S. Lee, D. Shim, W. Kim, C. Kim and S. Cho, “A dual-loop delay-
locked loop using multiple voltage-controlled delay lines/' IEEE Journal on
Solid-State Circuits, vol. 36, pp. 784-791, May 2001.
[139] J. Kim, S. Lee, T. Jung, C. Kim, S. Cho and B. Kim, “A low-jitter mixed­
mode DLL for high-speed DRAM applications,IEEE Journal on Solid­
State Circuits, vol. 35, pp. 1430-1436, Oct. 2000.
[140] G Dehng, J. Lin, and S. Liu, “A fast-lock mixed-mode DLL using a 2-b SAR
algorithm,“ IEEE Journal on Solid-State Circuits^ vol. 35, pp. 1464-1471,
Oct. 2001.
[141] B. Razavi, Design of Analog CMOS Integrated Circuits. New York:
McGraw-Hill, 2001.
[142] T. H. Lee, The Design of CMOS Radio-Frequency Integrated Circuits.
Cambridge, U.K.: Cambridge Univ. Press, 1998.
[143] T. Gabara, W. Fischer, J, Harrington, and W. Troutman, ^Forming damped
LRC parasitic circuits in simultaneously switched CMOS output buffers/5
IEEE Journal on Solid-State Circuits^ vol. 32, pp. 407-418, March 1997.
[144] M. Dolle, ^Analysis of simultaneous switching noise/' in Int. Symposium
Circuits Systems, May 1995, pp. 904-907.
[145] K. Kim et aL, "A 1.4 Gb/s DLL using 2nd order charge-pump scheme with
low phase/duty error for high-speed DRAM application/, in IEEE Int. Solid­
State Circuits Conf. (ISSCC) Dig. Tech. Papers, 2004, pp. 212-523.
[146] S. Lee et aL, "A 1.6 Gb/s/pin double data rate SDRAM with wave-pipelined
CAS latency control,M in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig.
Tech. Papers. 2004, pp. 210-213.
Appendix 405

[147] H. Fujisawa, T. Takahashi, H. Yoko, I. Fujii, Y. Takai, and M. Nakamura,441-

Gb/s/pin multi-gigabit DRAM design with low impedance hierarchical I/O
architecture,in VLSI Circuits Symposium Dig. Tech. Papers, 2002, pp.
118-119.
[148] C. Ybo, K. Kyung, K. Lim, H.-C. Lee, J.-W. Chai, N.-W. Heo, D.-J. Lee, and
C.-H Kim, “A 1.8-V 700-Mb/s/pin 512-Mb DDR-II DRAM with on-die
termination and oflf-chip driver calibration/* IEEE Journal on Solid-State
Circuits, vol. 39, pp. 941-951, June 2004.
[149] C. Y. Chung and A. Waizman, “Optimum signal integrity through
appropriate analysis of signal return path and power delivery/' in Elec.
Components and Tech. Conference, 2001, pp. 1402-1407.
[150] C. S. Choy, C. F. Chan, M. H. Ku, and J. Povazanec, ^Design procedure of
low-noise high-speed adaptive output drivers,^^ in IEEE International
Symposium on Circuits and Systemsy 1997, pp. 1796-1799.
[151] S. R. Vemuru, “Effects of simultaneous switching noise on the tapered
buffer design/' IEEE Transactions on Very Large Scale Integration (VLSI)
Systems, vol. 5, Sept. 1997.
[152] S. J. Jou, S. H. Kuo, J. T. Chiu, and T. H. Lin, “Low switching noise and
load-adaptive output buffer design techniques/' IEEE Journal ofSolid State
Circuits, vol. 36, Aug. 2001.
[153] L. Sumanen, M. Waltari, V. Hakkarainen, and K. Halonen, UCMOS dynamic
comparator for pipeline AID converters,'' in IEEE International Symposium
Circuits and Systems, ISCAS 2002, vol. 5, pp. 157-160.
[154] D. Somasekhar and K. Roy; ^Differential current switch logic: a low power
DCVS logic family,IEEE Journal on Solid-State Circuits, vol. 31, pp. 981­
991, July 1996.
[155] H. Tamura et al., <c5-Gb/s bidirectional balanced-line link compliant with
plesiochronous clocking,*' in \EEE Int. Solid-State Circuits Conf. Dig. Tech.
Papers、Feb. 2001, pp. 50-51.
[156] C. H. Lim and W. R. Dassch, ^Output buffer with self-adjusting slew rate
and on-chip compensation,'' in Proc. of IEEE Symposium on IC/Package
Design Integration, 1998, pp. 51-55.
[157] T. Shirotori and K. Nogami, "PLL-based, impedance-control led output
buffer,M in Symposium on VLSI Circuits Digest of Technical Papers、1991,
pp. 49-50.
[158] S.-J. Jou, W.-C. Cheng, and Y.-T. Lin, ^Simultaneous switching noise
analysis and low bouncing buffer design,“ in IEEE Conf, on Custom
Integrated Circuits. May 1998, pp. 545-548.
[159] T. Gabara, W. Fischer, J. Harrington, and W. Troutman, ^Forming damped
LRC parasitic circuits in simultaneously switched CMOS output buffers,M in
IEEE Conf, on Custom Integrated Circuits, May 1996, pp. 277-280.
406 Appendix

[160] R. Senthinthan and J. L. Prince, ^Application specific CMOS output driver

circuit design techniques to reduce simultaneous switching noise/' IEEE
Journal on Solid-State Circuits, vol. 28, pp. 1383-1388, Dec. 1993.
[161] H. I. Hanafi, R. H. Dennard, C.-L. Chen, R. J., and D. S. Zicherman, ^Design
and characterization of a CMOS ofF-chip driver/receiver with reduced
power-supply disturbance/' IEEE Journal on Solid-State Circuits, vol. 27,
pp. 783-791, 1992.
[162] E. Chioffi, F. Maloberti, G Marchesi, and G. Torelli, ^High-speed, low-
switching noise CMOS memory data output buffer,” IEEE Journal on Solid­
State Circuits, vol. 29, pp. 1359-1365, 1994.
[163] A. B. Dowlatabadi, UA robust, load-insensitive pad driver,M IEEE Journal on
Solid-State Circuits, 2000, vol. 35, pp. 660-665.
[164] J. Kennedy, R. Mooney, R. Ellis, J. Jaussi, S. Borkar, J. H. Choi, J. K. Kim,
C. K Kim, W.-S. Kim, C.-H. Kim, S. I. Cho, S. Loeffler, J. Hoffmann,
W. Hokenmaier, R. Houghton, and T. Vogelsang, t4A 3.6-Gb/s point-to-point
heterogeneous-voltage-capable DRAM interface for capacity-scalable
memory subsystems,^ IEEE Journal on Solid-State Circuits, vol. 40, pp.
233-244, Jan. 2005.
Glossary
1T1C A DRAM memory cell consisting of a single MOSFET access tran­
sistor and a single storage capacitor.
Bitline Also called a digitline or columnline. A common conductor made
from metal or polysilicon that connects multiple memory cells together
through their access transistors. The bitline is ultimately used to connect
memory cells to the sense amplifier block to permit Refresh, Read, and
Write operations.
Bootstrapped Driver A driver circuit that employs capacitive coupling to
boot, or raise up, a capacitive node to a voltage above Vcc.
Buried Capacitor Cell A DRAM memory cell in which the capacitor is
constructed below the digitline.
Charge Pump See Voltage Pump.
Clock Distribution Network, CDN Also called clock tree. A clock distri­
bution network delivers clock from one place to multiple locations. In
order to maintain signal quality and minimize timing skew, buffers are
generally inserted at various points, and matched routing is also applied.
CMOS, Complementary Metal-Oxide Semiconductor A silicon tech­
nology for fabricating integrated circuits. Complementary refers to the
technology's use of both NMOS and PMOS transistors in its construc­
tion. The PMOS transistor is used primarily to pull signals toward the
positive power supply VDD. The NMOS transistor is used primarily to
pull signals toward ground. The metal-oxide semiconductor describes the
sandwich of metal oxide (actually polysilicon in modem devices) and sil­
icon that makes up the NMOS and PMOS transistors.
COB, Capacitor over Bitline A DRAM memory cell in which the capac­
itor is constructed above the digitline (bitline).

407
408 Glossary

Columnline See Bitline.

Column Redundancy The practice of adding spare digitlines to a mem­
ory array so that defective digitlines can be replaced with nondefective
digitlines.
Current-Mode Logic, CML A small-swing, high-speed, fully differential
logic style.
DCSA, Direct Current Sense Amplifier An amplifier connected to the
memory array I/O lines that amplifies signals coming from the array.
Delay-Chain Logic A heuristic logic style comprised of a series of delay
elements that create a sequence of events. The delay elements tend to be
implemented using one-shot (monostable multi-vibrator) circuits and
high-load static logic chains. This type of circuitry issuited for post-sili­
con adjustments and is commonly found in the peripheral array logic
region.
Digitline See Bitline.
DCL, Read Latency or CAS Latency
DLL, Delay-Locked Loop A circuit that generates and inserts an opti­
mum delay to temporarily align two signals. In DRAM, a DLL synchro­
nizes the input and output clock signals of the DRAM to the I/O data
signals.
DRAM, Dynamic Random Access Memory A memory technology that
stores information in the form of electric charge on capacitors. This tech­
nology is considered dynamic because the stored charge degrades over
time due to leakage mechanisms. The leakage necessitates periodic
Refresh of the memory cells to replace the lost charge.
Dummy Structure Additional circuitry or structures added to a design,
most often to the memory array, that help maintain uniformity in live cir­
cuit structures. Nonunifbrmity occurs at the edges of repetitive structures
due to photolithography and etch process limitations.
EDO DRAM, Extended Data Out DRAM A variation of fast page mode
(FPM) DRAM. In EDO DRAM, the data outputs remain valid for a spec­
ified time after CAS* goes HIGH during a Read operation. This feature
permits higher system performance than would be obtained otherwise.
Efficiency A design metric, which is defined as the ratio of memory array
die area and total die area (chip size). It is expressed as a percentage.
Equilibration Circuit A circuit that equalizes the voltages of a digitline
pair by shorting the two digitlines together. Most often, the equilibration
Glossary 409

circuit includes a bias network, which helps to set and hold the equilibra­
tion level to a known voltage (generally VDDH) prior to Sensing.
Feature Size Generally refers to the minimum realizable process dimen­
sion. In the context of DRAM design, however, feature size equates to a
dimension that is half of the digitline or wordline layout pitch.
Folded DRAM Array Architecture A DRAM architecture that uses non­
crosspoint-style memory arrays in which a memory cell is placed only at
alternating wordline and digitline intersections. Digitline pairs, for con­
nection to the sense amplifiers, consist of two adjacent digitlines from a
single memory array. For layout efficiency, each sense amplifier connects
to two adjacent memory arrays through isolation transistor pairs.
FPM, Fast Page Mode A second-generation memory technology permit­
ting consecutive Reads from an open page of memory, in which the col­
umn address could be changed while CAS* was still LOW.
Gate-Delay Logic In this text, gate-delay logic refers to a section of dom­
ino logic that is first made functionally complete through a conversion of
static signals to dual-rail, monotonic signals. Following conversion, the
logic signals are used as precharge/evaluate terms to downstream domino
gates. This style of logic is used in the command decoder and address
register logic to avoid clock period dependencies fbr data access latency.
Helper Flip-Flop, HFF A positive feedback (regenerative) circuit for
amplifying the signals on the I/O lines.
I/O Devices MOSFET transistors that connect the array digitlines to the
I/O lines (through the sense amplifiers). Read and Write operations
from/to the memory arrays always occur through I/O devices.
Intrinsic Delay Also called intrinsic forward path delay. The intrinsic
delay represents a propagation delay of the critical timing path from
inputs (eg, CLK) to outputs (e.g., DQs), excluding the insertion delay
from timing adjustment circuits (e.g., DLL).
Isolation Devices MOSFET transistors that isolate array digitlines from
the sense amplifiers.
Matched Routing A method of maintaining an original timing relationship
after signal distribution and placement.
Mbit, Memory Bit A memory cell capable of storing one bit of data. In
modem DRAMs, the mbit consists of a single MOSFET access transistor
and a single storage capacitor. The gate of the MOSFET connects to the
wordline or rowline, while the source and drain of the MOSFET connect
to the storage capacitor and the digiti ine, respectively.
410 Glossary

Memory Array An array of memory or mbit cells.

Multiplexed Addressing The practice of using the same chip address
pins for both the row and column addresses. The addresses are clocked
into the device at different times.
On-Die Termination,ODT The impedance put on DRAM high-speed I/O
to reduce reflections and improve signal integrity. The termination value
depends on the system configuration and channel impedance.
Open DRAM Array Architecture A DRAM architecture that uses cross­
point-style memory arrays in which a memory cell is placed at every
wordline and digitline intersection. Digitline pairs, for connection to the
sense amplifiers, consist of a single digitline from two adjacent memory
arrays.
PDN, Power Delivery Network. The system of wiring and decoupling
elements connects individual transistors and devices to the system power
supply. In DRAM, the basic PDN includes on-die power bussing,
decoupling capacitors, and the device package power bussing. A more
complete PDN description would also include the DIMM, DIMM
connector, mother board, and related decoupling capacitors.
Peripheral Array Logic This section of logic is part of the helper flip-flop
(HFF) control and column timing chain logic. Often this logic is imple­
mented using the t€delay-chain^^ logic style described in this text.
Peripheral Logic Interface All of the logic necessary for implementing
the system level protocol. Generally, this logic resides between the I/O
circuits and the peripheral array logic.
Pitch The distance between like points in a periodic array. For example,
digitline pitch in a DRAM array is the distance between the centers or
edges of two adjacent digitlines.
RAM, Random Access Memory Computer memory that allows access
to any memory location without restrictions.
Read Pipe The partial multiplexor circuit following the HFF logic. This
logic acts as a FIFO to match the latency of array data accesses with the
programmed latency and throughput requirements of the output circuit
path. There are often data reordering functions as part of the logic in this
circuit or set of circuits.
Refresh The process of restoring the electric charge in DRAM memory
cell capacitors to full levels through Sensing. Note that Refresh occurs
every time a wordline is activated and the sense amplifiers are fired.
Rowline See Wordline.
Glossary 411

SDRAM, Synchronous DRAM A DRAM technology in which address­

ing, command, control, and data operations are accomplished in synchro­
nism with a master clock signal.
Sense Amplifier A type of regenerative amplifier that senses the contents
of memory cells and restores them to full levels.
Signal Integrity The quality of signal transportation during circuit opera­
tion.
Signal Return Path Shift, SRPS Return path shift may be caused when a
signal's reference plane changes during transmission due to being pulled
up or down by drivers at the near or far end.
Simultaneous Switching Noise, SSN I/O power supply noises generated by
simultaneously switching output drivers.
Source Synchronous Clock edges transmitted with the data used to launch
it are available at the receive end for capture.
Trench Capacitor A DRAM storage capacitor fabricated in a deep hole
(trench) in the semiconductor substrate.
Voltage-Controlled Delay Line, VCDL An adjustable delay chain in
which the delay of each stage is controlled by adjusting a bias voltage.
VCDL is generally found in analog delay-locked or phase-locked loops.
Voltage Pump Also called a charge pump. A circuit for generating volt­
ages that lie outside of the power supply range.
WL, Write Latency
Wordllne Also called a rowline. A wordline is a polysilicon conductor for
forming memory cell access transistor gates and connecting multiple
memory cells into a physical row. Driving a wordline HIGH activates, or
turns ON, the access transistors in a memory array row.
Write Pipe The partial demultiplexor circuit following the write data cap­
tures latches. This logic matches the data rate of the input circuit path
with the column cycle time of the array data path in the peripheral logic
interface. The demultiplexor is also commonly referred to as a serial-to-
parallel converter.
 path elements, 14, 126-130,
Index 131-134
predecode, 60-61, 100, 105,
Numerics 107-108,126, 128-134
1T1C memory cell, 10, 22-23, 33 transition detection, 131-133
3-transistor DRAM cell, 6-7 Address register logic, 316, 324-328
6F2, 36, 65-66, 70, 75, 84-86, 89 Address to Write delay time (以汾，5
8F2, 35-36, 66, 75, 77-79 Alignment
strobe, 195,207, 333, 342—345
A
Access time "we)，4, 18,68, 126, 135, clock and data, 194-195,207-211,
169,245, 301-303,313, 324 213, 216, 252, 337-338,
345, 352-354, 359-362
Access time
phase, 281, 352, 361
from C4S* (1℃)，14, 169, 301,
303 Arbiter, 142-143, 272-273
from column access (3), 169, Array, 2-A, 9
347, 349—350, 353, architecture, 31, 35,65-98
355-356 architecture, comparison, 93-97
107, 169 efficiency, 70-73, 76, 79-80,
ACT. See Active pull-up 93-94
Active pull-up (ACT), 26-30, 48-53,
157 B
Active to Read or Write delay (Z/?cd)» Bandgap reference, 153-154, 217
169,302 Bandwidth, 154,175-176,182-183,
Address 243,246,248,292,296-297, 304,
318, 323-324, 346, 368, 385-387
decode trees, 54,58-60,62,88,99,
102, 131 Binning, 170,314
multiplexing, 3,7-8, 126,180-181 Bias circuits, 44-46, 51, 113-114,
118-120, 153, 197, 278-279, 334

413
414 Index

Bilevel C/M input capacitance, 2

digitline, 83-97 CL. See CAS latency
digitline array architecture, 88-93 Clock cycle time (匕长)，169, 187, 200,
Bitline, 9 203, 210, 255,262—263, 268-269,
277, 281-282, 301-303, 327, 355,
capacitance number, limitations,
361,364
10
Clock (data) recovery circuit (CRC,
PRECHARGE voltage, 24,45
CDR), 208-209, 252, 292
Bitline over capacitor (BOC), 33
Clock distribution network (CDN),
BOC. See Bitline over capacitor 195, 202,208-209,213,231,251,
BOOST signal, 155-161 257,289-291,322,371
Bootstrap wordline driver, 54-58, Cmbit, Mbit capacitance, 26
60-61,73 CML. See Current mode logic
Breakdown voltages, 147 COB. See Capacitor over bitline
Bundled signal protocol, 328 Co・design, 373, 376, 382
Buried capacitor, 33, 39-42 Column address path, 14, 126, 131,
Buried digitline, 41-42 175, 187
Bum-in, 149-152, 161,242, 260 decoder, 4, 13-14, 325
Burst length, 21, 171, 176,219, 297, redundancy, 60, 99-104, 107, 131,
303, 322, 326, 329 178, 188
Column address strobe (G45), 3, 8,
C 13-17,29, 126, 169, 301,311,320
Capacitor Column cycle time Uccd)，170,
buried, 33, 39T2 175-176, 180,296-297, 303, 322,
trench, 42-44 326,331,348,355,357,372,385
Capacitor over bitline (COB), 41 Column hold time 8
Capacity, 159 Column decoder, 30,99-102,133,180,
189-190, 195, 306,318
Capture latch, 193, 203, 209, 288,
312-313,320, 324-325, 338-339, Column select (CSEL), 28,47, 51, 87,
365,371 100-102, 118, 133, 177-182, 189
CAS. See Column address strobe Column setup time 8
latency (CL), 20, 169,219, 245, Columnline, 6-7, 10, 13-14, 22
302-303,312-313 Conversion
C出?诂 digitline capacitance, 25 dual-rail, 320, 324
Charge pump, 147, 158 CML-to-CMOS, (CML2CMOS),
Charge-pump phase detector (CPPD), 283, 285
253,283 parallel-to-serial (P2S), 188, 219,
244, 346
Index 415

serial-to-parallel (S2P), 330-331, Delay-locked loop (DLL), 206, 227,

339 251,337,359-362, 371
signal, 196-197,217 all-digital, 253,254
static-to-dynamic, 312, 320 analog, 276, 286-290
time-to-digital. See Measure-con­ mixed-mode, 254, 276, 285-286,
trolled delay 289-291
voltage, 57 register-controlled, 135-145
voltage-to-time (VTC), 227 Delay variations, 196-197,202, 258,
CPPD. See Charge-pump phase detec­ 262,267,282, 286,298, 304,308,
tor 349,359, 361
CSEL. See Column select Demultiplexor, Write data, 330,
338-346
Current mode logic (CML), 193,
197-198,212,247-248, 277-278, Design examples, 79, 83-97, 150, 350
282, 371—372 De-skewing, 195,207, 251,253,
287-288
D Deterministic latency, 188, 333, 346,
Data-bus inversion (DBI), 194, 358
236-237 Digital filter, 226-227,269, 272
Data eye and diagram, 196,208-209, Digitline, 9-10, 22,33
222-223, 237,239-241,332
capacitance 23,25, 39,46,
Data input buffer, 111-112,127 66-68,91
Data path elements, 111 construction, 42
Data Read open digitline array layout, 24,
muxes, 122, 125 39-40, 67-80, 84-86, 89,
path, 118 92-96
DC sense amplifier (DCSA), 111, pair, 24,26, 36,44T7, 77, 85, 89
118-121, 125, 176,319 twist, 36-37, 76
DCD and DCC. See Duty-cycle distor­ voltages, 25
tion and correction Discontinuity, 261,267, 274
DCSA. See DC sense amplifier DLL. See Delay-locked loop
DDR. See Double data rate Domino logic, 309-314
DDR2,31, 171, 175, 180, 185, Donut gate, 56
221-222, 235, 243, 245, 296
Double data rate (DDR), 14,16-17,31,
DDR3,31, 177, 185, 194 124, 135-137,171,180,185, 193,
Decoupling capacitor, 231, 238-241, 221,249,252, 269, 296,303, 322,
243,258, 374 332-335,351, 359, 367
Delay-chain logic, 306-308, 329 DQ, 13, 17-18,31,52, 124-125,
416 Index

DQ, continued, 135-137, 175, second generation, 7, 9, 13, 17

184-185, 209, 230-233, 239, 245, synchronization, 135-145, 171,
255,351 187, 193, 245, 251-289,
mask {DQM), 18-19 346, 357
strobe (DQS), 16, 135-137, 194, synchronous, 15, 99-100, 107,
245 169-171, 174, 193, 259,
DQM. See DQ mask 298, 301,308,319
DQS. See DQ strobe test modes, 125-126, 273
DQS-DQ skew, (tDQSQ), 245 testing, 52, 314-315
Dual-rail logic, 312 types and operation, 1-21
Duty-cycle distortion and correction,
113, 124, 194, 208,211,217, E
251-252, 260, 278, 289-291,344 EDO. See Extended data out
Dual-loop, 274, 286, 289 Edge recovery, 254, 276, 285, 289
DVC2, 147, 165-166 EQ. See Equilibrate

Dynamic random access memory Equilibrate (EQ), 44-45

(DRAM), 1, 13 Equilibration, 44-46,177
1T1C memory cell, 10, 22-23, 33 Electrostatic discharge (ESD), 124,
3-transistor cell, 6-7 200,219, 242-244
address multiplexing, 3, 7-8, 126, Extended data out (EDO), 7, 14, 20,
301 99-100, 107, 126, 170, 301,303
array, 2-4, 9, 33
F
evolution, 1, 170-171,249, 301 Fast page mode (FPM), 7, 14, 20,
first generation, 1 99-100, 107,170
internal 匕工,91, 147—151 FBGA. See Fine ball grid array
layout, 24,31-33,35,39,41,46 FCIP. See Four-layer flip chip
Mbit, 22-23 FIFO
modes of operation, 7, 13-14, 20, asynchronous, 327-332, 357
301 orthogonal, 352-358
opening a row, 10, 11-13, 29-30 Fine ball grid array (FBGA), 231-232,
organization, 12 235
power consumption, 45-46 Folded array, 31, 35-36
power dissipation, 39, 48, 67-68, architecture, 77-85
77,90-91,385-386, 389 Folded digitline architectures, 77, 85
reading data out, 4-6, 135 Four-layer flip chip (FCIP), 373, 376,
Refresh, 5-6, 10, 12-13, 48-49, 381
67, 77,91, 128-129, 387
Index 417

FPM. See Fast page mode transmission line, 112,220-223

Frequency slewing, 262 Input buffer, 1,111-115,127-128,132,
Fully differential amplifier, 113, 197 136, 197, 255,258,371
Input capacitance (Cin ), 2
G Input phase selection, 262-265,281,
Gate-delayed logic, 320 285
GDDR. See Graphics double data rate Input timing adjustment, 193, 206-210
Graphics double data rate, 171 Interpolation, 259, 266-267,277,
GDDR2, 171 286-288, 292
GDDR3, 171,174,176-177,184-189, Inter-symbol interference (ISI), 194,
203, 228,238, 303-305, 355, 208, 221,249, 252, 333
372-382 Intrinsic delay, 255-257,262-263,266,
GDDR4, 171, 175, 177, 184,222, 272,286,290-291
231-232,245,252, 297, 305,357 Isolation, 42,44,46,48-51, 76, 87,
Global 89-90, 155,254,258
circuitry, 111-144
J
column select lines, 72, 178,189
JEDEC, 184,217
row decoders, 54, 57, 59, 70, 79,
Jitter, 140-141, 194-196, 208-211,
88,90, 93, 101
233-234,237, 241,245,251-253,
266, 271-272,282, 288,333, 360
Helper flip-flop (HFF), 31,118-119,
L
121-122,176-183, 306, 319, 325,
372 Latency, 18, 175, 187-189
HFF. See Helper flip-flop Leakage, 22,34,40,45,48-49, 56,58,
105, 147,246, 299, 309-310, 387
High-input trip point ), 2, 112,
127,194 Linearity, 225, 229,236, 277-278
Lock time, 140,252-253, 264,
268-270,284-285,289-291
I/O, 16-18,28,44, 116-124, 135, 193, Loop filter, 272, 277-278,283-284
219,248 Low-input trip point (P/t), 2,112,127,
I/O transistors, 28,31,46-53,100-101, 194
111, 116,118,176-178, 306
Impedance M
calibration, 219,224-230,243 Matched routing, 193,195,200-203,
207-208, 234,245, 315, 320, 323,
characteristic, 220-221, 233 335,365
high-impedance (tri-state), 116, Mbit, 22
125, 246
6F2, 38, 65-66, 70, 75, 84-86, 89
418 Index

Mbit, continued Output driver, 124-125,175, 183, 187,

8F2, 35-36, 66, 75, 77-79 219, 224, 228-230, 234-235, 243,
245-249
capacitance (Cmbit), 25
Oscillator circuits, 159
layout, 33-34
Output buffer circuit, 30,122,124-125,
Mbit pair layout, 34
242, 283
Measure-controlled delay (MCD), 253,
Oxide-nitride-oxide (ONO) dielectric,
270-271,285
40-43
Memory element, 2,4,10-11
Memory array, 2,4, 9-12, 23, 171, P
180-183, 190, 301 Page mode, 7,13-15, 99-100, 126,
layout, 24 132, 169-170, 301-302,304,315
size, 4, 12 Page size, 11-12
Monotonic signal, 310,320, 323, 325 PD. See Phase detector
Multiplexed addressing, 3, 8-9, 126 PDN. See Power delivery network
Multiplexor, Read data, 246-359 Peripheral array logic, 295-296,
308-309,318-319, 325-355
N Peripheral circuitry, 99-100
N sense amp latch (NLAT*), 26-30, Peripheral logic interface, 295-367
48-53
Phase detector (PD), 136-143, 227,
Negative bias temperature instability 254, 263, 267-273, 275, 278, 284,
(NBTI), 260 337, 359, 362
Nibble mode, 7,13-15 Phase drivers, 126, 131
NLAT*. See N sense-amp latch Phase generation, 264, 277,285,
289-290
o Phase-locked loop (PLL), 206,
On-die termination (ODT), 194, 208-210, 252—253,292, 337-338,
221-223, 228,236, 243,246 359,371
ONO dielectric. See Oxide-nitride-ox- Pipeline, 20,100,183, 188, 297,327,
ide dielectric 371
Operational amplifiers (op-amp), 229, Pitch, 34-35,78, 116
278
cells, 46
CMOS, 150
digitline, 34, 39, 68, 70, 78-79
power, 152, 154-156
wordline, 70, 79, 84, 88-93
standby, 154, 156
PLL. See Phase-locked loop
Open architectures, 65-75
PNP transistors, 153
Open array, 31, 38-39, 79
Polycide, 11,23,40
Opening a row, 10-13,29-30
Index 419

Power Read strobe (DQS) access time from

consumption, 45-46, 58, 140, 200, CLK/CLK^tDQSCK), 245
213,259, 261,291,334 Receiver, also see Input buffer
dissipation, 39, 48, 67-68, 77, clock, 195, 199, 210, 216, 255,
90-91, 140, 259-260, 287, 292, 371
264,372,385-386,389 command/address (C/A), 175,187,
Power delivery network, (PDN), 202, 227
372-376, 380, 382 data, 197, 209, 333-334
POWERDOWN, 148 differential, 114, 197-199, 323
POWERUP, 107, 138-139, 148-150 equalization, 249
PRECHARGE, 19,24,45, 105, 204, ESD. See Electrostatic discharge
309
input, 113-114, 196-200, 205,
Predecode logic, 126-134 209,312/334
Predecoding, 60, 128 ODT. See On-die termination
Pre-emphasis, 248-249 offset, 200, 205, 226, 300
Prefetch, 195, 297,300, 326, 330, 332, 限所,113, 152,193, 196-197, 199
346,351
wide-swing, 199
Pumps and generators, 124, 147,
Redistribution layer, 373, 376, 378
158-163
Reduced latency DRAM (RLDRAM),
R 188-190,318
R/W, 3-6, 13-15,329 Redundancy, 60,99-109, 126,
RAS*. See Row address strobe
131-132, 178, 187
RAS* chain, 60, 295
Refresh, 5-6,10, 12-13, 19-20,26,30,
46,48-49, 67, 77,91, 107,
Rate of activation, 49 128-129,387
RDL. See Redistribution layer Register-controlled delay-locked loop
RDLL. See Register-controlled delay (RDLL), 135-142
locked loop Retiming, 193, 313-314, 336-338,
Read cycle time (Frc), 4, 8,12 340-341,363-364
Read latency, 175, 187-188, 330, 350, RLDRAM. See Reduced latency
353-355, 359-367 DRAM
Read-Modify-Write, 13-14, 52-53, Row address strobe (7?^S*), 3, 8-19,
119 29-30,44, 60, 99, 104-105,
Read or Write cycle time (加0), 169, 107-108,118, 120,126-128,132,
301 155-157, 169,295, 301,311-312,
Reads, nonconsecutive, 352 317,319-321,324
Read pipes, 346, 350-359 Row decode, 36, 54, 57, 59, 65, 69,
71-78, 83, 86, 88-97
420 Index

Row hold time (他办 8 (SAR), 226-227, 270

Row redundancy, 102-109, 131 Supply noise, 194, 208, 230, 238, 258,
Row setup time (以8 265, 278, 282-283, 287-291, 337,
371,375,381
Rowline, 6-7, 9-10, 22—23, 30
Symmetrical delay line, 136-141, 261
s Sync point. See Synchronization points
Self-bias, 113-114, 197,278-281 Synchronization in DRAMs, 135-145,
Sense amplifier, 26-29, 36-38, 44-49, 171, 187, 193, 245, 251-289, 346,
69-76, 90-96, 177-182, 205 357
Serializer, 183, 350-351,355-356, Synchronous
358,366 DDR, double data rate, 14, 16-17,
Signal return path shift (SRPS), 31, 124, 135-137, 171,
238-242 180, 185, 193, 221, 249,
Silicide, 11,23,54 252, 269, 296, 303, 322,
332-335,351,359, 367
Simultaneous-switching noise, 194,
219, 230-237, 248 DRAM (SDRAM), 15-20, 135,
170-171, 176, 180, 185,
Skew-tolerant logic, 298, 311 187, 193, 269, 301 -303,
Skin effect, 221, 248 332, 359, 367
Slew rate, 155,196,219-220,222,234, Synchronous mirror delay (SMD), 135,
237, 245, 254, 260, 267 253, 270
SMD. See Synchronous mirror delay Synchronization points (SP), 187,
Source synchronous (SS), 193-195, 363—364
200, 208—210,219,315, 332, 336,
359 T
Speed grade, 170,259, 291, 305-306 tAA. See Access time from column ac­
cess
SRPS. See Signal return path shift
See Access time
SSN, SSO. See Simultaneous-switch­
ing noise tAsc- See Column setup time

SSTL. See Stub series terminated logic Z/SR. See Row setup time
Stability, 47, 51, 118, 154, 165, 253, 打小 See Address to Write delay time
268,270-271,284-285,290 (CAC- See Access time from CAS*
Static column mode, 7,13-14 tcAH- See Column hold time
Stub series terminated logic (SSTL), tccD- See Column cycle time
112-113, 222-223 See Clock cycle time
Subarray, 104, 107, 189-190 t[)QscK- See Read strobe (DQS) access
Subthreshold leakage, 48-49, 58 time from CLK/CLK^
Successive-approximation register tr)QSQ- See DQS-DQ skew
Index 421

S0ss. See Write command to first Z>0S V

latching transition pump, 158-160
Test modes, 125-126, 273 Vccp. 10,13,24,28, 30, 55-58,101,
tpc，See Read or Write cycle time 147, 157-163, 200
Me See Access time from RAS* Vm，See High-input trip point
Tracking Vn，See Low-input trip point voltage
latency, 340, 348, 352, 357, 360, Voltage-controlled delay line (VCDL),
366-374 253, 276-287
timing, 188, 266, 271-274, 285,
308 W
Wordline, 9-13, 22-26, 30-31, 33-35
voltage and temperature (VT), 253,
263,274, 286,291 Write command to first DQS latching
See Row hold time transition, (toQSs)^ 206-207, 252,
287,330,333—334, 341,347
Training, 195,207,209,252,288,292,
Write cycle time W 5
337
Write driver, 28,47, 112,115-118,
Transistor-transistor logic (TTL), 6-7,
306, 329
112
Write latency, 195,206,287,317, 326,
tRC See Read cycle time
330-338, 341,343,345
See Active to Read or Write delay
Write leveling, 207
tRP、302
Writes, nonconsecutive, 342-345
Trench capacitor, 42-44
Write operation waveforms, 28-29
TTL. See Transistor-transistor logic
Write pipes, 330, 346-347
ZffC- See Write cycle time
Write pulse width(/行>),5
Twisting, 36-38,68, 76-77, 85-90
See Write pulse width Y
Yield, 1,34,51, 54, 66, 86, 93, 102,
u 104-105, 125, 200, 243,298, 350
Unit delay, 136-138, 256,259-262,
266-267,271-272, 275,277-281 Z
Unit interval (W), 209-210, 341 ZQ calibration. See Impedance calibra­
tion