Confidential

HBM: Memory Solution for

Bandwidth-Hungry Processors
Joonyoung Kim and Younsu Kim
SK hynix Inc.
August 2014

2014 SK hynix Inc.

This material is proprietary of SK hynix Inc. and subject to change without notice.

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Memory requirement
HBM: Memory Solution for Density & Bandwidth-Hungry Processors
< Exa-scale Roadmap >
High-End Graphics

40G/100G Ethernet

Exa-scale HPC
Source : SciDAC,
www.scidacreview.org

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Memory bottleneck & solution - Speed, Density, Power & SFF

TSV is a revolutionary technology for overcoming the bottleneck

High Bandwidth

High Capacity

solution beyond

solution to overcome DRAM

DDR4/GDDR5/LPDDR4

Scaling Limit

Power Efficiency

Small Form Factor

solution to meet

solution to meet

system power budget

Size & Cost limitation

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

TSV(Through Silicon via)

TSV is the technology of 3D Stack
(High Density / Small size PKG / High speed)

Die Pad

Die Pad
Bond wire

TSV

PKG
Pad

PKG
Pad

Die #2
Die #1

< Wire bonding PKG >

< TSV PKG >

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Bottleneck 1) Bandwidth

DDR3

TSV(HBM)

Config.
soc

IO

64 DQ

1024 DQ

Speed

1.6G bps

1~2Gbps

Bandwidth
Compare

64 Gbps
Long Length

12.8GBps

1024 Gbps

RLC increase

Short Length

Max 256GBps
RLC decrease

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Bottleneck 2) Technology Limit

1. Capacity Limit

2. Scale Down Limit

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Bottleneck 3) Small Form Factor

Wire bonding - DDR3

TSV - HBM

Image
http://www.shmj.or.jp/english/
packaging/pac90s.html

PKG Size@die

100% (117mm2)

36% (42mm2)

mm2@128GB/s

100% (3744mm2)

11% (42mm2)

Power
Consumption*
@128GB/s

100% (6.4W)

51% (3.3W)

* Power Cal = IMPT

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

Bottleneck 4) Low Power

Lower Speed/pin and x1024 Wide IO

Lower Cio(0.6pF), No Termination

low power consumption per Pin.

small IO current consumption

Power consumption is decreased by 42% compared with GDDR5

Ratio
[mW/Gbps/Pin]

1.0

42%
0.63
0.55

0.32
DDR3 x16

DDR4 x16

GDDR5 x32

HBM

I/O Power Efficiency@IDD4R

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

HBM Overall Architecture

4 Core DRAM + 1 Base logic die (Chip on Wafer)

Items

Target

# of Stack

4(Core) + 1(Base)

Ch./Slice

Total Ch. for KGSD

IO/Ch.

128

Total IO/KGSD

1024(=128 x 8)

Address/CMD

Dual CMD

Data Rate

DDR

[1] D.U Lee, SK hynix, ISSCC 2014

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

HBM, What are the differences?

KGSD* Memory supporting for System in package
HBM in 2.5D SiP

FBGA

Side Molding

Side Molding

DRAM Slice
DRAM Slice
DRAM Slice
DRAM Slice
DA ball

TSV

PHY

SoC
PHY

Interposer

KGSD
Base Die
DFT Area

TSV Area

DA BALL

TSV OVERLAP

(TEST Logic)

ADD/CMD,
DQ (RX/TX),

(Direct Access)
DFT

PHY

DRAM POWER
Supply

SIGNAL
Connectivity
Test

* KGSD (Known Good Stacked Die)

10

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

HBM Market Segments

HBM market will scale-out to various segments
(Over 21 Design-Ins In Progress)

Low
Power
Capacity
& BW
Improve
RAS
POST
GDDR5
128GBps BW
JEDEC std

Density flexibility
>256GBps BW

VDD=1.XV
Good I/O power eff.
Digital Consumer,
Automotive, etc.

ECC
Post-package Repair
MBIST

205.132.242.85 / 2014. 07. 18 16 : 50 / B34047 / 2057897

HBM Overall specification

1nd Gen HBM
2Gb per DRAM die
1Gbps speed /pin
128GB/s Bandwidth
4 Hi Stack (1GB)
x1024 IO
1.2V VDD
Base Die

KGSD w/ Bump

2nd

Gen HBM

Interposer

8Gb per DRAM die

2Gbps speed/pin
256GBps Bandwidth/Stack
4/8 Hi Stack (4GB/8GB)

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Gen1 5mKGSD Structure

5mKGSD (molded Known Good Stacked Die)
- 1 Base + 4 Core (DRAM) with Side Mold -

<Cross Section View>

490mm
25um

5.48mm 25um

<Top View>

7.29mm 25um

Exposed Silicon

(a), (b), (c), (d)

<Bottom View>

Item

Value

(a)

uBump
Diameter

25um

(b)

uBump
Height

( 3.5 um)

(c)

uBump Pitch

55um

(d)

uBump Array
(MPGA)

JEDEC

Remark

( 3 um)

35um

Cu/Ni/SnAG
(17/3/15um)

JC11-2.883
JC11-4.884

Side Mold(190um)
205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

Comparison of HBM and other DRAMs

Item

DDR3 (x8)

GDDR5 (x32)

4-Hi HBM (x1024)

I/O

32

1024

Prefetch (Per IO)

8Byte

32Byte

256Byte

2GB/s

28GB/s

128~256GB/s

40~48ns

40ns(=1.6v, 1.5v)
48ns(=1.35v)

40~48ns

tCCD

4ns (=4tCK)

2ns (=4tCK)

2ns (=1tCK)

VPP

Internal VPP

Internal VPP,
(Opt. Ext. VPP)

Ext. VPP

VDD

1.5, 1.35

1.6, 1.5, 1.35

1.2

Single CMD

Single CMD

Dual CMD

Refresh Single Bank

DBI mode

O (DBI_DC)

O (DBI_AC)

Access Granularity
(=I/O x Prefetch)

Max. Bandwidth
tRC

CMD Input

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

Dual CMD interface

CMD efficiency increased by Semi-independent row/column input

Row/column input through different pins

RAS

CMD

Conventional drams share RAS/CAS CMD.

ADD

RA, CA

Bank

RA, CA

CMD

RAS/CAS/WE/CS

decoder

CMD
CAS

decoder

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

REF Single Bank

Single bank refresh and programmable tRAS

Concurrent read/write operation with single bank refresh allows data bus to remain active.

Refresh

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

REF Single Bank

Command bus efficiency can be maximized

- Dual Command & REF single bank -

CLK

T+1

T+2

T+3

BANK0

ACT

tRCD

BANK1

tRRD

ACT

T+5

T+6

T+7

WT

T+8
PCG

WT

BANK2

PCG

BANK3

WT

BANK4

tRRD

RD
WT

BANK5

RD

BANK6
BANK7

T+4

REFSB
RD

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Core Architecture

HBM single die has 2channels
1 channel consists of 128 TSV I/O with 2n prefetch
CH-Left
B0
YCTRL

B2
DWORD 0
32 I/O

B6

B0

YCTRL

YCTRL

B3

B2

DWORD 1
32 I/O

B4
YCTRL

B1

AWORD

CH-Right

DWORD 2
32 I/O

B5

B4

YCTRL

YCTRL

B7

B6

B1

B0

YCTRL

YCTRL

B3

B2

DWORD 3
32 I/O

DWORD 0
32 I/O

B5

B4

YCTRL

YCTRL

B7

B6

B1

B0

YCTRL

YCTRL

B3

B2

DWORD 1
32 I/O

B5
C

YCTRL

B7

AWORD

B1
C

YCTRL

B3

DWORD 2
32 I/O

DWORD 3
32 I/O

1 bank
B4 : 2 sub-banks(64Mb)
B5
non-shared I/O between
sub-banks
YCTRL

B6

YCTRL

B7

[2] D.U Lee, SK hynix, ISSCC 2014

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Base Die Architecture

Base die consists of 3 Areas PHY, TSV, Test Port Area

HBM ballout area

6,050x3,264 m
[3] D.U Lee, SK hynix, ISSCC 2014

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

Base Die Customization Future HBM Concept

Logic Layer

Host I/F + Memory I/F + Base Logic/IP Block

Overcome
Memory Scaling
Memory Die

Customization to
meet various
requirements

- Timing
- Refresh

Parallel-to-Serial(P2S)/S2P
JTAG, PMBIST
Configuration Registers
Error Handling

20

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Thermal Management

Temp. [[]

Thermal dummy bumps as well as well-designed device architecture are helpful for
thermal dissipation No mechanical reliability issues by thermal dummy bumps.

Temp. Saturation

TSV

Thermal
Dummy Bumps

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Long-term Roadmap(2) (Preliminary)

HBM product longevity is critical in several applications
SK hynix plans to address longevity requirement
2014

(1)

2015

2016

2017

2018

2019

2020

2021

2022

2023

Bandwidth: starts at 128GB/s & plans for higher

Note 1 anticipated future HBM density

Note 2 roadmap is subject to changes without prior notifications

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

HBM Summary
Perfect memory solution for various application requirement

High Bandwidth

High Density PKG

~256GB/s

Up to 8GB

HBM
Smaller
Form Factor

-65%

1 ~ 4 Cubes
per GPU

4 ~ 8 Stacks
KGSD

1)

Good Power
Efficiency

68%

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897

Thank You !

205.132.242.85 / 2014. 07. 18 16 : 51 / B34047 / 2057897