VLSI Physical Design Automation
Clock and Power Routing
Prof. David Pan
[email protected]Office: ACES 5.434
1
Routing of Clock and Power Nets
Different from other signal nets, clock and power are
special routing problems
For clock nets, need to consider clock skew as well as delay.
For power nets, need to consider current density (IR drop)
=> specialized routers for these nets.
Automatic tools for ASICs
Often manually routed and optimized for
microprocessors, with help from automatic tools
Clock Introduction
For synchronized designs, data transfer between
functional elements are synchronized by clock signals
Clock signal are generated externally (e.g., by PLL)
Clock period equation
clock peri od td tskew tsu tds
t d:
tskew:
tsu:
tds:
Longest path through combinational logic
Clock skew
Setup time of the synchronizing elements
Propagation delay within the synchronizing element
3
Clock Skew
Clock skew is the maximum difference in the arrival
time of a clock signal at two different components.
Clock skew forces designers to use a large time period
between clock pulses. This makes the system slower.
So, in addition to other objectives, clock skew should
be minimized during clock routing.
Clock Design Problem
What are the main concerns for clock design?
Skew
No. 1 concern for clock networks
For increased clock frequency, skew may contribute over 10% of
the system cycle time
Power
very important, as clock is a major power consumer!
It switches at every clock cycle!
Noise
Clock is often a very strong aggressor
May need shielding
Delay
Not really important
But slew rate is important (sharp transition)
5
The Clock Routing Problem
Given a source and n sinks.
Connect all sinks to the source by an interconnect
network (tree or non-tree) so as to minimize:
Clock Skew = maxi,j |ti - tj|
Delay = maxi ti
Total wirelength
Noise and coupling effect
Clock Design Considerations
Clock signal is global in nature, so clock nets are
usually very big
Significant interconnect capacitance and resistance
So what are the techniques?
Routing
Clock tree versus clock mesh (non-tree or grid)
Balance skew and total wire length
Buffer insertion // will be covered in EE382V (Optimization
issues in VLSI CAD)
Clock buffers to reduce clock skew, delay, and distortion in waveform.
Wire sizing // will be covered in Opt. Issues in VLSI CAD
To further tune the clock tree/mesh
7
Clock trees
A path from the clock source to clock sinks
Clock Source
FF
FF
FF
FF
FF
FF
FF
FF
FF
FF
Clock trees
A path from the clock source to clock sinks
Clock Source
FF
FF
FF
FF
FF
FF
FF
FF
FF
FF
9
H-Tree Clock Routing
Tapping Point
4 Points
16 Points
10
H-tree Algorithm
Minimize skew by making interconnections to subunits
equal in length
Regular pattern
The skew is 0 assuming delay is directly proportional to
wirelength
Is this always the case???
Can be used when terminals are evenly distributed
However, this is never the case in practice (due to blockage,
and so on)
So strict (pure) H-trees are rarely used
However, still popular for top-level clock network design
Cons: too costly to be used everywhere
Can you think of another shape if non-rectilinear wires are allowed?
11
Method of Means and Medians (MMM)
Applicable when the clock terminals are arbitrarily
arranged.
Follows a strategy very similar to H-Tree.
Recursively partition the terminals into two sets of equal
size (median). Then, connect the center of mass of the
whole circuit to the centers of mass of the two sub-circuits
(mean).
Clock skew is only minimized heuristically. The resulting
tree may not have zero-skew.
12
An Example of MMM
centers of mass
13
Geometric Matching Algorithm
(GMA)
MMM is a top-down algorithm, but GMA is a bottom-up
algorithm.
Geometric matching of n endpoints:
Construct a set of n/2 line segments connecting n endpoints
pairwise.
No two line segments share an endpoint.
The cost is the sum of the edge lengths.
The basic idea is to find a minimum cost geometric
matching recursively.
Time complexity is O(n2.5 log n) for n endpoints.
14
An Example of GMA
Tapping point
(not necessarily
the mid-point)
H-flipping
Post-processing
Apply geometric
matching recursively.
Can give clock tree of zero skew.
15
An Exact Zero Skew Algorithm
ICCAD 1991 and TCAD 1993, Ren-Song Tsay
A classic paper to manage clock skew
Use Elmore delay model to compute delay
Guarantee zero skew
Can easily to extended for zero skew or bounded skew
Can you think of a method to do it?
Try to minimize wire length, but not done very well
Lots of follow up works to minimize total wire length while
maintaining zero skew
DME and its extensions
16
Deferred Merge Embedding
As its name implies, DME defers the merging as late
as possible, to make sure minimal wire length cost for
merging
Independently proposed by several groups
Edahiro, NEC Res Dev, 1991
Chao et al, DAC92
Boese and Kahng, ASIC92
DME needs an abstract routing topology as the input
It has a bottom-up phase followed by a top-down
process (sounds familiar?)
17
DME:
18
Some Thoughts/Trend
Clock skew scheduling together with clock tree
synthesis
Schedule the timing slack of a circuit to the individual
registers for optimal performance and as a second
criteria to increase the robustness of the
implementation w.r.t. process variation.
Variability is a major nanometer concern
Non-tree clock networks for variation-tolerance
How to analyze it?
The task is to investigate a combined optimization
such that clock skew variability is reduced with
minimum wirelength penalty
19
Non-tree: Spine & Mesh
Spines
Clock sinks or local sub[Su et. al, ICCAD01]
networks
Clock sinks or local subnetworksin Pentium processor
Applied
[Kurd et. al. JSSC01]
Applied in IBM microprocessor
Very effective, huge wire
Clock sinks or local subnetworks
[Restle et. al, JSSC01]
20
Non-tree: Link Perspective
Non-tree = tree + links
How to select link pairs is the key problem
Link = link_capacitors + link_resistor
Key issue: find the best links that can help the skew variation
reduction the most!
u
C/2
Rl
w
u
C/2
Rl
C/2
C/2
w
[Rajaram et al, DAC04]
21
Power Distribution/Routing
22
Power Distribution
Power Distribution Network functions
Carry current from pads to transistors on chip
Maintain stable voltage with low noise
Provide average and peak power demands
Provide current return paths for signals
Avoid electromigration & self-heating wearout
Consume little chip area and wire
Easy to lay out
23
Power and Ground Routing
Each standard cell or macro has power and ground
signals, i.e., Vdd (power) and GND (ground)
They need to be connected as well
You can imagine that they are HUGE NETWORKS!
In general, P/G routings are pretty regular
They have high priority as well
P/G routing resources are usually reserved
When you do global and detailed routing for signal nets, you
cannot use up all the routing resources at each metal layers
Normally some design rules will be given (e.g., 40% of top metal
layers are reserved for P/G)
24
P/G Routing Main Objectives
Routing resource
Need to balance the routing resource for P/G, clock and signals
Voltage drop
Static (IR) and dynamic (L di/dt) voltage drops
More voltage drop means more gate delay
Usually less than 5-10% voltage drop is allowed
So you may need to size P/G wires accordingly
Electrical migration
Too big current may cause EMI problem
Others
25
P/G Mesh (Grid Distribution)
Power/Ground mesh will allow multiple paths from P/G
sources to destinations
Less series resistance
Hierarchical power and ground meshes from upper metal
layers to lower metal layers
All the way to M1 or M2 (stand cells)
Connection of lower layer layout/cells to the grid is through
vias
26
Using One Metal Layer
One tree for VDD and another tree for GND.
VDD
GND
27
Using Two Metal Layers
One 2D-grid for VDD and another one for GND:
VDD
GND
M5
M4
28
Gate Array & Standard Cell Design
Inter-weaved combs:
VDD
GND
29
Some Thoughts/Trends
P/G I/O pad co-optimization with classic
physical design
Decoupling capacitor can reduce P/G related
voltage drop
Need to be planned together with floorplanning and
placement
Multiple voltage/frequency islands make the
P/G problem and clock distributions more
challenging
30
