Title: Timingaware power noise reduction using prediction and correction
1Timing-aware power noise reduction using
prediction and correction
- Malgorzata Marek-Sadowska
- University of California, Santa Barbara
Chao-Yang Yeh Apache Design Solutions
2Outline
- Motivation
- Previous work
- Background
- Prediction
- Neighborhood-aware decap allocation
- Correction
- Gate sizing for power noise and timing
- Experiments
- Conclusion
3Motivation
- Power noise is a serious problem
20
16.7
0.3V
0.3V
0V
1.5V
1.8V
0V
- Power noise affects timing and even function
Cell delay
Voltage drop
- Decap is useful to reduce power noise
- Decap location has to be selected wisely
4Decap allocation problem
- How to allocate decaps among individual blocks of
a floorplan?
Decap
5Previous work
- Optimization during placement
- increases placement complexity
- lacks a global view at the early placement
iterations - Post-layout decap re-allocation
- Large decap re-allocation unattractive
Decap
Previous decap-allocation works do not consider
timing
6Proposed flow prediction and correction
Decap padding
Prediction
Cell decap padding assignment
( Reduce decap re-allocation )
Cell
Decap
Neighborhood Prediction
Placement Grid analysis
Gate sizing
Correction
Gate sizing for power noise and timing
( Use pad location/ grid information )
7Power grid/decap modeling (I)
- Regular power grid/pad distribution
- Uniformly partition of the chip
One partition
8Power grid/decap modeling (II)
Cell decap
Inserted decap
current
te
1ns
ts
time
9The prediction step
Decap padding
Decap_weight(n)
- Compute neighborhood current consumption (NCC)
n
Noise_weight(n)
( More noisy, more padding )
neighborhood
Timing_weight(n)
( More critical, less padding )
10Wirelength Prediction background Mutual
contraction
B. Ho, M. M. Sadowska, DAC03
- Based on circuit topology analysis
- Compute contraction weight
- Only predict distance between cells that have
connections
Mutual contraction MC(x, y) Wr(x, y) Wr(y, x)
11Mutual contraction in different placers
- Works for general wirelength-driven placers
Dragon
FengShui
12Neighborhood levels
n
0-th level Neighbors
1-st level Neighbors
(i1)-th level Neighbors
Cells with short connections to
Higher level neighbors, larger neighborhood area
13Neighborhood current consumption (NCC)
- Higher level NCC, more cells involved
- 0-th level NCC
- (Individual cell current consumption)
14Neighborhood current consumption (NCC)
Neighborhood levels
Level NCC computation
2-nd
c
11
1-st
b
5
9
0-th
n
d
10
a
8
e
2
8
2
5
8
9
10
8
9
Higher level NCC -gt Less highly switching cells
15Noise/timing weight
Decap_weight noise_weight timing_weight
Noise/timing weight in the range 01
16Decap allocation
Timing scale
Total decap area
17Prediction experiment (I)
Benchmark ex1010
(b)
(a)
NCC Level 1
NCC Level 0
55 cells
NCC Level 4
NCC Level 2
(d)
(c)
Decap will be allocated to those highly switching
cells
18Prediction experiment (II)
- Simulation results (Vdd 1.8v)
Decap EVEN
Decap WGT, NCC level 4
(a)
(b)
Min V1.64
Min V1.72
19Correction step gate sizing
- After placement and grid simulation
- Do gate sizing
- Re-allocate decaps
- Meet timing constraints
- Sequence of linear programs (SLP)
- Solve a linear programming (LP) problem in each
iteration
20Linear programming (LP) problem formulation
LP constraints
Divide chip into blocks
Timing_const
Minimize
Power consumption cost
Noise_const
Area_const
Noise cost
Noise cost factor
21LP timing constraints
- Using gain-based delay model
Cell power voltage
Size of cell k
22LP timing constraints (cont)
23Experimental setup
- Ideal chip voltage 1.8V
- MCNC benchmarks (cell 419923362)
- 0.18um technology
- Decap area 20 of total cell area
- 5 Vdd voltage noise margin
- Linux Pentium 2.4GHz
24Experiment flow
Prediction Allocate decap
Correction Gate sizing
Placement
Grid analysis
Grid analysis
Results after prediction optimization
Results after correction optimization
25Power noise measurement metrics
- 5 voltage drop margin
- Metrics
- Worst voltage drop (IRV)
- Violation count (vioC)
- Sum of excess noise area (SENA)
Voltage
Vdd
95 Vdd
Excess noise area
time
26Experimental results prediction
- Results after prediction optimization
NOC no decap, EVEN even decap distribution,
WGT predictive method
27Experimental results correction
- Results after correction optimization
28Experimental results voltage drop
- TS0, optimize power noise only
29Power noise reduction conclusion
- Proposed a flow for power noise reduction
- Timing-aware power noise reduction
- Prediction - correction flow
- Prediction neighborhood prediction
- Correction gate sizing for noise and timing
- Proposed techniques are effective and efficient
- Effective even without decap insertion
30Experimental results prediction (cont)
- Timing results after optimization