Moonsu Kim, Eun yeung Yu, Eunbyeol Kim Ayhan Mutlu
Samsung Electronics Co., Ltd. Synopsys Co., Ltd.
System LSI Division
• Process & Design trends
• Random variation is increasing in advanced process nodes
• Design size and performance gap are also increasing, but TAT limit is almost constant
• Requirement for process variation aware STA
• Advanced node requires higher level of STA accuracy for better PPA within tight TAT limit
• Accurate process variation will remove unnecessary pessimism and optimism at the same time
• Traditional AOCV(Advanced On Chip Variation) and constraint variation modeling has limitations
resulting in inherent pessimism
• Need to use POCV (Parametric OCV) to resolve traditional method’s limitation using advanced
modeling for delay/slew/constraint variation
• Challenge 1: slew/load impact to delay variation
• In Advanced OCV (AOCV), delay variation is function of logic depth and
distance, but does not consider slew/load impact
• In actual design, the magnitude of process variation will increase at higher
slew/load indices on advanced nodes
AOCV Variation is
independent to slew/load
AOCV Variation is
dependent on stage count
OCV_sigma (AOCV)
• Solution: Liberty Variation Format (LVF) for slew/load-based variation
• LVF includes delay variation based on actual slew/load
• The cell delay variation is modeled as a function of input transition and output
load per timing arc
POCV LVF delay variation σ = f (slew,load)
X1 X2 X3
Propagated path
X1+ X2+X3 arrival time distribution
EXAMPLE of LVF delay variation
ocv_sigma_cell_rise ("delay_template_4x4") {
sigma_type : "late";
index_1("0.001, 0.005, 0.010, 1.00");
index_2("0.001, 0.002, 0.003, 0.10");
values( "σ11, σ12, σ13, σ14", \
"σ21, σ22, σ23, σ24", \
Slew
"σ31, σ32, σ33, σ34", \
"σ41, σ42, σ43, σ44", );
} Load
• Samsung experiment results
• Extracted 300 paths from real design to compare arrival times using
PrimeTime POCV with MonteCarlo (MC) reference
• POCV has good correlation with MC using slew/load dependent model
• AOCV has large pessimism for some cases and therefore larger error spread
than POCV
POCV AOCV
error error
average -0.16% -0.85%
stdev 1.02% 7.45%
AOCV(green) : lumped slew/load model
POCV(red) : slew/load aware variation model
Monte Carlo(blue): reference
• Challenge 2: Common point optimism reduction for Min Pulse
Width(MPW) and half-cycle path
• Different transition at common point need to keep process variation at
common point
• In AOCV, process variation can’t be exactly calculated by statistical sum of
rising and falling paths and it can’t be separated with other variation models
Process variation will be removed from slack computation
cell1 cell2 A B C
A
B C ------------------------------
Launch clock : F R F
Capture clock: R F R
AOCV-based variation at point C Actual process variation at Point C
v.s. = sqrt( σ
cell1,rise +σcell2,fall +σcell1,fall +σcell2,rise
= max_derate – min_derate (CRPR) ^2 ^2 ^2 ^2 )
= 0 (max_derate = min_derate)
AOCV-based variation can be optimistic as
it may be removed by CRPR
• Solution: Accurate process variation for MPW using POCV
• Using POCV, accurate process variation for MPW and half cycle’s slack
computation will remove optimism
• Samsung experiment results
• For extracted 10K paths from real design, average 3.3% optimism reduction
can be reflected into MPW slack (or half cycle path) computation by POCV
model
cell1 cell2 POCV process variation at Point C (same as actual)
B C = sqrt( σcell1,rise^2+σcell2,fall^2+σcell1,fall^2+σcell2,rise^2 )
A
A B C
------------------------------ avg stdev
Launch clock : F R F
Capture clock: R F R MPW Delta -3.32% 0.79%
< MPW Delta: Optimism Reduction>
Delta = (new_pw – old_pw) / old_pw
new_pw: pulse width considering process variation at common point using POCV
old_pw: pulse width not considering process variation (traditional)
• Challenge 3: Delay variation is highly correlated with output slew
variation
• Output slew variation will affect delay variation in next stage
• Output slew variation has an increased impact at lower voltage
MonteCarlo Simulation for
Inverter Cell
? Dσcell2
Highly correlated
Dσcell1 ? Sσcell2 X axis: cell delay
Sσcell1 Y axis: output slew
Highly correlated
Sin,cell2
cell1
^^^^ ρ = 0.78
cell2
Dσcell1: cell1’s delay variation
Sσcell1 : cell1’s output slew variation
S in,cell1 : cell2’s input slew
• Solution: POCV considers correlation between slew and delay variation
• Better accuracy for lower voltages by reflecting correlation between slew and
delay variation in STA
• Samsung experiment results
• About 1% accuracy improvement by correlated slew variation modeling for
150 extracted paths
<POCV> Correlated delay and slew variation
Dσcell2
Dσcell1 Dσcell2(Sσcell1) Dσcell1: cell1’s delay variation w/o input slew variation
Sσcell1 Sσcell2 Sσcell1 : cell1’s output slew variation
Sσcell2 (Sσcell1) Dσcell2 : cell2’s delay variation w/o input slew variation
Dσcell2(Sσcell1) : cell2’s delay variation induced by input slew variation
Sσcell2 : cell2’s output slew variation w/o input slew variation
^^^^ Sσcell2 (Sσcell1) : cell2’s output slew variation induced by input slew variation
cell1 cell2
Delay variation only Delay variation + slew Delay variation + slew
variation (no correlation) variation (correlated)
Error(%) Mean =1.55% Mean =0.84% Mean = 0.56%
= (STA –MC)/MC Stdev = 0.70% Stdev = 0.56% Stdev = 0.55%
• Challenge 4: Traditionally, Flop’s constraint variation is added to nominal
constraint value as margin
Delay variation and constraint variation in STA becomes linear sum
• However, delay variation and constraint variation is independent and not
correlated
Delay variation and constraint variation should be statistical sum
• By separating constraint variation with nominal constraint value
theoretically maximum 40% pessimism reduction
A) Traditional linear sum: total variation = σpath + σH
σL B) POCV Statistical sum: total variation = SQRT ( σpath^2 + σH^2)
σH where σpath = SQRT ( σL^2 + σC^2)
Max difference between linear and statistical sum is when σpath= σH
Max { (A–B) / B } = (2 - 1.414) / 1.414 = ~40%
This is the pessimism from linear sum method
σC
By adopting separate constraint variation model, maximum 40%
pessimism of slack variation can be reduced
• Solution: POCV Support of Constraint Variation
• Uses statistical sum of delay and constraint variation
• Theoretical maximum of 40% pessimism reduction than traditional method
• Samsung experiment results
• Extracted 1000 paths from real design to evaluate real world pessimism
reduction when using separate modeling of FF constraint variation
• Linear sum pessimism = (linear sum slack – statistical sum slack)/linear sum
slack
Max ~40% pessimism reduction
• Majority of paths achieve close to 40%
avg stdev
pessimism reduction in total variation Linearsum
-0.37 0.04
Pessimism
• Average 37% pessimism reduction of
1000 paths
• Review potentially opportunities to improve accuracy challenges caused
by limitation of traditional variation (AOCV) and review solutions using
advanced method of PrimeTime POCV
• Slew/load aware delay variation modeling
- Accuracy improvement more opportunity of both robust and optimized design based on PBA
accuracy enhancement
• Common point optimism removal for Half-cycle and Min Pulse Width
- Average 3.3% optimism reduction robust design for MPW and half cycle path
• Correlated slew variation impact on accuracy
- Average 1% mean accuracy improvement by correlated modeling of delay and slew variation
• Statistical constraint variation modeling
- Average 37% pessimism reduction compared to traditional linear sum of delay variation and
constraint variation
• POCV based accurate analysis is necessary to meet tight PPA
requirement with limited resources for advanced nodes
• Results shown in this presentation based on Samsung Advanced node
design using Synopsys PrimeTime POCV Analysis
