© C. Visweswariah, 2002, do not use without permission 1
© C. Visweswariah, 2002, do not use without permission 1
© C. Visweswariah, 2002, do not use without permission 2
?
© C. Visweswariah, 2002, do not use without permission 3
?
© C. Visweswariah, 2002, do not use without permission 4
?
© C. Visweswariah, 2002, do not use without permission 5
!
© C. Visweswariah, 2002, do not use without permission 6
Death, Taxes and Failing Chips
Chandu VisweswariahIBM Thomas J. Watson Research Center
Yorktown Heights, NYTAU ’02
with thanks to my colleagues andcollaborators at IBM Fishkill,
IBM Burlington, IBM Yorktownand TU/Eindhoven
© C. Visweswariah, 2002, do not use without permission 7
Outline• “The era of probabilistic design”• Three aspects of the problem
– modeling– methodology– analysis + synthesis
• Characteristics of a good statistical timer• Our analysis efforts
© C. Visweswariah, 2002, do not use without permission 8
Catastrophic vs. parametric
Chip behavior in the face of environmentaland manufacturing variations
Chip behavior in the face of environmentaland manufacturing variations
Catastrophicyield loss
Catastrophicyield loss
Parametric or“circuit-limited”
yield loss
Parametric or“circuit-limited”
yield loss
Critical areaVoronoi diagrams
Redundant via insertionWire bending/spacing
Statistical timingYield predictionDesign centering
Design for manufacturability
DigitalASICs
© C. Visweswariah, 2002, do not use without permission 9
Bounded vs. probabilistic analysis
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1Slack (ns)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Yie
ld
Varying temperature, mean = 25, sigma = 25
BC NOMWC[Data courtesy K. Kalafala]
© C. Visweswariah, 2002, do not use without permission 10
Bounded vs. probabilistic analysis
-50 -25 0 25 50 75 100Temperature (Celsius)
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
Wor
st la
te s
lack
(ns)
PI->
Latc
h
Latch -> Gating
[Data courtesy K. Kalafala]
© C. Visweswariah, 2002, do not use without permission 11
Bounded vs. probabilistic analysis
- 0 .6
- 0 .5
- 0 .4
- 0 .3
- 0 .2
- 0 .1
0
0 .1
0 .2
0 .3
S la c k
Te m p e ra tu reVolta
ge
[Data courtesy K. Kalafala]
© C. Visweswariah, 2002, do not use without permission 12
Bounded vs. probabilistic analysis
Yield
Slack
© C. Visweswariah, 2002, do not use without permission 13
The era of probabilistic design
[T. Karnik, S. Borkar, V. De, ICCAD 2002]
© C. Visweswariah, 2002, do not use without permission 14
Statistical timer
Statictimer
Delay andslew models
Netlist+
assertions
1. Slack2. DiagnosticsStatistics of
the sourcesof variability
Dependenceon sources
of variability
1. Yield curve2. Diagnostics
Statisticaltimer
© C. Visweswariah, 2002, do not use without permission 15
The full picture
Mod
eling
Methodology
Analysis
© C. Visweswariah, 2002, do not use without permission 16
Methodology
Processmodels
Transistormodels
Gate delaymodels
Statictiming
Statisticalprocessmodels
Statisticaltransistormodels
Statisticalgate delay
models
Statisticalstatictiming
© C. Visweswariah, 2002, do not use without permission 17
Methodology issuesASIC Microprocessor
No at-speed test,often no AC test
Sorted
Large, flat Hierarchical
Library-based Custom circuits andlibrary-based
Focus on worst-casetiming
Focus on nominal(and best case!)timing
© C. Visweswariah, 2002, do not use without permission 18
Definition of yield• Risk management
– with PSROs (Performance-Sensitive RingOscillators) and appropriate sign-off criteria
– at multiple levels– with/without AC or at-speed test
• Environmental vs. manufacturing variations– require 100% yield in environmental window– guaranteed 100% yield in the manufacturing
window is overkill
© C. Visweswariah, 2002, do not use without permission 19
Other methodology implications• Number of timing runs is excessive
– early and late mode– LCD (Linear Combination of Delay) or “interval
delay” to model ACLV– CPPR (Common Path Pessimism Removal)– NBTI (Negative Bias Temperature Instability)– BEOL variations– coupling noise
• What is required is a reduction in the numberof timing runs while phasing new analysismodes into the current methodology!
Opportunity!
© C. Visweswariah, 2002, do not use without permission 20
Design methods• Examples:
– adaptivebody bias
– mixing oflogicfamilies
© C. Visweswariah, 2002, do not use without permission 21
Modeling• What are the sources of variation that really
matter?– mathematical vs. empirical answers
• What are the means, deviations andcorrelations of the sources of variation?
• What is the dependence of the delay andslew of each edge of the timing graph toeach source of variation? Is this computedduring the library characterization?
• What about custom circuits?
© C. Visweswariah, 2002, do not use without permission 22
Analysis wish list
© C. Visweswariah, 2002, do not use without permission 23
Number 1: path sharing
© C. Visweswariah, 2002, do not use without permission 24
Number 2: clock correlation
© C. Visweswariah, 2002, do not use without permission 25
Number 2: clock correlation• Importance of correlations
– consider a circuit with 50K latches, each with asetup and hold test, each of which has a 99.99%probability of being met
– if all tests are perfectly correlated,yield=99.99%
– if all tests are perfectly independent, yield is0.005%
– the truth is closer to the perfectly correlatedcase!
© C. Visweswariah, 2002, do not use without permission 26
Number 3: global correlation
L3 D
irec
tory
/Con
trol
L 2 L 2 L 2
L S U L S UIF UB X U
ID U ID U
IF UB X U
F P U F P U
FXU
FXUIS U IS U
© C. Visweswariah, 2002, do not use without permission 27
Number 3: global correlation
© C. Visweswariah, 2002, do not use without permission 28
Number 4: bounded vs. statistical• Bounded
– input vectors– environmental variables– PLL jitter
• Statistical– manufacturing parameters– coupling noise?
• Should be easy to switch betweencolumns
• Large vs. small number of randomvariables
© C. Visweswariah, 2002, do not use without permission 29
Number 5: slew/load dependence
© C. Visweswariah, 2002, do not use without permission 30
Number 6: deterministic vs.random ACV
[From M. Orshansky, L. Milor, P. Chen, K. Keutzer, C. Hu, ICCAD 2000]
© C. Visweswariah, 2002, do not use without permission 31
Number 7: the tail matters!
• Avoid pessimism• Capture correlations
© C. Visweswariah, 2002, do not use without permission 32
Numbers 8 and 9• Number 8
– fit well with rest of existing methodology– reduce number of timing runs required
• Number 9– provide diagnostics
© C. Visweswariah, 2002, do not use without permission 33
Number 10: flexibility
Quick and dirty Slow and accurate
For optimization For sign-off
Incremental Not incremental
Usually block-based,performance-spacemethods
Usually path-based,parameter-spacemethods
© C. Visweswariah, 2002, do not use without permission 34
Feasible region
Performance-space vs.parameter-space
oxt
effL
JPDF ofglobal
parameters
© C. Visweswariah, 2002, do not use without permission 35
Our analysis efforts• Three slides deleted• See DAC ’03 submission for details• Example: reduced run time from 68 hours
for repeated EinsTimer runs on a 200K gateASIC to about 15 minutes
© C. Visweswariah, 2002, do not use without permission 36
Conclusions• Brave old world of probabilistic design• Statistical considerations must influence all
stages of design• Comprehensive solution required
encompassing methodology, modeling,analysis, synthesis, test, design methods
• The computation will not prove to be thehard part; if nothing else, Monte Carlo withintelligent sampling will come to the rescue