1 Tau 2002
Explicit Computation of Performance as a Function of Process Parameters
Lou Scheffer
2 Tau 2002
What’s the problem?
• Chip manufacturing not perfect, so….
• Each chip is different
• Designers want as many chips as possible to work
• We consider 3 kinds of variation
‑ Inter-chip
‑ Intra-chip
‣Deterministic
‣Statistical
3 Tau 2002
Intra-chip Deterministic Variation (not considered further in this presentation)
• Optical Proximity Effects
• Metal Density Effects
• Center vs corner focus
You draw this: You get this:
4 Tau 2002
Inter-chip variation
• Many of the sources of variation affect all objects on the same layer of the same chip.
• Examples:
‑ Metal or dielectric layers might be thicker/thinner
‑ Each exposure could be over/under exposed
‑ Each layer could be over/under etched
5 Tau 2002
Interconnect variation
• Looking at chip cross section
• Pitch is well controlled, so spacing is not independent
These dimensions can vary indpendently
Pitch is well controlled
P1
P2
P3
P4
P0
P5
Width and spacing are not independent
6 Tau 2002
Intra-chip statistical variation
• Even within a single chip, not all parameters track:
‑ Gradients
‑ Non-flat wafers
‑ Statistical variation
‣Particularly apparent for transistors and therefore gates
‣Small devices increase the role of variation in number of dopant atoms and L
• Analog designers have coped with this for years
• Mismatch is statistical and a function of distance between two figures.
7 Tau 2002
Previous Approaches
• Worst case corners – all parameters set to 3
‑ Does not handle intra-chip variation at all
• 6 corner analysis
‑ Classify each signal/gate as clock or data
‑ Cases: both clock and data maximally slow, clock maximally slow and data almost as slow, etc.
• Problems with these approaches
‑ Too pessimistic: very unlikely to get 3 on all parameters
‑ Not pessimistic enough: doesn’t handle fast M1, slow M2
8 Tau 2002
Parasitics, net delays, path delays are f(P)
• CNET= f(P0, P1, P2, …)
• DELAYNET= f(P0, P1, P2, …)
Pitch is well controlled
P1
P2
P3
P4
P0
P5
9 Tau 2002
Keeping derivatives
• We represent a value as a Taylor series
• Where the di describe how the value varies with a change in process parameter pi
• Where pi itself has 2 parts pi = Gi + si,d
‑ Gi is global (chip-wide variation)
‑ si,d is the statistical variation of this value
N
iPdDD1
00
10 Tau 2002
Design constraints map to fuzzy hyperplanes
• The difference between data and clock must be less than the cycle time:
• Which defines a fuzzy hyperplane in process space
Global Statistical
(Hyperplane) (sums to distribution)
MAXTPCPD )()(
MAXNOMNOM TscsaGcaCAi
iciiaii
iii )()(
11 Tau 2002
Comparison to purely statistical timing• Two approaches are complementary
Propagate functions
Propagate distributions
Explicit computation
Statistical timing
12 Tau 2002
Similarities in timing analysis
• Extraction and delay reduction are straightforward, timing is not
• Latest arriving signal is now poorly defined
• If a significant probability for more than one signal to be last, both must be kept (or some approximate bound applied).
• Pruning threshold will determine accuracy/size tradeoff.
• Must compute an estimate of parametric yield at the end.
• Provide a probability of failure per path for optimization.
13 Tau 2002
Differences
• Propagate functions instead of distributions
• Distributions of process parameters are used at different times
‑ Statistical timing needs process parameters to do timing analysis
‑ Explicit computation does timing analysis first, then plugs in process distributions to get timing distributions.
‣Can evaluate different distributions without re-doing timing analysis
14 Tau 2002
Pruning
• In statistical timing
‑ Prune if one signal is ‘almost always’ earlier
‑ Need to consider correlation because of shared input cones
‑ Result is a distribution of delays
• In this explicit computation of timing
‑ Prune if one is earlier under ‘almost all’ process conditions
‑ Result is a function of process parameters
‑ Bad news – an exact answer could require (exponentially) complex functions
‑ Good news - no problem with correlation
15 Tau 2002
The bad news – complicated functions
a
01234567
891011121314151617181920
01234567 8910111213141516171819200
0.5
1
1.5
2 • Shows a possible pruning problem for a 2 input gate
• Bottom axes are two process parameters; vertical is MAX(A,B)
• Can keep it as an explicit function and prune when it gets too expensive
•Can cover with one (conservative) plane
0.8-0.2*P1+1.0*P2
0.7+0.5*P1
A
B
16 Tau 2002
The good news - reconvergent fanout
• The classic re-convergent fanout problem
• To avoid this, statistical timing needs to keep careful track of common paths – can take exponential time
17 Tau 2002
Reconvergent fanout (continued)
• Explicit calculation gives the correct result without common path calculations
D0+P1
P1 =
D1+P1D2+P1
D1+P1
Plug in distribution for P1
18 Tau 2002
Real situation is a combination of both
• Gate delays are somewhat correlated but have a big statistical component
• Wire delays (particularly close wires) are very highly correlated but have a small random component.
• Delays consist of two parts that combine differently
Distribution of statistical part is also a function of process variation
19 Tau 2002
So what’s the point of explicit computation?
• Not so worst case timing predictions
‑ Users have complained for years about timing pessimism
‑ Could be about 10% better (see experimental results)
‑ Could save months by eliminating unneeded tuning
• Will catch errors that are currently missed
‑ Fast/slow combinations are not currently verified
• Can predict parametric yield
‑ What’s the timing yield?
‑ How much will it help to get rid of a non-critical path?
20 Tau 2002
Predicted variations are always smaller
• Let C = C0 + k0p0+ k1p1 , , where p0 has deviation and p1 has deviation .
• Then worst corner case is:
• But if p0 and p1 are independent, we have
• So the real 3-sigma worst case is
• Which is always smaller by the triangle inequality
211
200 )()( kk
211
2000
211
2000 )3()3()()(3 kkCkkC
11000 33 kkC
21 Tau 2002
Won’t this be big and slow?
• Naively, adds an N element float vector to all values
• But, an x% change in a process parameter generally results in <x% change in value
‑ Can use a byte value with 1% accuracy
• A given R or C usually depends on a subset
‑ Just the properties of that layer(s)
• Net result – about 6 extra bytes per value
• Some compute overhead, but avoids multiple runs
22 Tau 2002
Experimental results for explicit part only
• Start with a 0.18 micron, 5LM, 144K net design
• First – is the linear approximation OK?
‑ Generated 35 cases with –20%,0,+20% variation of three most relevant parameters for metal-2 layer
‑ For each lumped C value did coeffgen, then HyperExtract, then a least-squares fit
‑ Less than 1% error for C = C0 + k0p0+ k1p1 + k2p2
• Since delay is dominated by C, this means delay will also be a (near) linear function of process variation.
23 Tau 2002
More Experimental Results
• Next, how much does it help?
‑ Varied each parameter (of 17) individually
‑ Compared to a worst case corner (3 sigma everywhere)
‑ Average 7% improvement in prediction of C
• Will expect a bigger improvement for timing
‑ Since it depends on more parameters, triangle inequality is (usually) stronger
24 Tau 2002
Conclusions
• Outlined a possible approach for handling process variation
‑ Handles explicit and statistical variation
‑ Theory straightforward in general
‣Pruning is the hardest part, but there are many alternatives
‑ Experiments back up assumptions needed
‑ Memory and compute time should be acceptable