It learns the equations
The network is trained against the governing equations themselves, so it holds up on structures it has never seen rather than interpolating between ones it has.
AI-accelerated reliability analysis
AbstractAccounting for manufacturing variation in an electromigration
analysis means running the same design many times over a distribution of material parameters.
Done numerically that is an overnight job, so it gets dropped and replaced by a guessed
margin. MetalStack-AI replaces the inner solve with a physics-informed neural network trained
on the governing equations rather than on a corpus of past results. Measured against the
numerical engine across five structure sizes, it returns the same analysis up to
86× faster — and up to 243× faster than a commercial tool —
while staying within 0.05% of the physics-exact answer. End to end, a 48-minute
commercial run becomes 20 seconds.
1 / WHY IT IS TRUSTWORTHY
Trained on the physics, not on the examples
Most AI in EDA is trained on solved examples, which makes it accurate on designs that resemble the training set and quietly wrong on designs that do not. That is the wrong failure mode for a sign-off flow, because the designs you most need to check are the unusual ones.
It is measured against the exact engine
Every number in the table below is a difference from a MetalStack Core run on the same structure. Speed with an unknown error bar is not speed you can use.
Training happens once, on our side
You get inference: seconds per run, no GPU cluster to provision and no data-collection project of your own to run first.
Variation is built into the answer
The result is a distribution, so the confidence range falls out of the same run rather than a separate statistical campaign.
2 / SPEED AND ACCURACY
Hours of analysis, done in seconds
The same full variation-aware reliability analysis run three ways, on structures from small to large.
| Structure size | Commercial tool | MetalStack Core | MetalStack-AI | Speedup | Difference |
|---|---|---|---|---|---|
| Small | 22 min | 7.6 min | 0.25 s | 86× | 0.02% |
| Medium | 37 min | 12.7 min | 0.43 s | 77× | 0.03% |
| Large | 50 min | 17.2 min | 0.39 s | 63× | 0.03% |
| Very large | 60 min | 22.2 min | 0.61 s | 46× | 0.04% |
| Largest | 69 min | 25.4 min | 0.80 s | 36× | 0.04% |
Four hundredths of a percent, worst case, on structures with hundreds of segments.
For every decision you would make from this analysis — does this net fail, when, do I need
to widen it — the fast answer and the exact answer are the same answer.
3 / END TO END
An hour of waiting becomes 20 seconds
The same comparison measured end to end rather than per structure, across the complete variation-aware analysis.
48 minCommercial tool, full analysis
17 minMetalStack Core — 2.8× faster
20 secMetalStack-AI — 145× faster
That is the difference between an analysis you schedule overnight and one you run every
time you change the design. The commercial-tool comparison in both this section and the
table above was run against COMSOL.
4 / IN PRACTICE
Where a 20-second analysis fits that a 48-minute one does not
Inside the optimization loop
A design-space search needs thousands of evaluations. At 20 seconds each that is a coffee break; at 48 minutes it is a quarter.
On every commit
Reliability stops being a gate at the end of the project and becomes a check that runs alongside everything else.
Across the whole corner space
Sweeping thermal maps and workloads stops being a budgeting exercise and becomes something you simply do.
Still signed off exactly
Anything the fast engine flags can be re-run on MetalStack Core before it goes in a report — see WP-01 for what that engine produces.
5 / METHOD AND SCOPE
Reference and what the benchmark establishes
The method behind MetalStack-AI is published and peer-reviewed as BPINN-EM: Fast Stochastic Analysis of Electromigration Damage using Bayesian Physics-Informed Neural Networks, ICCAD 2024 — read the paper.
Benchmark scope: five interconnect structure sizes, each run as a full variation-aware reliability analysis against MetalStack Core on the same structure and against COMSOL as the commercial reference. Reported timings exclude network training and are hardware-dependent. These are accuracy figures for the structures benchmarked, not guarantees for an arbitrary design outside the trained range.
From evidence to your design
Put the fast engine on your own grid.
Bring a design and the question you need answered, and we will show you both engines on it — the fast answer, the exact one, and the gap between them.