Power-grid reliability for the full chip
From the power grid you already have to the date it fails. Electromigration, thermomigration and IR-drop solved as one coupled problem, across the whole chip.
Actual MetalStack output on an industrial RISC-V power grid at 32/28 nm.
Electromigration, thermomigration and IR-drop solved together, under your chip’s real thermal map.
A physics-exact numerical engine for sign-off, and a PINN engine fast enough for every iteration.
Power grids straight from OpenROAD or Synopsys ICC / Fusion Compiler, scriptable end to end.
Capabilities
A rule-based electromigration check assumes one die temperature and one current limit per wire. Real power grids break both assumptions, over-building the wires that will never fail and missing the ones that will. MetalStack solves the governing physics instead, fast enough to run while the design is still moving.
01 / COUPLED PHYSICS
Electromigration, thermomigration and IR-drop are solved as one connected problem, because on real silicon that is what they are. Voids change the resistance, resistance changes the current, and the current changes the heat that drives both.
The validation study02 / THERMAL INPUT
Supply a measured or simulated temperature profile and MetalStack uses it directly, including the self-heating each wire generates. Where the hotspot sits relative to the current path decides the verdict; the average die temperature does not.
Thermal sensitivity03 / LIFETIME
The grid is run forward through its service life with degradation feeding back into the solve, so you see voltage drop drift year by year instead of a single day-one snapshot.
04 / STATISTICS
Manufacturing spread is carried through the analysis, so the answer is a distribution with real probabilities rather than one pessimistic number you have to guess a margin around.
Lifetime spread05 / INTEGRATION
Grids are read straight out of OpenROAD or Synopsys ICC / Fusion Compiler, with material and solver settings in a single YAML file. Nothing to hand-build and no methodology to adopt.
06 / AUTOMATION
Every input and output is scriptable, so MetalStack drops into automated and AI-driven design loops and closes EM and IR-drop sign-off without a human in the middle.
How MetalStack is built
Same governing equations, two ways to run them. Pick the one the task needs, or use both: explore with the fast engine and confirm with the exact one.
The numerical engine. Physics-exact analysis you can defend in a sign-off review, with every number tracing back to a first-principles solve on the full chip.
The PINN engine. The same analysis up to 86× faster, trained against the governing equations rather than a corpus of previously solved designs.
The fast engine is not a black box. It is benchmarked against Core on the same structures, so every quick answer carries a known margin and anything that matters can be re-run exactly.
The workflow
No model to hand-build and no new methodology to learn. MetalStack takes the power grid already in your layout and tells you where it fails, and when.
01 / INPUT
The power grid from your existing layout, a YAML file of material and solver settings, and a thermal map if you have one. Without a map, MetalStack works from self-heating over an ambient baseline.
02 / SCREENING
A fast first pass drops the nets that will never break, so the expensive compute goes only where the risk actually is.
03 / AGING
Heat, current and wear compound on each other across the service life, with the solve updating as the grid degrades rather than being evaluated once at time zero.
04 / OUTPUT
Where damage forms, how far voltage drop has drifted, the date the design crosses the limit you set, and degraded netlists you can feed back into the flow.
Switch to MetalStack-AI and the same four steps return in seconds rather than hours — same inputs, same outputs, nothing else about your flow changes.
White papers
This page stays short on purpose. The method, the designs each claim was measured on, and the full tables and figures live in a growing series of white papers — each a self-contained document you can read, share and cite in a design review.
WP-01 / VALIDATION
Six industrial power grids at 32/28 nm — voltage drop when new and after aging, time to failure and runtime — including two designs that pass every check you would run today and still fail inside a year.
Read WP-01WP-02 / AI ACCELERATION
What MetalStack-AI costs in accuracy for up to 86× in speed, measured against the numerical engine across five structure sizes. Published and peer-reviewed at ICCAD 2024.
Read WP-02WP-03 / IN PREPARATION
Driving EM and IR-drop sign-off from a scripted loop: what to expose to an optimizer, how to budget exact against accelerated runs, and what closing the loop changes about the grid you ship.
The full seriesTHE SERIES
Forthcoming papers will cover transient workload benchmarking and multi-corner sign-off. Tell us what you need to see, or ask to be notified when the next one is published.
Get notifiedStart with your power grid
Bring us a design — or just tell us what you are taping out and on which flow — and we will walk you through a run end to end, including the parts that look bad.