The AutoPDK platform

From first silicon
to a released PDK.
Automated.

AI agents that model the silicon your process makes — so a fab releases its process design kit in weeks, not months, with an engineer approving every stage.

The life of a chip

Years to build a process. Then every chip waits here.

01Crystal & waferSand to sliced silicon.months
02Process buildHundreds of tuned steps, each one a recipe.years
03Test siliconThe first wafers of test devices come out.weeks
04CharacterizeMeasure every device: thousands of curves, wafers, temperatures.~8 weeks
Where the months go05ModelFit physics models to the data — dozens of devices, hundreds of parameters.2–3 months
Where the months go06PDK releaseCorners, statistics, QA, versions: the design kit customers need.repeated every release
Where the months go07DesignCircuits built on the PDK — and re-verified every time it moves.months, repeated
08ProductionVolume wafers. Revenue.at last

No PDK, no customers. Every chip on every fab waits on steps 5 to 7 — and they run again after every process tweak.

And it is never done once

It loops. Every release.

v0.1 on first silicon. v0.5 when the process is tuned: refit everything. v0.9: refit again, add statistics. v1.0, and every tweak after that. One node, five releases — and today each one is a quarter of work, split across teams and toolsets with files passed between them.

Today, per PDK release
Characterize · 8 weeksModeling & release · 2–3 months
~5 months
With AutoPDK
Characterize · 8 weeks
~2 months
≈ 11weeks backon every PDK release
≤ 1week of modelingdown from 2–3 months
1engineerwhere modeling often took several

Mature-node PDK release. Modeling times from hands-on extraction experience; characterization time unchanged.

Two moments every fab knows

So we put one AI layer across both.

01 · Model & release

The process has moved. The extraction method lives with one senior engineer.

It lives in one engineer's head and four hundred lines of macro.

AutoPDK makes the method itself the asset. Described once, in plain language, it becomes a plan the AI executes on every new technology: decks generated, stages fitted in physics order against real SPICE, every stage checked — and nothing fitted before an engineer says go.

Months → 1 week of modeling per release · as few as one engineer
AutoPDK › AutoPDK fit: measured versus extracted model
02 · The same change, with one data chain

A process change lands in the fab.

Re-measured, re-modeled, re-released. In days, not a quarter.

The change flows from the new measurements into the models and on into the released kit your designers pull. Your teams keep their tools and their judgment; the modeling hand-offs are gone.

≈ 11 weeks back on every PDK release
Why us ›
What the names mean

Auto, as in autonomous.

AutoPDKAutonomous process design kit

The compact models and the design kit a fab ships to its customers — extracted, checked and released by AI agents, with your engineers approving every stage.

The product

One product. One data chain.

From your characterization data to a released, verified design kit — one chain, no hand-offs.

One data chain

Silicon in. Released PDK out.

Measured siliconyour characterization data
ModelsAutoPDK
PDK releaseAutoPDK
Who it is for

Built for the teams that ship process design kits.

New fabs coming online

A first PDK, then a release every few weeks while the process settles — without a modeling department to match.

Foundries launching nodes

Every node loops measure → model → release many times before v1.0. Each loop gets shorter.

Specialty fabs

GaN, SiC, RF and power processes with their own compact models — any standard model, one engine.

Fabless design teams

Models you can trust for your corners, and designs re-verified the moment the PDK moves.

How we build it

AI that does the work. Engineers who decide.

Physics

Not a black box

Real compact-model parameters, real SPICE, real sign-off decks. No surrogate model standing in for the device.

Control

The expert stays in charge

Plan first, then fit. Lock, re-run or override any stage; nothing is fitted before the engineer says go.

Deployment

Runs where your data lives

On-premises or cloud, with a hosted or fully local language model. Data never has to leave the fab.

Memory

Institutional memory

Methods, model cards, bounds and PDK versions are tracked — the next release starts where the last one ended.

Release your next PDK in weeks.

AutoPDK is ready today. Bring your data and your method — we run the first extraction with you.