✎ AI News

AI Rewrote AMD's Circuit Code, and Some of It Is Already in Silicon

OpenAI reports up to 14% PPA improvement on AMD flagship GPU IP against heavily optimized baselines, about a year of design progress in under five months, with some changes already in a chip sent for manufacturing.

By AIToolsRecap November 22, 2026 7 min read 18 views
Home › Articles › AI News › ChatGPT › OpenAI Says Its Models Cut 14% From AMD Chip De...
AI Rewrote AMD's Circuit Code, and Some of It Is Already in Silicon
The short version
● What OpenAI says its models rewrote AMD GPU circuit code and found up to 14% PPA gains.
● Speed About a year of design progress in under five months.
● Shipped AMD put some changes into a design already sent for manufacturing.
● Caveat Every figure is OpenAI's. AMD has not published its own.

OpenAI published an account on 18 November of its models rewriting the code that describes AMD chip circuits, and it is the most concrete claim anyone has made about AI doing silicon design work that ships.

The headline number: up to 14% improvement in PPA — power, performance and area — on AMD flagship GPU IP. The qualifier matters and OpenAI includes it: that is against "heavily optimized baselines", not against a naive starting point.

What the model actually did

It rewrote SystemVerilog RTL, the code that specifies a circuit's behaviour before synthesis. Synopsys tools then verified that each rewrite preserved the original behaviour, which is the part that makes the claim checkable rather than a demo.

That is a meaningful distinction. The model is not drawing transistors. It is editing hardware description code and having a formal toolchain confirm the edit is equivalent.

The numbers OpenAI published

BlockResult
Flagship GPU IPUp to 14% PPA vs heavily optimized baselines
Jalapeno vector unit~13% standard-cell area reduction
Jalapeno matrix unit~10% standard-cell area reduction
Matrix multiplier~15% area, latency cut from two clock cycles to one
Next-gen flagship CPUSingle-digit % PPA across key blocks, verified by AMD
Timeline~1 year of progress in under 5 months

The multiplier result is the one worth pausing on. Cutting latency from two clock cycles to one is not an incremental area saving; it changes the timing budget of everything downstream of it.

Some of it is already in silicon

OpenAI says AMD included some of these changes in a design it recently sent for manufacturing, and added others to its production design files for use in future versions. It also says the Jalapeno optimisations were silicon proven during that chip's bring-up, and that an early model — GPT-o3 mini — was working on Jalapeno from early 2025.

So this is not a projection. Parts of it are on a wafer.

What to hold lightly

Every figure here is OpenAI's. There is no AMD press release and no AMD quote anywhere in the post, and the single-digit CPU gains are OpenAI's account of AMD's verification rather than AMD stating it. Coverage elsewhere has already dropped the "heavily optimized baselines" qualifier, which is the phrase that makes 14% a real number instead of a marketing one.

OpenAI also does not name which model did the AMD work. It says only "our model". The named models — GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol — appear in a separate RISC-V test where all five scored higher designing from a specification, at roughly 112,000 tokens per run.

FAQ

Did AI really design an AMD chip?
OpenAI says its models rewrote SystemVerilog RTL for AMD circuit blocks and found up to 14% PPA improvement. AMD engineers kept the work; the models did not design a chip end to end.
What does PPA mean?
Power, performance and area — the three things chip designers trade against each other. A PPA gain means better on at least one without losing the others.
Has any of it shipped?
OpenAI says AMD included some changes in a design already sent for manufacturing, and that the Jalapeno optimisations were silicon proven during bring-up.
Has AMD confirmed the 14%?
Not publicly. The figure comes from OpenAI's post, which contains no AMD statement.
Which OpenAI model did the work?
Not named. OpenAI says only "our model". GPT-6.2 Astra and three others are named in a separate RISC-V design test.
Tags
AI NewsOpenAINvidiaGenerative AI2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →