---
title: "OpenAI Rewrote Part of AMD's Circuit. Proving It Was the Hard Part."
date: 2026-11-19
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-openai-amd-chip-1119
language: en
---

# OpenAI Rewrote Part of AMD's Circuit. Proving It Was the Hard Part.

> On Advanced Micro Devices' (NASDAQ: AMD) flagship GPU IP, one of the three PPA metrics — power, performance, area — improved by up to 14%. A model did the rewriting.

On Advanced Micro Devices' (NASDAQ: AMD) flagship GPU IP, one of the three PPA metrics — power, performance, area — improved by up to 14%.

A model did the rewriting.

OpenAI disclosed the work on November 18; verification tools came from Synopsys (NASDAQ: SNPS). Its framing: about a year of progress in under five months.

AMD saying so publicly is the signal: model-assisted design now sits inside a mainstream chip company's flow. That label sells better than 14%.

## Where the 14% sits

Best of three metrics, not a gain on all three. PPA pulls against itself: faster burns more power, smaller is harder to route. A quoted PPA gain usually means a weighted composite, or one metric up with the rest flat. The announcement does not say which.

The baseline matters more. This was not a rough draft but a heavily optimized design, every block converged many times over. Taking 14% off that is not taking 14% off a sketch. The number is small because it moved where experienced engineers had already conceded.

## What the model does

Chip design runs on people writing RTL, code that describes circuit behavior in SystemVerilog. It states logic: which signal arrives, when a block turns on and off. The loop was a person writing, tools checking, the person revising.

The model took the middle step: read existing RTL, rewrite it, hand it to tools. It sounds like code completion, but that aims at something that compiles. RTL rewriting aims at timing, area and power that do not degrade together, in a form tools can close. The model has to know what the logic does before choosing which layer can move. Move the wrong one and the function changes.

OpenAI says the model reasons about RTL "at the level of a senior microarchitect." That claim is OpenAI's, unreproduced.

One boundary: module work inside a flagship GPU, not a model designing a chip from scratch. Between a module and a full chip sit partition, clock tree, power budget and packaging. Blur that and 14% reads larger than it is.

OpenAI was using GPT-4o in Jalapeño design in early 2025; by GPT-6.2, under two years later, it had gone from participant to main contributor.

## Verification costs more than generation

Many models write code. Few are trusted to change circuits before tape-out. The difference is verification.

Every rewrite must be proven equivalent with Synopsys tools, then judged on PPA. Behavior is the floor: faster or smaller is negotiable, a changed function is scrap. Nobody skips it: the time saved will not cover one bad tape-out. Each edit needs a run. Generation is parallel; verification is serial.

AMD re-verified independently. Some changes went into designs taped out recently; others into volume production files for later revisions. Tape-out is a trial batch; volume production ships at scale. Clearing the first gate does not mean clearing the second.

The numbers on Jalapeño, AMD's in-house inference chip, are finer. Standard-cell area fell about 13% in the vector unit and 10% in the matrix unit. A redesigned multiplier cut area about 15% and latency from two cycles to one; an adder cut area about 15% at the same latency. Area and latency trade off, so the multiplier gaining both is the hardest result.

These optimizations are already up on silicon. Key blocks of AMD's next flagship CPU showed single-digit percentage PPA gains. From GPU to CPU, the method is spreading.

## Reading the scores

OpenAI also ran public evaluations on two open 32-bit RISC-V cores, PicoRV32 and OpenTitan's Ibex. Each run had a budget of about 112,000 tokens; the full evaluation produced 9,881 valid scores. The five models: GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol, GPT-5.5. GPT-6.2 Astra had the highest average.

Nearly ten thousand scores are there to shrink variance, not to pick a winner. One result runs against intuition: all five scored higher designing a core from a spec than optimizing an existing one. A blank page has no legacy and none of the features real IP must keep. The models dropped performance counters, the core got simpler, the scores improved. OpenAI concedes it.

So the scores measure whether a model can write logic that runs, not whether it can deliver manufacturable IP. Between a core that runs and a core that sells sit verification, timing closure and yield. None is in the score. Fewer counters make area and timing look good: a hole in the test.

## The moat moved

Five months against twelve: 40% of the time, seven months saved. That buys another architecture exploration, or a thicker verification campaign. But generation speeding up does not speed verification up. A model can propose a hundred revisions a day; there is one flow to prove them. The bottleneck moves from writing it to proving it.

So the moat shifts from how many senior microarchitects you employ to whether you have a high-quality verification flow. Salaries buy the first. The second accumulates through projects: every verification pass or fail is a label, and whoever runs more accumulates faster. That gap is harder to close than "smarter model," because it needs real projects, not more GPUs.

The Synopsys partnership points the same way: the two are training GPT-Synopsys to put model capability inside the EDA toolchain. That puts the referee and the player in one room: the tool judges, the model proposes. The interface belongs to the tool vendor; who owns the data will be in the contract.

One question left: if changing a circuit costs a model plus a verification flow, how large an RTL team does a chip company still need? Hiring plans will answer first.

The real barrier may no longer be whether the model can change the circuit.

It is whether anyone will sign the tape-out.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-openai-amd-chip-1119
Free to quote with attribution and a link to the original.
