Skip to header Skip to main navigation Skip to main content Skip to footer
Scott Lawson
Your next upgrade lives here

OpenAI Says Its Models Rewrote AMD GPU Circuits for Up to 14 Percent Better Efficiency

A top-down photograph of an AMD Ryzen 7 7800X3D processor showing the AMD Ryzen branding on its metal heat spreader.

OpenAI's models are no longer just running on advanced chips. According to a technical report the company published on November 18, 2026, titled "Advancing chip design with OpenAI models," they are helping to design those chips as well, and some of the changes they proposed are already on their way into production hardware.

The headline result comes from work with AMD. OpenAI says its models rewrote parts of the SystemVerilog RTL, the code that describes a chip's circuits, for AMD's flagship GPU intellectual property. Each rewrite was checked with Synopsys design software to confirm that it still performed the same job, and then scored for its effect on power, performance and area, a combined measure chip designers call PPA. Against baselines that AMD had already heavily optimized, the models achieved up to a 14 percent improvement in PPA. OpenAI describes that as roughly a year of design optimization progress compressed into less than five months.

That figure needs careful reading. OpenAI does not name the AMD GPU involved, does not break the 14 percent down between power, performance and area, and reports no measurements from finished AMD silicon, so it does not mean the GPU itself runs 14 percent faster. What makes the claim harder to dismiss is what happened next. According to OpenAI, AMD included some of the model-proposed changes in a design recently sent for manufacturing, and added the rest to its production design files for use in later versions. AMD also verified single-digit percentage PPA gains on key blocks of its next flagship CPU using model-written designs. As of November 19, AMD had published no statement of its own about the work.

The AMD work grew out of OpenAI's own silicon. Jalapeño, developed with Broadcom, is OpenAI's first custom inference chip, and its engineers began using the company's models to help design it in early 2025. The chip moved from design to tape-out in nine months. OpenAI estimates that the accepted model-proposed changes reduced standard-cell area by about 13 percent in the vector compute unit and about 10 percent in the matrix compute unit, and it says those savings were confirmed when the real chip was brought up in the lab. One multiplier redesign used about 15 percent less cell area while cutting its latency from two clock cycles to one.

To test whether the skill is improving from one model generation to the next, OpenAI ran 9,881 scored evaluations on two open-source 32-bit RISC-V processor cores, PicoRV32 and Ibex. Under a budget of roughly 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score on both processors. OpenAI is open about a caveat in that test: models scored higher when designing a core from a written specification than when improving an existing design, partly because they could leave out features, such as performance counters, that the benchmark did not reward.

The next step is to give that loop better tools. OpenAI says it is working with Synopsys to train a model called GPT-Synopsys, intended to use professional electronic design automation software the way an expert chip designer would: propose a rewrite, prove that it works, score the trade-offs, and keep or discard it. If gains like these hold up in finished silicon, the implication is significant. Better models would help design better chips, and those better chips would then run the next generation of models.

Photo: An AMD Ryzen 7 7800X3D processor photographed by Fritzchens Fritz, via Wikimedia Commons, CC0 1.0.

Artificial Intelligence
OpenAI
Semiconductors
AI Infrastructure
Technology
Innovation
Big Tech

Copyright © 2026 Rocky Mountain Madman LLC - All rights reserved

Developed and Designed by Rocky Mountain Madman LLC