OpenAI announced Wednesday that its models improved power, performance and area metrics of AMD’s flagship GPU designs by up to 14%. The company also reports modifications integrated into a chip sent to manufacturing, while maintaining systematic verification of the changes proposed by the AI.
Tests on two processors and verified changes
OpenAI evaluated five generations of models on two small open source processors. According to the company, its latest model, GPT-6.2 Astra, comes out on top both for optimizing an existing design and for creating a processor from written instructions. However, she specifies that this last test was easier: the models could leave aside certain functions of the original processors.
Last month, Richard Ho, head of chips at OpenAI, said he did not yet have complete confidence in the design changes suggested by AI. He explained that everyone was therefore subject to verification.
Gains announced for AMD and Jalapeño
The announced comparison concerns designs that AMD had already heavily optimized. Power, performance and surface area are the metrics that designers must balance between. According to OpenAI, the model refactored the code describing the circuits, then used Synopsys software to monitor their operation after each change. According to the company, the progress made in less than five months corresponds to approximately a year of work. AMD has also confirmed, it reports, single-digit gains on key elements of its next flagship CPU.
The evening letterFree · 7 p.m.
Are you following the race for AI models?
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Jalapeño, OpenAI’s first in-house designed chip, benefited from this method, according to the company, and went from design to manufacturing in nine months. According to the company, the suggested models have reduced the size of two key parts of this chip by approximately 13% and 10%. Their effectiveness has been confirmed on manufactured silicon, she says.
A model with Synopsys and a fundraising for Architect Labs
OpenAI is currently training GPT-Synopsys in collaboration with Synopsys to drive the tools that engineers already use to design their chips. Synopsys refocuses on AI-assisted design. In this same area, Architect Labs obtained $24 million in funding in June to support companies in creating personalized chips using AI.

