OpenAI President Greg Brockman said the company's AI models helped AMD achieve the equivalent of "an extra year" of chip development progress in less than five months, with some optimizations already incorporated into production-ready chip designs.
In a post on X on Wednesday, Brockman said that with the help of OpenAI's models, "AMD's chips achieved nearly an additional full year of development progress in under five months," adding that some results have already been used in a chip design that recently completed tape-out, calling it "an unprecedented achievement at the frontier of chip design."
In a research report released on Wednesday, OpenAI disclosed that through its model optimizations, AMD's flagship GPU design achieved improvements of up to 14% across power, performance, and area (PPA). OpenAI's models rewrote chip circuit instructions, and these changes were verified through Synopsys software, confirming that the modifications optimized the chip while preserving its original functionality.
AMD incorporated multiple improvements into a chip design that was recently completed and sent for manufacturing, while also adding more optimizations to subsequent production plans. OpenAI also revealed that AMD has validated single-digit percentage improvements in PPA metrics in key components of its upcoming flagship central processing unit (CPU).
This progress further deepens the partnership between the two companies. The collaboration has drawn significant attention since October 2025, when the two companies announced plans to deploy AMD processors with a total computing scale of 6 gigawatts (GW). Since then, AMD has partnered with multiple leading AI developers to optimize software performance across its hardware ecosystem.
OpenAI also improved its custom "Jalapeno" inference chip, reducing the circuit area of one computing component by approximately 13% and another component by approximately 10%. The company previously disclosed this chip, developed jointly with Broadcom, in August.
Multiple independent evaluations show that GPT-6.2 Astra outperformed previous models in chip design tasks involving the open-source Ibex and PicoRV32 processors; however, OpenAI also acknowledged limitations in its testing methodology.
In September of this year, reports said OpenAI partnered with Synopsys to develop "GPT-Synopsys," a system that uses specialized engineering software to optimize semiconductor design. Brockman said this collaboration is expected to shorten development cycles by "weeks or even months," while traditional verification tools will continue to be used to check AI-generated designs.
Comments