AMD's flagship GPU design, with changes identified by OpenAI's models, has been sent for manufacturing, the AI company said Nov. 18, marking a step for model-assisted chip design.
OpenAI said AMD's flagship GPU intellectual property showed gains of up to 14% across power, performance and area (PPA), measured against baselines it described as already extensively optimized. Reworking circuit-description code revealed additional improvement opportunities, it said.
OpenAI said it equated the results from a project lasting under five months with about a year of design progress. Further revisions, it said, are recorded in design files for future versions.
OpenAI said AMD confirmed single-digit PPA gains across key blocks in its next-generation flagship CPU, using model-produced designs. OpenAI also applied the method to Jalapeño, its first custom inference chip, and said the improvements were verified on silicon.



