OpenAI models cut chip area on Jalapeño and improve AMD GPU design
Advancing chip design with OpenAI models
OpenAI published research on models helping with chip design. Standard cell area on two Jalapeño compute blocks fell about 13% and 10%, and silicon debugging confirmed the optimizations. Work with AMD on a heavily optimized flagship GPU IP baseline reached up to 14% better power, performance and area (PPA), with some changes already in designs sent to manufacturing.
Why it matters: In OpenAI's evaluation, designing a processor from a spec scored higher than optimizing existing circuits, and dropping some original features may explain that edge.