OpenAI said on Wednesday, Nov. 18, that its AI models had helped improve parts of the design for AMD's next-generation flagship GPU. GPUs are advanced AI chips, the hardware AI models are trained and run on. OpenAI says that, working with AMD, it saw gains of up to 14% in power use, speed and size, measured against designs that had already been heavily optimized. Chip designers have to balance those three against one another, and OpenAI says the model weighed those trade-offs. OpenAI did not say which of the three the 14% refers to.
OpenAI's model, which it did not name, worked by rewriting the code that describes the chip's circuits, OpenAI says. It used chip design software from Synopsys to check that each rewrite still did the same job. OpenAI calls the result roughly a year of progress in less than five months, without saying what it compares against. It says AMD has already put some of the changes into a recent design it sent for manufacturing, and added others to its design files for future versions.
OpenAI says AMD also confirmed smaller gains, under 10%, in key parts of its next flagship CPU, a computer's main processor, using designs the model wrote. OpenAI says it expects more gains on both chips.
OpenAI first said in August that its models helped design Jalapeño, its first chip of its own for running AI models. Its new post gives figures. OpenAI started using an older model, GPT-o3 mini, on the design in early 2025. It estimates that the models' changes shrank the space taken by the chip's basic circuit building blocks by about 13% in the part that runs many calculations at once, and 10% in the part that does arithmetic. It says those savings held up when the first chips were made and tested.
To track progress from one model to the next, OpenAI asked five of its models to speed up two small, freely available processor designs, and also to design the same processors from written requirements and a template for how each connects to the rest of the system. When each model was given about the same amount of text to work with per attempt, GPT-6.2 Astra had the highest average score on both processors at the speed-up task. At that same allowance, all five models scored higher when designing from written requirements than when improving the existing designs. OpenAI notes that starting from requirements let them leave out features the original designs had, which could explain part of that gap. It says stricter tests would better reflect real chip design.
OpenAI gives only its best result for AMD's chip, not a typical one. It describes the work as on AMD's next-generation flagship GPU, but has not named the chip or said when it will ship.
OpenAI is working with Synopsys to train GPT-Synopsys, a model meant to use Synopsys' design software the way expert chip designers do. The two companies announced it on Sept. 30, said early work with leading chipmakers was under way, and have not said when customers can use it.