AlphaGazette
Menu

Tech

OpenAI models improved AMD's flagship GPU design by up to 14% PPA

AMD has already placed some of those circuit changes in a design sent to be manufactured. Gains on its next flagship CPU stayed in the single digits, OpenAI said.

1 min read AI-generated

OpenAI says models it has not named improved the circuitry of AMD's forthcoming flagship GPU by as much as 14% on electric power, performance and silicon area together. AMD has already placed some of those changes in a design it sent to be manufactured. On AMD's next flagship processor, the same effort produced only single-digit gains, OpenAI said.

The edits were made in SystemVerilog, a language used to describe high-end chip circuits. OpenAI said a model found further gains by rewriting that circuit description, condensing about a year of ordinary design progress into less than five months, a pace it called unprecedented for this work. It said the model reasoned over the design at an expert microarchitect's level, and that it expects more improvement on both the GPU and the CPU from the new baseline.

OpenAI has applied models to silicon before. With Broadcom it designed Jalapeño, a chip aimed at inference, and engineering samples were running real workloads in June 2026. The company had expected only small Jalapeño deployments around late 2026, with a wider rollout in 2027. Benchmarks already out for the chip purport lower power draw and latency than Nvidia Blackwell GPUs on single-token prediction.

Researchers also scored high-end models on the open PicoRV32 and Ibex cores. Each attempt got 112,000 tokens and a goal of higher throughput, meaning instructions per clock together with the length of a clock cycle, and the models were not told which workloads the chips would run. A design scored zero if it failed or used more than 16x the earlier area, and 1.0 if it was the best result inside the test limits. GPT-6.2 Astra posted both the top score and the lowest cost. A second test found stronger results when models rebuilt a core from its original specification instead of tuning code that already existed.