OpenAI models helped redesign AMD chips with up to 14% gains in power, performance and area

OpenAI says its models improved AMD's flagship GPU design by up to 14% on combined power, performance and area — about a year's design progress in under five months — and cut area by about a tenth on its own Jalapeño inference chip.
OpenAI said in a research update published on November 18 that its AI models can now deliver real improvements in chip design. Working with AMD, the models rewrote the code describing the circuits of AMD's flagship GPU, improving the combined measure of power, performance and area (PPA) by up to 14% against baselines that were already heavily optimized. OpenAI described that as roughly a year of progress achieved in less than five months.
Its own chip first: Jalapeño's measured gains
The work grew out of OpenAI's first custom inference chip, Jalapeño. The team began using GPT-o3 mini on the design in early 2025 and went from initial design to tape-out in nine months. Model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, including one multiplier redesign that used about 15% less cell area while cutting latency from two clock cycles to one. The optimizations were proven in silicon bring-up, and engineering samples are running machine-learning workloads, including GPT-5.3-Codex-Spark, at target frequency and power, with initial deployment planned by the end of 2026.
The AMD collaboration: flagship GPU and next CPU
With AMD, the models rewrote SystemVerilog RTL — the code that describes a chip's circuits — and used Synopsys tools to verify that each rewrite preserved the original behavior and to measure the trade-offs in power, performance and area. AMD has already included some of the model-proposed changes in a recent design it sent for manufacturing, and added the rest to its production design files for future versions. AMD also verified single-digit percentage PPA gains on key blocks of its next flagship CPU using model-written designs.
Open-source processor tests: the newest model leads
To measure how chip-design ability improves from one model generation to the next, OpenAI ran 9,881 scored evaluations on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex, asking models either to optimize the existing designs or to build cores from written specifications. With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra recorded the highest average score on both cores, and GPT-6.1 Sol also outscored GPT-5.6 Sol. OpenAI was candid about a caveat: models scored higher designing from a specification partly because they could omit features of the original cores, such as performance counters, that the test workloads did not reward.
What comes next, and the limits
OpenAI said it is working with Synopsys to train GPT-Synopsys, a new model that uses electronic design automation tools the way expert chip designers do. The results still have limits: as of November 19, AMD had published no statement of its own and named no product carrying the model-proposed changes; 14% is a combined power, performance and area figure, not a 14% speed gain, and there are no measurements yet from finished AMD silicon. In the same week, Anthropic said its Claude Mythos Preview model got Claude running on AMD MI355X GPUs in about three days — a sign that AI-designed chips are becoming a new front in the industry's competition.
(Illustration: AMD and OpenAI logos over a circuit board. Image: TechReport)
N環球 
💬 Comments
No comments yet. Be the first!