OpenAI Models Squeeze Fourteen Percent From AMD GPU Designs
The gains are per block, the numbers are OpenAI's own, and the benchmark's fine print is printed in the same post.
The models aren't drawing chips. They're editing the code that describes them, and the edits survived a tape-out.
OpenAI says its models rewrote AMD's flagship GPU code and improved power, performance and area by up to 14 percent.
The unit matters. PPA means power, performance and area scored together, and the 14 percent is measured per block against baselines AMD had already heavily optimized. It isn't a 14 percent faster GPU. The models rewrote SystemVerilog RTL, the code that describes the circuits, and OpenAI estimates the work equals roughly a year of progress done in under five months.
AMD's part is checkable in one direction. OpenAI says AMD has already included some of the changes in a design it sent for manufacturing, and added others to its production design files for future versions. AMD also verified single-digit percentage gains across key blocks of its next-generation flagship CPU. Every figure in the post is OpenAI-reported, and AMD hasn't published a note of its own beside it.
The same method built OpenAI's own silicon first. On Jalapeño, the company's first custom inference chip, model-proposed changes cut standard-cell area by about 13 percent in the vector unit and about 10 percent in the matrix unit, and those numbers were proven at silicon bring-up. One multiplier redesign used about 15 percent less cell area and cut latency from two clock cycles to one.
The benchmark half of the post runs on two open-source RISC-V cores, PicoRV32 and Ibex. At a budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both. The full evaluation covered 9,881 scored runs, with correctness checked in simulation and clock speed estimated by the open-source Yosys tool rather than measured on fabricated parts.
OpenAI prints its own caveats, and they're load-bearing. Models that designed a core from a written specification outscored models improving an existing core, and OpenAI notes the from-scratch designs could simply leave out features the originals carry, such as performance counters and extra instructions. It also says tighter area limits and timing checks after physical synthesis would make the test more representative of production work.
The verification step is what separates this from a demo. Synopsys tools checked that each rewrite preserved the block's behavior, and OpenAI is working with Synopsys to wire its models into the EDA software engineers already use, including a model called GPT-Synopsys. A model that proposes edits is a curiosity. A model whose edits pass the industry's own checking tools and reach a tape-out is a workflow.
The conversation.
Loading comments…
A little more perspective.
Reading is free. Verified members can comment and reply for $2 a month.