OpenAI's Models Cut Chip Area on Jalapeño and Rewrote AMD GPU Circuits in Under Five Months
OpenAI said on November 18 that its models cut Jalapeño standard-cell area by about 13 percent in the vector unit and delivered up to 14 percent better PPA on AMD flagship GPU IP, with some changes already sent for manufacturing.
OpenAI's own models are now rewriting the circuits that will run the next ones. In a post dated November 18, the company says GPT class models cut standard-cell area on its first custom inference chip, Jalapeño, and produced up to 14 percent better power, performance and area on AMD's flagship GPU IP against baselines AMD had already spent years tightening.
The work started small. OpenAI began feeding GPT-o3 mini into Jalapeño in early 2025. In the vector unit, the models tried different adder and multiplier layouts, conversion paths, reduction logic and special-function hardware, and looked for places to share circuitry instead of storing intermediate results. In the matrix unit, one multiplier redesign used about 15 percent less cell area in synthesis and dropped latency from two clock cycles to one. A separate adder redesign also cut area by about 15 percent while keeping its two-cycle latency.
Across the broader Jalapeño effort, OpenAI estimates the accepted changes reduced standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit. Those numbers are the company's, and it says they were later proven in silicon during Jalapeño's bring-up rather than left as synthesis estimates.
AMD's production files, not a lab demo
The AMD results are the part that travels outside OpenAI's own tape-out. OpenAI says it partnered with AMD on PPA improvements and saw up to 14 percent against "heavily optimized baselines" on the flagship GPU IP, by rewriting the SystemVerilog RTL that describes the circuits. It calls that roughly a year of progress in less than five months. AMD has already put some of those changes into a design that went out for manufacturing, OpenAI says, and parked others in production files for later versions.
The model used Synopsys tools to check that each rewrite still did the same job, then to score the power, performance and area trade-offs. AMD also verified single-digit percentage PPA gains on key blocks of its upcoming next-generation flagship CPU, according to the same post. OpenAI does not name the GPU or CPU products.
Measuring the models against each other
To show the skill is moving from one generation to the next, OpenAI ran 9,881 scored attempts on two open-source 32-bit RISC-V cores: PicoRV32, a compact FPGA core, and Ibex, the embedded processor in OpenTitan's security chip. Each model got the original hardware description and a brief to raise throughput, the instructions the core can finish per second. Designs that failed a simulation check, or that ballooned past 16 times the original area, scored zero.
With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both cores. On the existing PicoRV32 design, its score at that budget was about twice GPT-6 Astra's. When the models were instead handed written requirements and asked to design a core from scratch, all five of them scored higher than when they tried to improve the original, which OpenAI notes may partly reflect dropping features the test workloads do not reward.
The EDA seat
The next step is to put that loop inside the tools chip designers already trust. OpenAI says it is partnering with Synopsys to train a model, GPT-Synopsys, to use Synopsys electronic design automation software the way an expert does: propose a rewrite, prove it, score it, keep or throw it away.
The claim worth watching is not that a chatbot can doodle a schematic. It is that AMD already let some of those doodles into a chip that is being made.
Some offers on this page may be paid placements or contain affiliate links.
Subscribe to Techpresso
Free daily newsletter, read in 5 minutes.
Subscribe free