OpenAI Says Its Models Found Up to 14% PPA Gains in AMD's GPU Design; Figures Are OpenAI's
OpenAI said on November 18 that its models found up to 14% better power, performance and area (PPA) on AMD's flagship GPU design, against heavily optimised baselines, in under five months, and that AMD has put some of the changes into a design sent for manufacturing. The figure is OpenAI's: AMD chief executive Lisa Su called the results fantastic on X, and we found no numbers published by AMD. OpenAI also reports that model-proposed changes to its own Jalapeño chip were silicon proven, and that it is training a GPT-Synopsys model with Synopsys. OpenAI's own RISC-V tests are not a production benchmark, it says.

Chip design is one of the few places where an AI lab's claim eventually meets silicon, and OpenAI's new post sorts its results by how far they have travelled. On its own Jalapeño chip, OpenAI says model-proposed changes were silicon proven. On AMD's flagship GPU, it says its models found up to 14% better power, performance and area, known as PPA, than heavily optimised baselines in under five months. That figure is OpenAI's. AMD chief executive Lisa Su called the results "fantastic" on X, and we found no numbers published by AMD.
OpenAI says AMD has already put some of the changes into a recent design it sent for manufacturing and added others to its production design files. If that holds, the models' work is in AMD's pipeline, not only in a lab demonstration. The post does not say how the 14% splits across power, performance and area, or which baseline it is measured against beyond "heavily optimized."
What OpenAI reports#
On the AMD work, OpenAI says its model rewrote the SystemVerilog code that describes the chip's circuits and used Synopsys tools to check that each rewrite preserved the block's behaviour and to weigh the trade-offs in power, performance and area. OpenAI calls the result roughly a year of progress achieved in less than five months. It also says AMD verified single-digit PPA gains across key blocks of its upcoming flagship CPU, using designs the model wrote.
On Jalapeño, OpenAI's first custom inference chip, OpenAI says it began using GPT-4o in early 2025, and that the team estimates model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit. It says those optimisations were silicon proven during bring-up. One multiplier redesign used about 15% less cell area in synthesis and cut latency from two clock cycles to one.
OpenAI also tested successive models on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, across 9,881 scored runs in the full evaluation. Next, OpenAI is working with Synopsys to train GPT-Synopsys, a model meant to use Synopsys's design tools the way expert chip designers do.
Claims and checks#
Everything above is OpenAI's account. Su's post endorses the work but adds no figures. The Next Web reports that OpenAI's own chip chief, Richard Ho, said last month he still does not fully trust AI design changes and so checks every one; we did not find the original remark.
OpenAI is candid about its RISC-V test. When models designed a core from written requirements, all five scored higher than when they tried to improve an existing design, but OpenAI says the models could leave out features of the original cores, which may explain part of the advantage. It says requiring those features, setting a tighter area limit and checking timing after physical synthesis would make the test more representative of production chip design.
The AMD result also sits beside a commercial tie: TensorWave says it will supply OpenAI 457 MW of AMD GPU capacity, none expected online before the first half of 2027. Nothing in OpenAI's post links the two.
Limits#
- Every AMD figure is OpenAI's account. We found no AMD-published numbers or methodology, and Su's post gives none.
- "Up to 14%" is a best case. The post does not break PPA into its three measures or describe the baseline in detail.
- That some changes are in a design sent for manufacturing is OpenAI's statement; we have not seen the design or an AMD confirmation.
- The RISC-V comparison covers OpenAI's own models only, and OpenAI says the from-specification task was easier.
- The Jalapeño figures are OpenAI's team estimates.
Sources#
- OpenAI: Advancing chip design with OpenAI models (November 18, 2026)
- OpenAI on X: AMD and Jalapeño chip-design results (November 18, 2026)
- Lisa Su on X: Proud of these fantastic results (November 18, 2026)
- The Next Web: OpenAI says its AI found 14% gains in AMD’s flagship GPU design (November 18, 2026)
- The Decoder: OpenAI says its models delivered a year’s worth of chip optimization on top of AMD’s own engineering (November 18, 2026)