OpenAI’s AI Redesigned AMD’s Flagship GPU and Squeezed Out 14% More Performance

OpenAI’s models are no longer just writing code. They are redesigning the silicon that runs them. In a technical report published on November 18, “Advancing chip design with OpenAI models”, the company says its frontier models rewrote the circuit code of AMD’s flagship GPU and improved it by up to 14% — against a baseline AMD’s own engineers had already tuned hard.
OpenAI calls that roughly a year of design progress compressed into less than five months. And this is not a lab demo: AMD has already put some of the model-proposed changes into a chip design it sent for manufacturing, and written the rest into production design files for future products.
What OpenAI actually did
The models worked at the circuit level, rewriting SystemVerilog RTL — the code that describes a chip’s actual logic — rather than just suggesting tweaks. Each rewrite had to survive checks with Synopsys electronic design automation tools, which confirmed the rewritten block still behaved exactly as specified and then measured the trade-offs between power, performance and area. Engineers call that combined measure PPA.
OpenAI says its models can now reason about RTL the way an expert microarchitect does: restructuring computation, reusing hardware and rerouting data paths. The headline result is the up-to-14% PPA gain on AMD’s flagship GPU intellectual property. AMD also verified single-digit percentage PPA gains across key blocks of its next-generation flagship CPU using designs the models wrote.
The work builds on OpenAI’s own chip project. The company first said in August that its models had helped design Jalapeño, its first custom inference chip. It began using GPT-4o on that project in early 2025, and model-proposed changes cut standard-cell area by about 13% in the chip’s vector unit and about 10% in its matrix unit — numbers measured in real silicon, not simulation. One multiplier redesign used about 15% less cell area while cutting latency from two clock cycles to one.
Why 14% matters — and what it does not mean
A 14% PPA gain on an already heavily optimized design is a big deal in chip engineering, where single-digit gains normally take months of expert work. Squeezing that out of AMD’s flagship GPU IP is why OpenAI frames it as a year’s progress in under five months.
But the number needs its caveats, and OpenAI states them. First, 14% is an “up to” figure on specific GPU IP, not an improvement across a whole chip. Second, OpenAI named no AMD product and gave no breakdown of the gain between power, speed and area. Third, it reported no measurements from finished AMD silicon — the gains are measured in the design flow, and the chips carrying them are still on their way to manufacturing.
Fourth, and most important, all of this is vendor-reported. Neither OpenAI’s figures nor AMD’s verification have been independently reproduced by a third party. The results are still striking, but they are claims from the companies involved, not settled fact.
Jalapeño, OpenAI’s own chip
The AMD work gets the headlines, but Jalapeño may matter more for OpenAI itself. The company designed its first inference chip with AI assistance from the start, moving from initial design to tape-out in nine months. For a company that spends heavily on compute, designing its own silicon — with its own models doing part of the design — is a direct way to cut costs and reduce its dependence on outside suppliers.
OpenAI also ran 9,881 scored tests across two open-source 32-bit RISC-V processors, PicoRV32 and Ibex, with models spanning several generations up to the newly released GPT-6.2 Astra. The newer models did better, which suggests the capability is still improving fast rather than flattening out.
What comes next with Synopsys
OpenAI is not stopping at one-off projects. It has partnered with Synopsys, the electronic design automation giant, to train a new model called GPT-Synopsys that can use Synopsys’s tools the way expert chip designers do. The deal, announced on September 30, is a multi-year preferred partnership with joint go-to-market plans and revenue sharing.
The direction is clear: AI models are moving from helping with chip design to doing parts of it directly, with professional tools checking every change. If the AMD results hold up in finished silicon, chip design could be one of the first industries where AI does expert-level engineering work at scale — and the chips it designs will run the next generation of AI models. The loop is closing fast.
The chip world is moving quickly on all fronts — NVIDIA’s Rubin HPC GPU arrived just this week, and NVIDIA ended H200 chip sales in China days ago. AI is now reshaping both sides of the silicon business: who buys the chips, and how they get designed.
