OpenAI’s models rewrote AMD’s GPU circuits — and found a year’s worth of design gains in under five months
OpenAI’s models are now doing the kind of chip design work that used to belong to a human microarchitect — and the results are landing in real silicon. In a research update published Tuesday, the company says its models rewrote the circuits of AMD’s flagship GPU design for up to 14% better power, performance and area, a gain OpenAI calls roughly a year of progress achieved in less than five months. AMD has already folded some of the changes into a design it sent for manufacturing.
- The headline number: up to 14% improvement in PPA — power, performance and area — on AMD’s flagship GPU intellectual property, measured against baselines AMD had already heavily optimized.
- The timeline: OpenAI calls it roughly a year of design progress in less than five months. Some changes are already in a design AMD sent for manufacturing; others went into AMD’s production design files for future versions.
- The CPU side: AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by OpenAI’s model.
- OpenAI’s own chip: on Jalapeño, its first custom inference chip, model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit — confirmed in silicon during bring-up.
- The catch: OpenAI gives no breakdown of the 14% between power, performance and area, names no AMD product, and reports no measurements from finished silicon. AMD has published no statement of its own on the work.

The loop is now closed. The chips that train and run frontier AI models are themselves being designed, in part, by frontier AI models — and the work is no longer a lab demo. OpenAI’s update, published November 18, describes model-proposed changes that made it into production hardware at two companies: its own Jalapeño inference chip and AMD’s flagship GPU and CPU lines.
The AMD result is the one that travels. OpenAI says it partnered with AMD on power, performance and area — PPA, the three-way trade-off chip designers live by — and saw up to 14% improvement on the flagship GPU intellectual property. The model worked at the RTL level, rewriting the SystemVerilog code that describes a chip’s circuits before they are laid out physically. Each rewrite was checked with Synopsys tools to confirm the changed block still behaved as the original, then scored on the PPA trade-offs.
What the model actually did#
OpenAI describes the model reasoning about RTL the way an expert microarchitect would — the designer who decides how a processor block is organized inside — exploring alternative logic and different ways of dividing computations across pipeline stages. The point is not that the model wrote code faster; it found better circuits than the human-optimized baselines it started from.
That claim is what makes the manufacturing detail matter. According to OpenAI, AMD has already included some of the model’s changes in a recent design it sent for manufacturing, and added others to its production design files for future versions. A 14% figure on a slide is one thing; a change that survives AMD’s own verification and lands in a tape-out is another. The CPU result matters too: AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs the model wrote.

Jalapeño: where the loop started#
The AMD work grew out of OpenAI’s own chip. Jalapeño, its first custom inference processor, built with Broadcom, was the test bed: OpenAI says it started using GPT-o3 mini on the design in early 2025, and even that early model found nontrivial improvements.
The numbers OpenAI gives for Jalapeño are more concrete than the AMD ones, because they are its own silicon. Across the broader design effort, the team estimates model-proposed changes reduced standard-cell area by about 13% in the vector unit, which manages parallel computing tasks, and 10% in the matrix unit. One multiplier redesign used about 15% less cell area and cut latency from two clock cycles to one. These were not simulations: OpenAI says the optimizations were proven in silicon when Jalapeño was brought up in the lab. The company first described the Jalapeño work in August; Tuesday’s post fills in the engineering behind it.
How good are the models, really?#
To measure whether each generation is actually getting better, OpenAI tested successive models on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex, asking them to increase throughput while keeping the designs correct. The full evaluation ran 9,881 scored runs across both cores and both tasks. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol on both — evidence, OpenAI argues, that the capability is improving with each generation rather than being a one-off.
OpenAI is candid about one awkward result: every model scored higher when it designed a core from written requirements than when it tried to improve the existing design. Starting from a specification gives the model room to leave out features the original cores carry — performance counters, extra instructions — that do little for throughput on the test workloads. OpenAI says requiring those features would make the test more representative of production chip design — a useful admission that the benchmarks flatter the models.

The Synopsys partnership#
The last piece points at where this goes next. OpenAI is partnering with Synopsys, whose electronic design automation tools are the industry standard, to train GPT-Synopsys — a model built to use those tools the way expert chip designers do. The Jalapeño and AMD work already leaned on Synopsys software for verification; the partnership would bake that workflow into the model itself.
What to watch#
First, the caveats: no breakdown of the 14% between power, performance and area, no named GPU or CPU products, and no measurements from finished AMD silicon — so the figure should not be read as a 14% faster graphics card. AMD has published no statement of its own, so the account rests on OpenAI’s telling until AMD ships a product carrying the changes, or talks about them itself.
Second, the race: every frontier lab now has reason to point its models at silicon, and whoever gets a design-assistant model into production EDA workflows first sets the terms for the rest.
Third, the feedback loop itself. Jalapeño was designed with model help, and it will run the models that design its successor. That is the story underneath the percentages: the industry that builds AI’s hardware is starting to be built by AI — and the question is whether the verification tools, and the humans who sign off on a tape-out, can keep up.
Sources: OpenAI, “Advancing chip design with OpenAI models” (November 18, 2026); FreshSynapse, “OpenAI’s Models Redesigned AMD’s Flagship GPU — 14% Better in Under Five Months”; DirtyStatic, “OpenAI Models Improve AMD Chip Designs”.