OpenAI Says Its Models Improved AMD’s Flagship GPU Design by Up to 14%

OpenAI AMD chip design work improved a flagship GPU by up to 14%, the company said, and AMD has already sent some of those changes out for manufacturing.

Share this article

Featured image: AMD CEO Lisa Su speaking at Oak Ridge National Laboratory on May 7, 2019, in a file photo that is not from OpenAI’s Nov. 18, 2026, announcement. Photo: Genevieve Martin, OLCF at ORNL, Wikimedia Commons, CC BY 2.0. Cropped to 16:9.

OpenAI said Wednesday, Nov. 18, 2026, that its models improved AMD’s flagship GPU intellectual property by up to 14% on power, performance and area against heavily optimized baselines, a result the company described as part of its OpenAI AMD chip design work. Some of those changes, OpenAI said, have already been included in a design AMD sent for manufacturing, with others added to production design files for later versions.

Advertisement
AMD GPU IP, PPA
Up to 14%
vs. heavily optimized baselines, OpenAI said
Time to those gains
Under 5 months
OpenAI compared that with about a year of conventional work
Jalapeño multiplier
About 15%
less cell area, and latency cut from two cycles to one, OpenAI said
AMD next-gen CPU
Single-digit %
PPA gains on key blocks, OpenAI said AMD verified

The claims come from OpenAI, which is both the model developer and a partner of AMD’s. They have not been independently audited in the materials OpenAI published, and AMD did not issue a separate, detailed confirmation of the 14% figure in the documents reviewed for this report. Tom’s Hardware, writing on Nov. 20, similarly treated the percentages as OpenAI’s figures.

What OpenAI says the models changed

OpenAI said its models rewrote SystemVerilog RTL, the code that describes a chip’s circuits, and used Synopsys electronic design automation tools to check that each rewrite preserved the block’s behavior and to score the trade-offs in power, performance and area. The company said the model can rewrite circuits that perform calculations, control operations or move data, and that it “reason[ed] about RTL at the level of an expert microarchitect.”

On AMD’s upcoming next-generation flagship CPU, OpenAI said AMD verified single-digit percentage PPA gains across key blocks from designs written by the model. The company said it expects further gains on top of those GPU and CPU baselines. It did not name the GPU or CPU products, and it did not publish the baseline designs or the full set of verification results.

The AMD work sits alongside OpenAI’s own custom inference chip, Jalapeño. OpenAI said in August that engineers used its frontier models to speed that project, and Wednesday’s post offered more detail. The company said it began using GPT-o3 mini on Jalapeño in early 2025. Across the broader design effort, the team estimated that model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and that those optimizations were “silicon proven” during Jalapeño’s bring-up.

One multiplier redesign, OpenAI said, used about 15% less cell area in synthesis and reduced latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while keeping two-cycle latency. Those are OpenAI’s measurements of its own chip, not AMD’s.

By the numbers

The 14% AMD GPU figure and the Jalapeño area estimates are not the same kind of measurement. Power-performance-area is a combined trade-off; cell area is one component of a block. OpenAI presented them as separate results, and this chart keeps them separate rather than adding them together.

Horizontal bar chart of OpenAI-reported chip design improvements: Jalapeño vector unit cell area about 13 percent, matrix unit about 10 percent, a multiplier redesign about 15 percent, and AMD flagship GPU IP power-performance-area up to 14 percent
Source: OpenAI, “Advancing chip design with OpenAI models,” Nov. 18, 2026. The bars use different metrics and are not directly comparable. Chart: The Boston Report.

To compare models against one another, OpenAI also tested successive systems on two open-source 32-bit RISC-V cores, PicoRV32 and OpenTitan’s Ibex. With a comparable budget of about 112,000 tokens per run, OpenAI said GPT-6.2 Astra had the highest average score on both cores when the task was to raise throughput on an existing design. On PicoRV32, OpenAI said Astra’s score at that budget was about twice GPT-6 Astra’s. A second task asked models to design a core from a written specification; OpenAI said all five models scored higher on that task than when they tried to improve the existing design. The full evaluation, the company said, included 9,881 scored runs. Designs that failed checks, or that used more than 16 times the original core’s area, scored zero.

Advertisement

OpenAI itself noted limits in that test. Starting from written requirements can let a model leave out features in the original cores, such as performance counters, that may not help the measured workloads. Tighter area limits and timing checks after physical synthesis, it said, would make the test closer to production chip design.

The Synopsys partnership

OpenAI said it is working with Synopsys to train GPT-Synopsys, a model meant to use Synopsys tools the way expert designers do. That partnership was announced on Sept. 30, 2026, in a joint Synopsys-OpenAI release that described a multi-year agreement, a shared revenue framework and early technology engagements with semiconductor customers. Synopsys said customer design data is not used to train the model. Wednesday’s post presented GPT-Synopsys as the next step for evaluating and broadening the design work, not as a newly launched product.

For more on how companies are using AI in product development, see The Boston Report’s AI news hub.

Related: OpenAI Says Coding Agents Now Finish 46% of Two-to-Four-Hour Research Tasks on Their Own; California Panel Urges Emergency Power to Shut Down Dangerous AI Systems.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up