OpenAI says AI found up to 14% gains in AMD GPU designs, with edits sent to manufacturing
The research describes circuit rewrites checked with established design tools. Its separate processor tests show stronger models—but also how dropping features can improve benchmark scores.
OpenAI’s chip-design work is moving beyond code generation into hardware changes that appear in AMD’s design sent for manufacturing and files for future versions. In its November 18 publication, OpenAI reports up to 14% power, performance and area improvement in selected GPU circuits—not whole-GPU speed—and single-digit gains in blocks of AMD’s next flagship CPU. Initial tests of OpenAI’s Jalapeño silicon provide separate evidence of area reductions, while a processor benchmark’s results are qualified because models could omit features. OpenAI and Synopsys are developing GPT-Synopsys for electronic design automation workflows.
01
Models rewrote SystemVerilog RTL, with Synopsys tools checking functional equivalence and evaluating power, performance and area tradeoffs.
02
OpenAI estimates Jalapeño’s vector and matrix units use about 13% and 10% less standard-cell area, respectively, based on initial silicon testing.
03
Across 9,881 processor-design runs, GPT-6.2 Astra had the highest average score on both tested cores at comparable budgets of about 112,000 tokens.
OpenAI says its models found further improvements in AMD chip designs that engineers had already heavily optimized—and some edits are now in a design sent for manufacturing. Its November 18 research publication reports up to 14% gains in power, performance and area, a circuit-design result rather than a claim of 14% faster whole GPUs.
Rewriting circuits, then checking the tradeoffs
The AMD work starts with SystemVerilog RTL: code describing a chip’s circuits. OpenAI’s model rewrote that code to change how circuits calculate, control operations or move data. It used Synopsys software to check that each rewrite preserved the block’s behavior, then evaluate the tradeoffs in power, performance and area, collectively called PPA.
OpenAI describes the GPU improvement as roughly a year of progress achieved in less than five months. That is the company’s comparison, not an independently measured engineering timetable. Alongside changes included in the design sent for manufacturing, AMD added others to its production design files for future versions, according to OpenAI.
The collaboration also extends to CPUs. OpenAI says AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using model-written designs. These are improvements to specified parts of the designs; the publication presents no corresponding whole-processor performance figure.
For evidence from tested silicon, OpenAI points to Jalapeño, its custom inference chip. The company estimates that model-proposed changes reduced standard-cell area—the area occupied by circuit building blocks—by about 13% in its vector unit and 10% in its matrix unit. It says those optimizations were proven during initial silicon testing.
Editorial illustration for OpenAI says AI found up to 14% gains in AMD GPU designs, with edits sent to manufacturing.
The Jalapeño examples show different kinds of gains. Models explored alternative logic and ways to divide calculations across pipeline stages, sharing hardware and reducing intermediate storage. One matrix-unit multiplier became both smaller and lower-latency. A separate adder used about 15% less cell area while retaining its existing two-cycle latency.
A processor benchmark with room to leave features out
To compare model generations, OpenAI ran a separate evaluation using PicoRV32 and Ibex, two open-source 32-bit RISC-V processor cores for embedded, microcontroller-class work. The evaluation covered 9,881 scored runs across two tasks and multiple budgets for model output. It tested both revising an existing implementation and starting from requirements.
Improve an existing core: models received its original hardware description and tried to increase instructions executed per second.
Design from a specification: models received written requirements and connection templates, then chose how data moved and work was divided into stages.
A design scored zero if it failed a simulation check or exceeded 16 times the original core’s area. Passing designs were assessed on hidden workloads. OpenAI combined instructions per clock cycle with an estimated maximum clock frequency to calculate throughput, then normalized scores against the best observed result for each core.
At comparable budgets of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. All five tested models scored better when designing from written requirements than when improving existing code. But that apparent advantage comes with an important qualification: models could leave out features present in the originals.
Those omissions could include performance counters and additional instructions that contributed little to the tested workloads. OpenAI says that freedom may explain part of the score advantage. Requiring those features, tightening the area limit and checking timing after physical synthesis would make the evaluation more representative of production chip design, the company acknowledges.
Taking the workflow into design software
OpenAI’s stated next step is to ground model development in chip-design evaluators and expand the challenges it addresses. It is working with Synopsys to train GPT-Synopsys to use electronic design automation tools—the software engineers use to design and verify chips. That work targets the tool-based workflow behind the reported circuit improvements, not just the ability to generate hardware code.
Sources
openai.comAdvancing chip design with OpenAI models
Reader comments
Newest comments first. Replies stay oldest first.