
Illustration
Companies3 min read
OpenAI says its models improved AMD's flagship GPU design by up to 14%
OpenAI's models improved the power, performance and area of AMD's flagship GPU design by up to 14% over heavily optimised baselines, OpenAI said on 18 November. AMD has already included some of the changes in a recent design it sent for manufacturing, the company added.
Key facts
- On 18 November 2026 OpenAI reported gains of up to 14% in power, performance and area on AMD's flagship GPU design.
- On OpenAI's own Jalapeño chip, model-proposed changes cut cell area by about 13% (vector unit) and 10% (matrix unit), OpenAI said.
- In 9,881 runs on RISC-V cores GPT-6.2 Astra scored highest, according to OpenAI; GPT-Synopsys comes next.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
The model did this by rewriting the code that describes the chip's circuits (SystemVerilog RTL), according to OpenAI's post. Each rewrite was checked with tools from chip-design software maker Synopsys to confirm the block still behaved the same and to measure the trade-offs in power, performance and area. OpenAI calls it “roughly a year of progress” achieved in under five months.
AMD added other changes to its production design files for future versions, according to OpenAI. AMD also verified single-digit percentage gains in key blocks of its upcoming next-generation flagship CPU using designs written by the model. All figures come from OpenAI's post; AMD had not published its own account when this article went live. The two companies are closely tied: under an October 2025 agreement, AMD is to supply OpenAI with 6 gigawatts of GPUs, and OpenAI holds a warrant for up to 160 million AMD shares, AMD said at the time.
OpenAI's Jalapeño chip as the first testbed
OpenAI says it started with Jalapeño, its own inference chip unveiled in August. Even GPT-4o proposed useful improvements as early as the start of 2025, OpenAI said. Across the project, the team estimates that model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit. One redesigned multiplier used about 15% less area while cutting latency from two clock cycles to one. The gains were proven in silicon, OpenAI said.
To track progress across model generations, OpenAI asked its models to speed up two open-source RISC-V cores, PicoRV32 and Ibex, in 9,881 scored runs. GPT-6.2 Astra had the highest average score on both; on PicoRV32 at about 112,000 tokens per run it scored roughly twice as high as GPT-6 Astra. When asked to design a core from written requirements, all five models tested scored higher than when improving the existing design. OpenAI notes that the models could drop features such as performance counters in that task, and that tighter area limits and timing checks would make the test more realistic.
Next, OpenAI is training a dedicated model with Synopsys, GPT-Synopsys, meant to operate electronic design automation (EDA) tools the way expert chip designers do. The two companies announced the partnership on 30 September; engineers set the goals and review and approve the output, The Decoder reported. The model that led the chip tests, GPT-6.2 Astra, launched only on 17 November.
Analysis
The 14% figure is OpenAI's account of work with a supplier in which OpenAI can itself acquire a stake. More telling than the number is a detail in the report: every rewrite went through Synopsys verification tools before AMD adopted it. Chip design has had automatic checkers for decades that reliably tell whether a circuit still does the same thing. That lets a model try many variants, and only verified ones reach manufacturing. Security showed a similar pattern this week: in the Java4Shell case, AI made finding the flaw cheap, but the find only paid off once someone verified it and owned the fix.
For companies outside the chip industry, this suggests a sober selection rule. AI agents pay off most where a defined check scores the result, such as tests in software development, calculation rules in accounting or approval criteria for ads. Without that step, the model judges its own work, which it still does poorly according to Epoch's test with real office work. Teams planning a first agent project should first define how a good result is recognised and who signs it off, and only then pick the model. What remains open is how much of AMD's gain comes from the model and how much from the engineers who chose which proposals to keep.
Got AI news? Submit news →
FAQ
What did OpenAI do on AMD's chip design?
According to OpenAI, its models rewrote the circuit code (RTL) of AMD's flagship GPU and improved power, performance and area by up to 14%. Every change was verified with Synopsys tools.
What is GPT-Synopsys?
A model OpenAI is training with Synopsys to operate chip-design (EDA) software like expert engineers. The partnership was announced on 30 September 2026; engineers review and approve the output.
Source
OpenAI: Advancing chip design with OpenAI models
More sources
- The Decoder – Ankündigung GPT-Synopsys am 30.09.2026, läuft auf OpenAI-Infrastruktur, Ingenieure prüfen und geben frei
- AMD – Liefervertrag 6 Gigawatt vom 6.10.2025, Optionsschein auf bis zu 160 Millionen AMD-Aktien



