OpenAI says its AI models rewrote parts of AMD's flagship graphics chip design and made it better — up to 14% better on power, performance and area, the three measures chip designers care about most. OpenAI says that amounts to roughly a year of progress in under five months.
Every figure here is vendor-reported: AMD has published no figures of its own, and none has been independently reproduced. Here is what OpenAI says happened, how it works, and where the limits are.
What OpenAI says happened
On 18 November 2026, OpenAI published a blog post titled "Advancing chip design with OpenAI models". Its frontier models rewrote SystemVerilog RTL — the code describing a chip's circuits — for AMD hardware, delivering up to 14% PPA improvement on AMD's flagship GPU IP, against heavily optimised baselines.
AMD has used some changes in a design recently sent for manufacturing, and added others to production design files. The work also covered AMD's next-generation flagship CPU, where AMD verified single-digit percentage PPA gains across key blocks. No exact number was given for the CPU work.
PPA, RTL and tape-out, in plain words
PPA stands for power, performance and area: how much energy a chip uses, how fast it works, and how much silicon it takes up. Improving all three at once is hard, since faster often means bigger or hungrier — so 14% PPA means a meaningfully better balance across all three.
RTL stands for register-transfer level, written in a language called SystemVerilog — the detailed blueprint of a chip, describing its circuits and how data moves between them. Tape-out is the moment a finished design is sent to the factory to be made: the point of no return.
Jalapeño, OpenAI's own chip
The same approach was used on Jalapeño, OpenAI's first custom inference chip, built for running AI models rather than training them. Work began in early 2025, with GPT-o3 mini helping design it, and the design reached tape-out in nine months.
OpenAI reports that model-proposed changes cut standard-cell area — the basic building blocks of a chip's logic — by about 13% in the vector unit and 10% in the matrix unit. The gains were silicon-proven during bring-up, when a newly manufactured chip is first tested, so they held up on real hardware.
One example is a multiplier redesign: the new version used about 15% less cell area in synthesis while cutting latency from two clock cycles to one.
How the AI actually redesigns a chip
According to OpenAI, the model reasons about RTL like an expert microarchitect. It restructures computation, reuses hardware instead of duplicating it, and reroutes data paths so information travels more efficiently.
Nothing goes straight into a chip. Synopsys EDA tools — electronic design automation software — verify each rewrite behaves identically to the original module, then measure the power, performance and area trade-off.
OpenAI is now training GPT-Synopsys with Synopsys, a model designed to drive EDA tools directly, under a multi-year agreement announced on 30 September 2026.
What to be careful about
First, all the figures are vendor-reported, with no independent reproduction yet, and AMD has published no figures of its own — a claim to be tested, not settled fact.
Second, OpenAI itself reports limits. The models scored better when designing from written requirements than when improving existing code, and they could drop features such as performance counters, the small circuits that track how a chip behaves. Timing was also not checked after physical synthesis.
Third, OpenAI's chip chief, Richard Ho, said in October 2026 that he does not yet fully trust automated circuit changes, and every proposed change is checked. Even its builders do not treat its output as ready to use unchecked.
Key takeaways
- OpenAI says its models rewrote SystemVerilog RTL and delivered up to 14% PPA improvement on AMD's flagship GPU IP, against heavily optimised baselines.
- OpenAI calls it roughly a year of progress in under five months; AMD has used some changes in a design recently sent for manufacturing.
- For AMD's next-generation flagship CPU, single-digit percentage PPA gains were verified across key blocks, with no exact number given.
- On Jalapeño, OpenAI's first custom inference chip, model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, proven on real silicon.
- The design reached tape-out in nine months; one multiplier redesign used about 15% less cell area and cut latency from two clock cycles to one.
- All figures are vendor-reported and unreproduced; models could drop features such as performance counters, and timing was not checked after physical synthesis.
Conclusion
If the figures hold up, this is a genuine shift: AI models finding improvements experienced engineers missed, on flagship hardware, in months rather than a year. But the proof that matters most is still to come: independent reproduction and AMD's own figures.
FAQ
What does a 14% PPA improvement mean?
PPA stands for power, performance and area — how much energy a chip uses, how fast it works, and how much silicon it takes up. OpenAI reports up to 14% on AMD's flagship GPU IP, against heavily optimised baselines, but the figure is vendor-reported and unreproduced.
What is Jalapeño?
Jalapeño is OpenAI's first custom inference chip, built for running AI models rather than training them. Work began in early 2025 with GPT-o3 mini helping design it, and the design reached tape-out in nine months. OpenAI says model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit.
Can AI design chips on its own now?
Not on its own. OpenAI's models propose code rewrites, Synopsys EDA tools verify each one behaves identically to the original, and engineers check every proposed change. OpenAI chip chief Richard Ho said in October 2026 that he does not yet fully trust automated circuit changes.
Sources
- OpenAI — Advancing chip design with OpenAI models
- Data Center Dynamics — OpenAI claims its models have delivered chip design improvements
- Silicon Report — OpenAI says its models improved AMD flagship GPU designs
- Daily Tech & AI — NVIDIA Q3: Five Numbers That Matter
- Daily Tech & AI — AI Tools Collection
Researched with AI assistance and reviewed for accuracy.
Follow on WhatsApp
Follow Daily Tech & AI on WhatsApp for one tech and AI update every day: Daily Tech & AI WhatsApp channel.
Comments
Post a Comment