AI Chip Design: What OpenAI and AMD Actually Showed

Researcher examining a reflective silicon semiconductor wafer
A researcher inspects a silicon wafer used in semiconductor fabrication.

OpenAI says its newest models improved real AI chip design blocks at AMD and produced working changes that were sent toward manufacturing. The results are notable, but they come from a company case study rather than an independently replicated benchmark.

In a November 18 technical report, OpenAI described AI chip design assistance across standard cells, CPU and GPU intellectual-property blocks, and open RISC-V processors. AMD collaborated on part of the work.

What OpenAI and AMD report about AI chip design

For a standard-cell project called Jalapeño, OpenAI says the model-assisted workflow reduced area by about 13% for one vector cell and 10% for one matrix cell while preserving required behavior. The company says the cells were fabricated and validated in silicon.

On AMD blocks, the report claims improvements of up to 14% in combined power, performance and area against highly optimized flagship GPU intellectual property. It also reports single-digit gains on upcoming CPU blocks. OpenAI says some resulting changes entered designs sent for manufacturing.

Those statements should be attributed precisely. OpenAI published the results; AMD’s detailed independent validation, test methodology and full block-level data were not provided in the public report. “Up to 14%” describes the best reported case, not an average across an entire processor.

Why power, performance and area are linked

Chip designers rarely optimize a single number. A smaller circuit may use less power, but a change that slows timing can make it unusable. A faster block may require more area or energy. Engineers therefore evaluate trade-offs among power, performance and area—often shortened to PPA—while enforcing functional correctness.

Models can help by proposing code changes, running design tools, inspecting errors and trying again. The design tools remain essential: they synthesize hardware descriptions, simulate behavior and measure whether a proposal meets constraints. An eloquent answer is worthless if the resulting circuit is wrong.

The open RISC-V evaluation adds a reproducible test

OpenAI also evaluated models on the PicoRV32 and Ibex open-source RISC-V processors. Across 9,881 scored runs, candidate designs received no credit if they failed correctness checks or used more than 16 times the original area. At a budget of roughly 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score in the reported comparison.

The evaluation is more transparent than a private customer case because the processors are open. It is still a bounded benchmark. Simulation does not capture every physical implementation issue, and the report notes that its specification-from-scratch task can omit features. It also did not use full physical-synthesis timing for every comparison.

What the results mean—and what they do not

The evidence supports a narrower claim than “AI designs chips.” Models can assist skilled engineering teams on selected optimization tasks when connected to compilers, simulators and verification tools. Human engineers define constraints, review changes and remain responsible for correctness.

The report does not show that a model can independently produce a competitive modern GPU. Full chips contain many interacting blocks, verification requirements, physical-layout constraints and manufacturing risks. Improvements on selected modules may not translate linearly to a whole processor.

OpenAI also announced GPT-Synopsys work with electronic-design-automation company Synopsys. That partnership could move model assistance closer to standard engineering workflows, but product availability and real-world adoption will be separate tests.

Future & Fact’s overview of AI infrastructure explains why accelerator architecture, memory, networking and data-center systems all affect model performance.

The most useful next evidence would be independent replication: fixed design problems, disclosed tool versions, comparable engineering time, complete correctness checks and results across many blocks rather than selected successes.

The Future & Fact newsletter

Fresh discoveries.
Delivered to your inbox.

Explore AI, science and space with clear explanations and stories grounded in primary sources.

Coming soon — signups are not open yet.

Read our Privacy Policy for information about how we handle personal data.