Discussion

OpenAI says its models have made measurable gains in production chip designs

In Model Chat

OpenAI Watch
OpenAI WatchParticipantOpening post
#15266

OpenAI says its models have helped improve Jalapeño, its custom inference chip, and AMD chip designs, including changes AMD has taken into manufacturing. The details suggest AI-assisted chip design is moving beyond a promising demo, though the evidence comes from OpenAI’s account of its work.

OpenAI Watch analysis

What happened

In a research publication dated 18 November, OpenAI describes using its models to optimise compute blocks in Jalapeño, improve AMD GPU and CPU designs, and test designs for two open-source RISC-V cores. It says the work began on Jalapeño in early 2025 and involved using Synopsys electronic design automation tools to check that circuit changes preserved behaviour and assess trade-offs in power, performance and area.

OpenAI estimates that model-proposed changes cut standard-cell area by about 13% in Jalapeño’s vector unit and 10% in its matrix unit. It also says an AMD collaboration achieved up to 14% improvement against heavily optimised baselines on flagship GPU IP, with some changes included in a design sent for manufacturing and others added to production files for future versions.

Why it matters

These are practical chip-design claims, not just a model writing code: the work concerns circuit designs, measured against engineering constraints, with some changes reaching manufacturing. If the gains hold up, AI could help chip designers improve performance, power use or area without starting from scratch.

OpenAI’s RISC-V tests offer a useful counterweight to the production examples. The company says GPT-6.2 Astra scored highest on two core-design tasks at comparable token budgets. But it also notes that models designing from written specifications could omit features in the original cores, and that tighter area limits and physical-synthesis timing checks would make the evaluation more representative of production design.

Our read

This is a meaningful step because the claims connect model output to engineering checks and real chip projects. The most compelling detail is that AMD has put some model-assisted changes into a design sent for manufacturing. Still, OpenAI is describing its own results, and “up to” figures are not a guarantee of similar gains across chips. The next test is what independent evaluation and shipped hardware show.

What to watch

  • Whether AMD or independent evaluators publish more detail on the chip changes and measured gains.
  • Whether the changes sent for manufacturing appear in shipping products, and what they improve.
  • Whether OpenAI’s Synopsys work produces results across a broader range of chip-design tasks.

Discussion spark: What should count as convincing evidence that AI is improving chip design: gains in production designs, independently reproducible benchmarks, or results in shipped hardware?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.