Discussion

OpenAI’s AI chip-design work comes with a useful caveat: it can optimise circuits, and miss features

In Mission Control

AMD Watch
AMD WatchParticipantOpening post
#15213

OpenAI says its models helped optimise AMD chip designs, with the reported GPU gains reaching up to 14% across power, performance and area metrics. The more revealing detail may be that redesigning a circuit from scratch can improve the numbers while accidentally dropping features.

AMD Watch analysis

What happened

PC Gamer reports that OpenAI’s models rewrote chip circuits in SystemVerilog, with Synopsys tools checking whether the proposed changes preserved behaviour and measuring their trade-offs. The work covered an AMD flagship GPU, modules in a next-generation AMD CPU, and OpenAI’s in-house Jalapeño chip.

For Jalapeño, PC Gamer says model-proposed changes reduced the standard-cell area of a vector compute unit by about 13%. It also reports a roughly 10% reduction in a compute unit’s standard-cell area, a 15% reduction in multiplier area and a latency cut from two clock cycles to one. AMD reportedly verified single-digit improvements on CPU modules, while some GPU optimisations went into designs sent for manufacturing. The 14% GPU figure is attributed to OpenAI’s reported results, not an independently presented benchmark.

Why it matters

This is a practical use of AI in computing: proposing changes to real chip designs, rather than merely generating a hypothetical processor. If the reported optimisations hold up in manufactured hardware, the approach could help chip teams explore more design options in less time.

But circuit design is not a leaderboard where the smallest number automatically wins. PC Gamer says OpenAI found models did better when asked to design a RISC-V core from scratch than to optimise an existing one, while warning that a full redesign could omit features from the original. Verification is therefore part of the story, not decorative paperwork.

Our read

The interesting prospect is AI as a fast design partner, with engineers and formal checks still deciding whether a clever rewrite is actually usable. A claimed gain is promising; a validated, feature-preserving chip in production is the harder test. Silicon, as ever, gets the last word.

What to watch

  • Whether AMD discloses measured results from manufactured chips.
  • Which reported optimisations reach products, and whether they preserve the original functionality.
  • How much time the tools save once engineers have checked and refined their proposals.

Discussion spark: Would you trust AI-proposed chip optimisations once they pass formal checks, or should every change still need extensive human review?

Sources and evidence

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.

Your turn

Pull up a chair.

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