Advancing chip design with OpenAI models

OpenAI’s claim that its models can advance chip design, including partnerships with AMD and Synopsys and RISC‑V CPU experiments, is met with skepticism over vague metrics, lack of released designs, and what some see as IPO‑driven marketing. Commenters debate how well large models can really handle RTL, floorplanning, and verification given scarce high‑quality training data, and whether any productivity gains will be outweighed by the persistent bottleneck of validation. Many also raise concerns about IP exposure and the long‑term risk of AI vendors using insight from customers’ proprietary designs to compete with them directly.

Perception of the announcement / marketing context

  • Many see the post as pre-IPO hype and “marketing noise” rather than substantive progress.
  • Others argue marketing is a legitimate part of competition and that rivals are also boosting media presence.
  • Some say these kinds of OpenAI-authored case studies would be more convincing if third‑party companies independently reported results.

Skepticism about metrics and claims

  • Several commenters criticize use of vague metrics like “mean normalized reward” without clear definitions.
  • The AMD claim of “single-digit PPA gains” is viewed as hedged: limited to “key blocks” and “designs written by our model,” leaving room for minimal or accidental benefit.
  • Lack of released RISC‑V designs or concrete y‑axis units is taken by some as evidence that results may not be impressive.

Can LLMs really design chips? Data, methods, and scope

  • Debate over feasibility: some think RTL optimization and logic-level work are highly amenable to LLMs + RL; others insist there’s nowhere near enough high‑quality RTL and layout data.
  • Distinction is made between:
    • RTL / logic synthesis (where tools and metrics are mature, and LLMs already seem useful), and
    • Floorplanning, routing, analog/RF, and advanced ASIC work, which are data-poor, highly spatial, and still largely manual.
  • Past systems like DEC’s SID and modern EDA tools already automate large fractions of logic synthesis; AI is seen as an incremental, not revolutionary, step there.

Verification as the real bottleneck

  • Practicing designers stress that design time is not the main limiter; verification and validation dominate schedules.
  • Even if spec‑to‑RTL time dropped to zero, throughput of finished chips might barely change.
  • More autogenerated RTL may actually worsen verification load.

IP, partnerships, and data exploitation concerns

  • Suspicion around how AMD/Synopsys share RTL with OpenAI and whether it is mined despite “no training” promises.
  • Comparisons made to Amazon/Apple using partner or platform data to build competing products.
  • Broader worry that AI providers will both supply tools and eventually compete directly with their own customers.

Other notes

  • RISC‑V is seen as the natural choice to avoid licensing friction.
  • Some report modest success using LLMs to generate EDA scripts, glue code, and simple boards, but not high‑end layouts.
  • One subthread discusses the Jalapeño chip name and its layered meanings, mostly as light relief.