Testing Generative AI for Circuit Board Design

Overall reaction to LLMs on PCB design

• Many see this as a clear demo of current LLM limits: models generate schematics and footprints that look plausible but have critical errors (missing connections, wrong footprints, bad decoupling, extra components).
• Several EEs stress that a single such error can make an entire board worthless; “70% correct” designs are often harder to fix than starting from scratch.
• Others are impressed that frontier models can do anything at all in this domain and view the results as “amazing for 2024,” but still not production‑ready.

Where LLMs seem useful today

• Strong agreement that LLMs shine at:
  • Parsing and summarizing datasheets and PDFs.
  • Turning informal requirements into high‑level code or DSLs (e.g., SKiDL, tscircuit‑style React DSLs).
  • Automating boilerplate tasks: power nets, decoupling suggestions, pin tying, reminders like “you forgot I²C pullups.”
• Several comments argue that LLMs are best used as “vocabulary expanders” and design‑space explorers, not as final designers.

Alternative approaches and future directions

• Multiple posters argue pure LLM/next‑token prediction is the wrong tool for holistic, highly constrained design problems like PCB layout or high‑speed transmission lines.
• Suggestions include:
  • Diffusion or other generative models for combinatorial optimization and placement/routing.
  • Domain‑specific solvers for length matching, SI, etc., with LLMs as a front‑end.
  • Fine‑tuning on netlists, footprints, or synthetic datasets; RL or self‑play using rule‑based checkers.
• Some report promising results using diffusion and autoregressive models on combinatorial problems (MaxCut, MIS, MaxClique) but note neural search and exploration remain open challenges.

Data, components, and verification issues

• A major bottleneck: poor, inconsistent, NDA‑gated datasheets and third‑party libraries. Juniors are taught not to trust existing symbols/footprints.
• Proposals include “datasheet‑to‑component” and “datasheet‑to‑SPICE” pipelines, but others warn verification is hard and AI‑generated libraries will not be trusted without strict checks.

Meta‑discussion: model quality, hype, and “goalposts”

• Mixed experiences comparing Claude, GPT‑4o, GPT‑4‑turbo, Gemini; performance is highly task‑dependent.
• Debate over “goalpost shifting”: some say expectations unrealistically jump from “works at all” to “zero hallucinations,” others argue reliability is a legitimate new requirement.
• Broad consensus: LLMs are transformative and useful in many workflows, but far from replacing domain experts in safety‑critical or nuance‑heavy design tasks.