EEBench: Can AI Design Circuit Boards Yet?
Original: Can AI design circuit boards yet?
Why This Matters
Standardized hardware AI benchmarks are critical as models increasingly target physical engineering workflows.
EEBench, a new benchmark for AI-generated electronics, launched alongside OpenAI's GPT-6 Astra demo in KiCad. The benchmark uses atopile's declarative code format to test whether AI models can produce functional, real-world circuit designs — not just draw schematics.
When OpenAI featured a GPT-6 Astra demo operating in KiCad on the front page of its launch post, the team behind EEBench saw it as a timely moment to introduce their benchmark for evaluating AI-generated circuit designs. Rather than testing a model's ability to click through a GUI, EEBench uses atopile — a declarative code-based circuit design format — so agents can work directly with components, connections, and electrical constraints, running simulations and inspecting failures within the same environment.
One public benchmark task is modeled on a residential energy meter hold-up circuit: when the 5V supply drops, the design must keep a processor alive for 20ms above a 3.0V brownout threshold. While most models correctly identify the need for a capacitor, real-world physics complicate things. Ceramic capacitors lose significant capacitance under DC bias — one submitted design nominally used 22µF, but measured only 11.4µF effective capacitance at 4.7V bias, far short of the 545µF required. The circuit failed at 0.85ms, well before the 20ms requirement. EEBench checks voltage throughout the outage, capacitor effective capacitance, recovery behavior, package constraints, and cost. More advanced tasks include designing multiple-feedback low-pass filters around op-amps.