Research · Updated 6 Oct, 09:30 am IST
SGAnalog: benchmark of open-source analog tapeout circuits for model evaluation
Why it matters for readers: It demonstrates how researchers measure whether AI systems really learned circuit skills instead of memorizing examples.
- The benchmark collects 273 topologically distinct top-level designs from Tiny Tapeout manufacturing shuttle submissions.1
- Each source is processed at its recorded revision in a fixed containerized environment and exports schematic images plus SPICE netlists from the same source file.1
- SGAnalog evaluates both schematic-to-netlist transcription and device sizing tasks; on 66 transcription tasks the strongest model reached 56.1% exact graph isomorphism.1
- Across model sizes six of seven models drop sharply in performance from the small to the medium tier, indicating sensitivity to model scale or data.1
- On 17 sizing tasks the leading model converged on all proposals and scored 91.2 out of 100 against the human reference; removing author-chosen labels reduced exact matches but preserved structural F1 for one frontier model.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started