Hugging Face Releases SGAnalog: Revolutionizing Analog Circuit Design Benchmarking

Summary:Hugging Face has released SGAnalog, an open-source benchmark for analog integrated circuit design, addressing the challenges of assessing whether models have learned transferable circuit skills and if their outputs are viable under defined process and test conditions. The benchmark comprises 273 topologically distinct designs, evaluating schematic-to-netlist transcription and device sizing. Experiments demonstrate SGAnalog's effectiveness in benchmarking model performance, highlighting distinct

---

Background and Challenges

Analog circuit design has long faced two key challenges: whether models have truly learned transferable circuit design skills and if their outputs are viable under defined process and test conditions. Existing benchmarks struggle to comprehensively address these issues.

Innovations of SGAnalog

Hugging Face's SGAnalog benchmark addresses these challenges through the following features:

• Open-Source Circuit Designs: Based on human-designed, open-source circuits associated with Tiny Tapeout manufacturing shuttles, the benchmark includes 273 topologically distinct top-level designs.
• Precise Processing Pipeline: Each circuit source is retrieved at the revision recorded for its shuttle submission and processed in a fixed containerized environment, ensuring an exact structural reference for transcription.
• Multi-Dimensional Evaluation: Covers schematic-to-netlist transcription tasks and device sizing optimization tasks.
• Model Performance Analysis: Evaluates the performance of seven models on a fixed set of tasks, revealing differences in visual and design capabilities.

Experimental Results and Findings

• Transcription Tasks: The strongest model achieved 56.1% exact graph isomorphism on 66 transcription tasks, but most models showed a sharp decline from small to medium task sets.
• Role of Labels: For a frontier model, removing author-chosen labels reduced exact matches but preserved aggregate structural F1, suggesting labels aid connectivity tracing.
• Sizing Tasks: The leading model converged on all 17 proposals and outperformed the human reference with a score of 91.2/100, while Claude models struggled with proposals lacking sizes.

Industry Impact and Developer Recommendations

SGAnalog provides a more reliable evaluation framework for analog circuit design, helping developers better understand model performance across different tasks. Its open-source nature also fosters community collaboration and innovation.

Future Outlook

As SGAnalog becomes widely adopted, the automation and intelligence of analog circuit design will be further enhanced, driving rapid advancements in the field.

---

Source: ArXiv AI (cs.AI) (2026-10-07)

— END —

Tags: #Hugging Face Follow #Open-Source Models Follow #Analog Circuit Design Follow #AI Benchmarking Follow