Technical Name SiliconMind-V1: Multi-Agent Distillation and Debug-Reasoning Workflows for Verilog Code Generation
Project Operator National Taiwan University
Project Host 洪士灝
Summary
SiliconMind-V1 is an open-source LLM and AI agent that automates Verilog generation, testing, and debugging for chip design. It features efficient data distillation (using 1/9 compute of CodeV-R1), three dynamic inference modes (Regular, Deep Thinking, Agentic), and EvolVE optimization, which boosts PPA by 17% to 66% over expert designs. Offering 4B-8B lightweight models, SiliconMind-V1 enables secure, low-cost on-premise deployment to enhance industry competitiveness.
Scientific Breakthrough
Integrating multi-agent distillation and self-correction, SiliconMInd-V1 creates a on-premise deployable Verilog generation framework. Compared to CodeV-R1, it delivers 9x higher training efficiency, with lightweight models achieving a world-class 82% Pass@1. The built-in three-tier inference reduces costs and enhances test-time scaling capability, while the EvolVE evolutionary mode automatically optimizes chip PPA by 17% on average and up to 66%.
Industrial Applicability
SiliconMind-V1 automates RTL generation, testing and debugging via lightweight, on-premise models to 100% safeguard corporate IP. It further incorporates EvolVE mode leverages AI compute to drastically optimize chip PPA. Experimental results show that it significantly increases test pass rates while cutting API costs, making it the premier high-ROI solution for deploying trustworthy AI design automation in semiconductors.
  • Contact
  • Shih-Hao Hung