Reproduce “UVMarvel: an Automated LLM-aided UVM Machine for Subsystem-level RTL Verification” on OpenResearch

About this paper

Junhao Ye, Dingrong Pan, Hanyuan Liu, Yuchen Hu, Jie Zhou, Ke Xu, Xinwei Fang, Xi Wang and 2 more

UVMarvel introduces an automated framework utilizing large language models to streamline Universal Verification Methodology (UVM) for subsystem-level RTL designs. It reduces verification time by 20.17x, achieving an average code coverage of 95.65% on industrial benchmarks by providing LLMs with structured design intent and focused RTL context.

Key insights

• Providing LLMs with structured, verification-oriented design intent via an Intermediate Representation significantly improves their ability to build comprehensive UVM testbenches for complex subsystems.
• Domain-specific templating, such as a Bus Protocol Library, is critical for guiding LLMs in accurately synthesizing correct and robust UVM components for intricate bus protocols.
• Extracting and presenting minimal, dependency-preserving RTL context (Filtered DUT) for specific coverage gaps empowers LLMs to effectively generate targeted stimuli for deep corner cases in large, complex designs.

Summary by alphaXiv.