AMD's OpenSIL Firmware Now Has an AI That Writes Its Own Tests — And Fixes Them Too
New AI agent system generates 98.8% line coverage for firmware unit tests
Researchers from AMD and academic collaborators have developed an AI-driven automated testing system for firmware development that significantly reduces manual debugging effort. The approach, detailed in a new arXiv paper, focuses on AMD's Open-Source Silicon Initialization Library (OpenSIL), a critical component in firmware development.
The system employs a large language model (LLM) guided multi-agent pipeline that combines three key components: automated test scaffold generation, library-aware creation of stubs/mocks/fakes, and an iterative compile-dispatch repair loop driven by build logs and line-coverage feedback. This addresses a major challenge in low-level C firmware testing where strict build constraints often prevent compilation due to missing headers or dependency mismatches. Across 76 tested functions, the system generated compilable unit tests for 73 functions, with mean line coverage reaching 98.8% when using line-coverage guidance.
- AMD's OpenSIL firmware now has AI-powered unit test generation achieving 98.8% line coverage on 48 tested functions
- The multi-agent LLM pipeline uses library-aware mocks and iterative repair loops to handle strict build constraints
- System generated compilable unit tests for 73 out of 76 tested functions with reduced manual debugging effort
Why It Matters
This breakthrough could reduce firmware development cycles by automating tedious test maintenance in constrained environments