Normal Computing Builds Open-Source Verilog Simulator Using AI Agents: 580K Lines in 43 Days

Summary

Normal Computing has demonstrated the potential of agentic AI in engineering by building a comprehensive open-source Verilog simulator in just 43 days. The project added 580,430 lines of code across 3,846 files to CIRCT (Circuit IR Compilers and Tools), an LLVM-based infrastructure for hardware design. The effort included implementing an event-driven simulator, VPI/cocotb integration, UVM runtime support, bounded model checking, logic equivalence checking, and mutation testing—creating a practical verification stack capable of simulating real-world protocol testbenches end-to-end.

The development utilized AI agents across 2,968 commits, with Claude models (Opus 4.5 and 4.6) handling 54% of the work and OpenAI's Codex models contributing 46%. The pace accelerated from approximately 25 commits per day in the first week to a peak of 124 commits per day by week seven. The project expanded CIRCT's test files from 987 to 4,229—a 4.3x increase—and achieved 100% compliance with the sv-tests suite, surpassing established free simulators like Verilator (94%) and Icarus (80%).

Formal verification and mutation testing accounted for over 1,000 commits (34% of total work), while the Verilog frontend received 461 commits for critical lowering passes. The project maintained detailed engineering logs tracking 1,554 iteration cycles, demonstrating a systematic approach to AI-assisted development. Normal Computing's experiment highlights how agentic AI can tackle well-specified, labor-intensive engineering problems where specifications are public (IEEE 1800-2017) and infrastructure exists, but significant implementation volume is required. The result challenges the commercial EDA toolchain market, which typically costs teams millions of dollars annually.

• The simulator includes event-driven simulation, formal verification, mutation testing, and UVM runtime support—creating a practical alternative to million-dollar commercial EDA toolchains
• Test coverage expanded 4.3x from 987 to 4,229 test files, with formal verification and mutation testing accounting for 34% of total development effort

Editorial Opinion

Normal Computing's achievement represents a watershed moment for agentic AI in complex engineering tasks. By targeting a domain with rigorous public specifications (IEEE 1800-2017) and existing compiler infrastructure, they've demonstrated that AI agents can now tackle million-line codebases with sufficient quality to exceed established open-source tools. The 43-day timeline and 100% sv-tests compliance rate suggest we're entering an era where AI can compress years of traditional development into weeks for well-specified problems. This has profound implications not just for semiconductor tooling, but for any mature engineering domain where specifications exist but implementation requires massive human effort.