What happens when you run Artemis against production-grade CUDA libraries? You do not only optimise code. You uncover defects that existing test suites never exercised.
The libraries
We applied Artemis to two core NVIDIA GPU libraries: cuCollections, used for high-performance data structures, and NCCL, the multi-GPU communication layer behind most distributed training. Both are widely deployed, heavily reviewed, and well tested.
What surfaced
Artemis flagged real bugs, tightened edge-case logic, and made critical paths safer and more maintainable. These were not style nits. They were behaviours that only appear when code is executed across the range of inputs an evolutionary search will try, rather than the range a human test author thought to write down.
Optimisation as inspection
This is the part teams tend to underestimate. An optimiser that builds, runs, and benchmarks thousands of variants is also, incidentally, one of the most thorough readers your codebase will ever get. Correctness findings come free with the performance work, because both come from the same source: actually running the thing.
