FastKernels: Benchmarking GPU Kernel Generation in Production

FastKernels is a production-aligned benchmark and inference framework for GPU kernel generation.

It spans 46 representative architectures across 8 categories and is aligned with real serving stacks.

Results show top kernel agents achieve only limited or even negative speedups versus production baselines.

The work exposes a gap between benchmark wins and real-world throughput, and the limits of current kernel-generation agents.