Benchmark RadarFastKernels
FastKernels evaluates GPU kernel generation for production inference across 46 representative architectures spanning 8 categories. It measures correctness and speedup of candidate kernels against reference implementations at multiple abstraction levels, from single-kernel to end-to-end serving.
- Released
- 2026-05-22
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing kernel benchmarks reward sandbox optimizations that fail in production. FastKernels aligns evaluation with real inference frameworks, providing a benchmark where scores translate to throughput improvements in production codebases.
Motivation
LLM-based agents for GPU kernel generation are advancing rapidly, yet their progress is fundamentally constrained by the benchmarks they optimize against.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.