Benchmark
KernelBench
Test whether language models write correct GPU kernels that outperform reference implementations.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The original suite contains 250 PyTorch workloads spanning operators, fused operations, and model architectures.
Methodology
Check generated kernels against randomized inputs, then time them against PyTorch. Report the fraction both correct and above a specified speedup threshold.