SYSTEM Cited by 1 source
KernelBench¶
Definition¶
KernelBench is a Stanford-authored benchmark suite of 250 kernel-optimization problems spanning three difficulty levels, where each problem asks for a GPU kernel that is (a) functionally correct and (b) faster than a PyTorch reference implementation. It is the external benchmark Meta's KernelEvolve reports against to validate generalization beyond Meta-internal production workloads (Source: sources/2026-04-02-meta-kernelevolve-how-metas-ranking-engineer-agent-optimizes-ai-infrastructure).
KernelEvolve result¶
Meta reports 100% pass rate on KernelBench — "all generated kernels are both functionally correct and faster than their PyTorch reference implementations." This is the headline external-benchmark disclosure in the 2026-04-02 post.
Alongside the KernelBench 250-problem score, Meta reports 480 configurations validated (160 PyTorch ATen operators × 3 hardware platforms, 100% correctness) as the internal correctness benchmark covering PyTorch's standard-operator coverage.
Seen in¶
- Meta KernelEvolve (2026-04-02, canonical). The external Stanford benchmark Meta's agentic kernel system reports perfect score on. (Source: sources/2026-04-02-meta-kernelevolve-how-metas-ranking-engineer-agent-optimizes-ai-infrastructure)
Caveats¶
The 2026-04-02 post does not characterize KernelBench's three difficulty levels in detail, nor the specific problem set. Stanford's publicly-released benchmark is the canonical source.
Related¶
- companies/meta — KernelEvolve author company.
- systems/kernelevolve — the agentic system that achieves 100% pass rate on KernelBench.
- systems/tritonbench — Meta's complementary internal benchmark harness.