Skip to content
Eval Suites50

KernelBench

Asks models to write GPU kernels for PyTorch workloads and checks both correctness and speedup over the reference.

KernelBench at a glance

Maintainer
Stanford Scaling Intelligence
Size
250 workloads[1]
Category
Coding & Systems
Runs with
Python / CUDA
License
MIT
Tier
frontier

Failure modes KernelBench catches

Incorrect kernels
No speedup
Numerical drift

Starter command

A starting point, not a guarantee. Commands change between releases, so confirm against the official README.

KernelBench starter
git clone https://github.com/ScalingIntelligence/KernelBench
# follow the repo README for setup and run commands

Sources

  1. 1. KernelBench official page or repository · Stanford Scaling Intelligence · Repository · Sources checked
  2. 2. KernelBench paper · arXiv · Paper · Sources checked