# ScalingIntelligence/KernelBench

KernelBench: Can LLMs Write GPU Kernels? - Benchmark + Toolkit with Torch -> CUDA (+ more DSLs)

Repository: https://github.com/ScalingIntelligence/KernelBench
Canonical: https://ross.abutalabs.com/products/kernelbench
Homepage: https://scalingintelligence.stanford.edu/blogs/kernelbench/
Language: Jupyter Notebook
License: NOASSERTION
License Family: other
Topics: benchmark, codegen, evaluation, gpu, tooling, rl-environment
Last push: 2026-03-24T04:29:13+00:00

## Health v2 (maintenance only)
Score: 55/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 73, release rhythm 35, longevity 48
- inputs: {"age_days": 678, "days_push": 162, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1214, forks 192 (observed 2026-08-28T04:04:00.798616+00:00)

## What it is
KernelBench is a benchmark and toolkit from Stanford's Scaling Intelligence Lab that evaluates whether LLMs can generate correct and efficient GPU kernels from PyTorch programs. It provides 250+ tasks across four difficulty levels, evaluation scripts for correctness and speedup checks, and a HuggingFace dataset.

## Use cases
- benchmark llm code generation for cuda kernels
- evaluate if a model can write efficient gpu kernels
- measure speedup of llm-generated kernels against pytorch baselines
- use kernel optimization tasks as an rl environment for training models
- compare codegen models on gpu kernel transpilation
- optimize pytorch operators with custom cuda kernels

## When to choose
- you need a standardized benchmark for llm gpu kernel generation
- you want to evaluate or train models on cuda code generation tasks
- you are researching agentic kernel optimization with profiling feedback

## When to avoid
- you need a production tool to auto-generate kernels for your app
- you have no gpu hardware available
- you want a turnkey agentic scaffold rather than a benchmark toolkit

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, llm-training, machine-learning, gpu-computing
- domain: large-language-models, machine-learning, gpu-computing, developer-tools, artificial-intelligence
- platform: python, cli
- tags: llm-benchmark, cuda-codegen, kernel-generation, code-evaluation, rl-environment, pytorch, evaluation, gpu, linux

## Member repositories
- ScalingIntelligence/KernelBench (main) score 55

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:00.798616+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:17:20.510300+00:00, confidence not recorded.
  - readme: https://github.com/ScalingIntelligence/KernelBench (fetched 2026-08-28T04:04:00.798616+00:00, sha b64c5ac676fb)
  - homepage: https://scalingintelligence.stanford.edu/blogs/kernelbench/ (fetched 2026-08-29T12:25:31.981956+00:00, sha 694d214d031c)
- Data as of 2026-08-30T08:39:29.467469+00:00.