#174174th of 242 in ML frameworks and libraries

thunderkittens

Header-only CUDA library of tile primitives for writing fast GPU kernels

Stars
3.8k
License
MIT
Last commit
Sep 2026
Language
Cuda
Stars, last 29 weeks+495
25 Apr 2026: 3,309 stars2 May 2026: 3,319 stars9 May 2026: 3,330 stars16 May 2026: 3,343 stars23 May 2026: 3,360 stars30 May 2026: 3,384 stars6 Jun 2026: 3,400 stars13 Jun 2026: 3,413 stars20 Jun 2026: 3,450 stars27 Jun 2026: 3,475 stars4 Jul 2026: 3,497 stars11 Jul 2026: 3,518 stars18 Jul 2026: 3,539 stars25 Jul 2026: 3,555 stars1 Aug 2026: 3,573 stars8 Aug 2026: 3,609 stars15 Aug 2026: 3,623 stars22 Aug 2026: 3,634 stars29 Aug 2026: 3,652 stars5 Sept 2026: 3,662 stars12 Sept 2026: 3,672 stars19 Sept 2026: 3,716 stars26 Sept 2026: 3,727 stars3 Oct 2026: 3,736 stars10 Oct 2026: 3,744 stars17 Oct 2026: 3,756 stars24 Oct 2026: 3,783 stars31 Oct 2026: 3,790 stars7 Nov 2026: 3,799 stars14 Nov 2026: 3,804 stars

3,309 on 25 Apr 20263,804 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

ThunderKittens is a header-only C++/CUDA library that provides register and shared-memory tile and vector types, plus operations on them, for writing deep learning kernels. It wraps tensor core calls (WGMMA on H100, TCGEN05 on B200), TMA loads and stores, and NVLink transfers. The repo includes pre-implemented kernels (GEMM, attention) with per-kernel Makefiles and PyTorch bindings, and demos for LLM training and inference.

Who it is for: CUDA engineers writing custom kernels for Hopper and Blackwell GPUs

Strengths

  • Header-only: clone and include kittens.cuh, no install step
  • Covers Hopper and Blackwell features: WGMMA, TCGEN05, TMA, MXFP8, NVFP4
  • Compile-time layout checks catch invalid tile operations early
  • Example H100 GEMM in under 100 lines reaches about 855 TFLOPs (86% of peak)

Weaknesses

  • NVIDIA only; AMD GPUs need the separate HipKittens project
  • Ampere no longer actively supported
  • Requires CUDA 12.8+ and C++20; older gcc causes compile errors
  • No top-level Python package since 2.0; each kernel is built separately with make

What it needs

  • GPU required
  • Needs CUDA 12.8+, C++20 compiler (gcc 11+), PyTorch 2.8+ and PyBind11 for bundled kernels

Also in ML frameworks and libraries

See all 242
Also in ML frameworks and libraries
RankProjectScore
1transformersModel-definition library for pretrained text, vision, audio and multimodal models169.3k starsApache-2.094 out of 100
2ultralyticsPython library and CLI for training and running YOLO vision models63.5k starsAGPL-3.094 out of 100
3pytorch-image-modelsPyTorch library of image models, pretrained weights and training scripts37.3k starsApache-2.094 out of 100
#4pytorchPython tensor library with GPU support and tape-based autograd105.8k starscustom license93 out of 100
#5supervisionPython toolkit for loading, annotating and counting computer vision detections Live demo ↗ (opens in a new tab)51.7k starsMIT92 out of 100