HazyResearch/ThunderKittens

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

3659 stars · 318 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

ThunderKittens is a C++/CUDA framework of tile-based primitives for writing fast deep learning GPU kernels. It embeds natively into CUDA so developers can build high-performance kernels like FlashAttention-3 with simple, extensible code.

Use cases

• write fast CUDA kernels for deep learning
• implement attention kernels like FlashAttention
• optimize matrix multiply operations on NVIDIA GPUs
• build custom GPU kernels for LLM training and inference
• prototype high-performance kernels without writing raw CUDA
• use low-precision formats like MXFP8 and NVFP4 on Blackwell GPUs

When to choose

• you need to write custom high-performance CUDA kernels for NVIDIA GPUs
• you want kernel performance comparable to hand-tuned implementations with simpler code
• you target modern NVIDIA architectures like Hopper or Blackwell
• you need tile-level primitives for matrix operations in deep learning workloads

When to avoid

• you target AMD GPUs (use HipKittens instead)
• you rely on Ampere GPU support, which is no longer actively developed
• you want a Python package with pip installation, since kernels must now be compiled individually
• you need a general-purpose linear algebra library rather than kernel primitives

Facets

framework · maturity active

gpu-computing deep-learning machine-learning llm-inference llm-training deep-learning gpu-computing machine-learning large-language-models developer-tools cpp cross-platform cuda kernels tile-primitives flash-attention nvidia blackwell hopper performance dsl gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="HazyResearch/ThunderKittens")