A hands-on tutorial series for writing high-performance CUDA kernels with ThunderKittens.
The first few levels can be completed on any Ampere or newer GPU architectures. Starting at level X (to be announced), however, you will need the H100s to run the tensor core matrix multiplications.
Clone with submodules:
git clone --recurse-submodules <repo-url>If you already cloned without submodules:
git submodule update --init --recursive- Harness — Raw CUDA GEMM and correctness check / benchmarking harness