cutlass

Why we included this project

CUTLASS is the go-to reference for writing fast GEMM kernels on NVIDIA GPUs. It breaks matrix multiplication into modular, reusable pieces, tiling, data movement, and tensor-core operations, so you can assemble and specialize your own kernels instead of starting from a blank file. The library has shipped CUDA C++ template abstractions since 2017, and it covers a wide set of precisions, from FP32, TF32, and FP16 to FP8, block-scaled formats, narrow integers, and even binary data, across architectures from Volta through Blackwell. The Python DSLs introduced in version 4, starting with CuTe DSL, expose the same underlying concepts with much shorter compile times and easier integration into deep learning frameworks, which makes them far more approachable for teams that don't want to live in C++ template metaprogramming. For performance engineers and framework developers building custom kernels, most of the hard low-level work is already done here.