ThunderKittens GitHub repository (HazyResearch)

README, MIT. 'ThunderKittens is mainly built and tested for Hopper and Blackwell GPUs'; ARCH can be SM80, SM90, SM100, SM103 or SM120. Names the primitives it exposes: 'asynchronous WGMMA calls on H100 GPUs and TCGEN05 calls on B200 GPUs' and 'asynchronous copies and address generation with TMA'. Two design constraints worth recording: the warpgroup (4 contiguous warps) is the level from which async MMA is issued, so many ops only exist at warpgroup scope; and grid size 'should be equal to (or slightly less than) a multiple of the number of SMs on the GPU to avoid tail effects' - a scheduling rule, not just an instruction wrapper. Current release adds Blackwell support with MXFP8 and NVFP4 precision.