ThunderKittens
A lightweight deep learning kernel authoring framework targeting GPUs. Its core objective is to enable researchers and engineers to write high-performance kernels approaching hardware limits through concise abstractions, organizing computation and data movement around the tile as the central abstraction.
Timeline (last 90 days)
ThunderKittens has been ported to the NVIDIA Vera Rubin NVL72 platform, with NVFP4 GEMM kernels reconstructed around the new hardware.
ThunderKittens' NVFP4 GEMM performance has increased from previously only 42% of the roofline theoretical peak to over 22 PFLOPS.
The reconstructed ThunderKittens kernels can now compete head-to-head with officially highly optimized solutions such as cuBLAS and the CuTe DSL.
ThunderKittens is a lightweight deep learning kernel writing framework for GPUs that organizes computation and data movement around the core abstraction of tiles.
ThunderKittens prioritizes readability and research iteration speed, sacrificing optimal performance in some extreme cases in exchange for allowing developers to write runnable custom kernels within a few hours.
All Facts (5)
ThunderKittens' NVFP4 GEMM performance has increased from previously only 42% of the roofline theoretical peak to over 22 PFLOPS.
50%UnverifiedThe reconstructed ThunderKittens kernels can now compete head-to-head with officially highly optimized solutions such as cuBLAS and the CuTe DSL.
50%UnverifiedThunderKittens is a lightweight deep learning kernel writing framework for GPUs that organizes computation and data movement around the core abstraction of tiles.
50%UnverifiedThunderKittens prioritizes readability and research iteration speed, sacrificing optimal performance in some extreme cases in exchange for allowing developers to write runnable custom kernels within a few hours.
50%UnverifiedThunderKittens has been ported to the NVIDIA Vera Rubin NVL72 platform, with NVFP4 GEMM kernels reconstructed around the new hardware.
50%