Lightning Talk: Harnessing NVIDIA Tensor Cores: An Exploration of CUTLASS & OpenAI..
summary
CUTLASS v3 targets NVIDIA Tensor Cores with a layered C++ library and a new Python front end that cuts boilerplate to about one third the parameters, adds an epilog visitor tree for composable store/load epilogues on Ampere and Hopper, and shows profiles hitting ~90% peak utilization on H100 with CUDA 12.2. Upcoming 3.3 supports mixed-input GEMMs (A fp16, B int8 upcast to bf16), grouped GEMMs and pointer-array GEMMs later in the year, and Hopper Triton support is already released.
AI-written summary · about
related talks
35mCUTLASS: A CUDA C++ Template Library for Accelerating Deep Learning...
08mLightning Talk: Achieving SOTA GEMM Performance: A CuTeDSL Backend for PyTorch Induc...
05mLightning Talk: GPU Programming with C++ and Triton
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
14mLightning Talk: Every Millisecond Counts: The Fine-tuning Journey of an Ultra-Eff...
24mKeynote: Next-generation frameworks for Large-scale AI
19mTask-Based Tensor Computations on Modern GPUs
12mKeynote: Build Accelerated LLM Applications with Open Source Tools from NVIDIA