OpenAI Triton: writing GPU kernels in a Python-like language
In one sentence OpenAI releases Triton 1.0, an open-source Python-like language and compiler for writing efficient GPU kernels without CUDA experience: an FP16 matrix multiplication on par with cuBLAS in under 25 lines.
Writing fast code for NVIDIA GPUs is historically hard: you need to know CUDA and manually handle memory transfers, shared memory and scheduling.
OpenAI releases Triton, a language (and compiler) that looks like Python and lets researchers with no CUDA experience write GPU kernels that, according to OpenAI, are most of the time on par with what an expert would produce.
The project grew out of a 2019 paper by Philippe Tillet, who later joined OpenAI. At launch it only runs on recent NVIDIA GPUs: CPUs and AMD GPUs are not supported.
Companies
OpenAI
Tools
Triton
Tags
Sources