OpenAI introduces Triton, an open-source GPU programming language

OpenAI released Triton 1.0, an open-source, Python-like language and compiler for writing GPU code, under an MIT licence. It was built by Philippe Tillet, who had originated the design during his PhD at Harvard and then re-implemented it at OpenAI, and was aimed, in his words, at “machine learning researchers and engineers who are unfamiliar with GPU programming despite having good software engineering skills.”

Writing efficient GPU kernels by hand normally requires CUDA, Nvidia’s low-level programming toolkit, which demands expertise most machine-learning researchers do not have; most neural-network code instead relies on kernels pre-written by library authors, leaving custom operations slow unless a specialist hand-tunes them. OpenAI’s post gave, as an example, an FP16 matrix-multiplication kernel written in Triton in under 25 lines that matched the performance of cuBLAS, Nvidia’s own hand-optimised library — a task that would ordinarily take a CUDA expert far more code and tuning effort to match.

Triton did not replace CUDA or challenge Nvidia’s hardware position; it targeted the layer above it, letting researchers write custom, performant kernels without becoming GPU specialists. The project was open-sourced rather than kept as an internal OpenAI tool, and it was later adopted well beyond OpenAI — most notably as the kernel-writing backend behind PyTorch’s torch.compile, making it one of the more consequential pieces of infrastructure to come out of a frontier lab’s engineering rather than research side.