NVIDIA opens two early ways to write GPU programs in Rust

NVIDIA published two early ways to write GPU programs in the Rust language that compile straight to the chip, instead of only calling code written in other languages.

A lot of AI infrastructure is already written in Rust. This is NVIDIA saying, on its own blog, that the tiny programs that run on the GPU can be written in Rust too. Both tracks are early. NVIDIA says they are not production-ready.

NVIDIA’s technical blog, by Sri Koundinyan, Melih Elibol, and Jonathan Bentz, announces CUDA Rust. CUDA is NVIDIA’s system for running code on its graphics chips. A kernel is a small program that runs on the GPU. NVIDIA says you write those kernels in Rust and compile them natively to PTX, the GPU’s own instruction set — not just launch someone else’s kernels from Rust.

There are two tracks. cuda-oxide follows the usual CUDA style: many threads running the same instruction (SIMT). NVIDIA calls it early alpha — first-draft software. cutile-rs uses a newer tile style and is on crates.io, the public Rust package site. NVIDIA says Hugging Face’s Grout engine and mistral.rs already use that second track.

NVIDIA says both tracks are early-stage and neither is production-ready. Coverage is incomplete and the programming interfaces will change.

Both tracks try to catch a class of memory bugs at compile time, before the program runs.

NVIDIA’s context: other NVIDIA pieces already use Rust (the Nova Linux driver, Dynamo, NVTX bindings). Kernels were the gap. Sharing work with CUDA C++ and Python is planned.

Compiler internals, pinned toolchains, Rust versions, and GPU-generation cutoffs are on the NVIDIA blog in Sources. They are not the news.