CUDA Rust

CUDA Rust is NVIDIA's official toolkit for writing GPU code in Rust, and it's free under Apache 2.0. It grew out of cuda-oxide, an NVIDIA research project, into a bigger package: shared host crates for talking to the GPU driver, plus two ways to write kernels (the small programs that run on the GPU).

cuda-oxide is the hands-on option. A custom compiler backend turns plain Rust into PTX, with threads, shared memory and warp operations exposed the way CUDA C++ programmers expect. cutile-rs is the high-level option: you work on tiles of tensors, NumPy or PyTorch style, and the compiler handles scheduling.

Nothing to host and no paid tier. Solo, small teams, larger teams: all free, and the bill is the GPU. You need Linux, an NVIDIA card and the CUDA toolkit. cutile-rs builds on stable Rust but wants CUDA 13.2 or newer and an Ampere or newer GPU. cuda-oxide needs a pinned nightly Rust and Clang 21.

The catch: NVIDIA calls it an early-stage research project: expect bugs, incomplete features and API breakage. Keep production kernels on CUDA C++. And install cargo-oxide from NVIDIA's git repo, because the cuda-oxide name on crates.io is an unrelated 2021 crate.