NVIDIA’s CUDA Rust platform includes tools for writing GPU kernels in Rust. Its current repository brings together the SIMT and Tile approaches and shared host-side runtime crates.
CUDA Rust offers two kernel models
The cuda-oxide track gives programmers explicit control over threads and blocks. The cutile-rs track expresses work as tensor tiles and lets the compiler map those tiles onto the GPU architecture. NVIDIA’s repository explains these as choices for different kernel requirements, rather than a replacement for every existing CUDA workflow.
https://github.com/NVIDIA/cuda-rust
Compilation differs between the tracks
The current repository describes ahead-of-time Rust-to-PTX compilation for cuda-oxide. For cutile-rs it describes just-in-time compilation at first launch, through Tile IR to a device binary. Calling the whole platform a single stable-Rust-to-PTX compiler would erase this distinction.
https://github.com/NVIDIA/cuda-rust
Development plans are not a readiness guarantee
NVIDIA’s September blog said it would grow and mature CUDA Rust into the following year and beyond. The current repository still calls the platform an early-stage research project and warns of bugs, incomplete features and API breakage. That is not a guaranteed production-readiness date.
https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/
Sources and further reading
https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/
https://github.com/NVIDIA/cuda-rust