NVIDIA introduces two tracks for CUDA Rust GPU kernels
NVIDIA’s CUDA Rust work targets both major CUDA programming models: cuda-oxide compiles SIMT-style Rust kernels to PTX through a custom rustc/Pliron/LLVM backend, while cutile-rs offers tile-based GPU programming from stable Rust via CUDA Tile IR.
The first path is early-alpha and needs a pinned nightly toolchain; the second targets stable Rust and is already used in Hugging Face’s Grout and mistral.rs, according to NVIDIA. Both projects describe compile-time approaches to aliasing and memory safety, with planned interoperability across CUDA Rust, C++, and Python. Lobsters had no substantive comments.