NVIDIA Brings GPU Kernels to Rust With CUDA Rust's Two Tracks
NVIDIA introduced CUDA Rust: cuda-oxide compiles SIMT kernels straight to PTX on a pinned nightly, while cutile-rs brings tile-based GPU programming to stable Rust 1.89+ and is already used by HuggingFace's Grout engine.
What NVIDIA announced
NVIDIA introduced CUDA Rust, opening two official tracks for writing GPU kernels in Rust. cuda-oxide is a custom rustc codegen backend that compiles SIMT-style kernels straight to PTX through the Pliron IR framework and LLVM. cutile-rs brings tile-based GPU programming to stable Rust, letting the compiler manage thread mapping and memory layout via CUDA Tile IR.
Why it matters
Both tracks enforce memory safety at compile time — cuda-oxide with DisjointSlice and launch contracts, cutile-rs through tensor partitioning and ownership. cutile-rs already runs on stable Rust 1.89+ with CUDA 13.3 and is used in HuggingFace's Grout inference engine and mistral.rs, while cuda-oxide remains in early alpha on a pinned nightly toolchain. NVIDIA plans interoperability between CUDA Rust, CUDA C++ and CUDA Python.