NVIDIA introduces early-stage Rust tracks for GPU kernel programming
CUDA Rust allows GPU kernels to be written in Rust and compiled natively to PTX, but NVIDIA says neither of its new tracks is production-ready.
NVIDIA has announced CUDA Rust, a programme for writing GPU kernels in Rust and compiling them natively to PTX. The company introduced two early-stage tracks: cuda-oxide for single instruction, multiple thread programming and cutile-rs for tile-based programming.
NVIDIA’s technical blog describes cuda-oxide as a custom rustc code-generation backend that routes Rust kernel code through intermediate representations to PTX, while leaving other Rust code with the standard compiler backend. The track requires Linux, a GPU with compute capability 8.0 or later, CUDA 12.x or newer and a pinned nightly Rust toolchain, among other dependencies.
cutile-rs operates at a higher level, using CUDA Tile IR to let the compiler determine how tile operations map to GPU threads. It requires Linux, CUDA 13.3 and stable Rust 1.89 or newer, but does not require a nightly toolchain or separately installed LLVM.
NVIDIA says both tracks provide compile-time safeguards against certain aliasing and launch-configuration errors. cuda-oxide uses types and launch contracts to constrain access to output buffers and validate configurations, while cutile-rs uses ownership and tile partitioning to manage writable tensors.
The company positions CUDA Rust alongside its mature, enterprise-grade CUDA C++ and CUDA Python toolchains. It says Rust is already used in parts of its systems software, including the Nova Linux driver and the core of NVIDIA Dynamo.
Neither project is production-ready. NVIDIA describes cuda-oxide as early alpha, while cutile-rs has been published and is reported to have external use. The announcement provides no comparative benchmarks against CUDA C++ or CUDA Python, and says APIs remain subject to change.
