NVIDIA's CUDA Rust Brings Native GPU Kernel Writing in Rust

NVIDIA has announced CUDA Rust, a native path for writing GPU kernels directly in Rust instead of wrapping CUDA C++ or Python code. The move extends Rust's growing footprint across NVIDIA's stack, seen already in the Nova Linux driver and the Dynamo serving core, and closes one of the last remaining gaps: the kernel itself.

Two separate projects implement the two CUDA programming models. cuda-oxide targets the classic SIMT model, where you describe what a single thread does and launch thousands of them. It works as a custom rustc codegen backend, routing #[kernel] functions through Rust MIR, the Pliron IR framework, and LLVM IR down to PTX, and requires a nightly toolchain plus a compute capability 8.0+ GPU. cutile-rs targets the newer Tile model, where you describe what one tile of data does and let the compiler decide how it maps onto real threads. It ships as an ordinary crate, needs only stable Rust and CUDA 13.3, and JIT-compiles kernels through CUDA Tile IR the first time they run.

For engineers, this means GPU kernels can now live in the same crate, with the same safety guarantees and build pipeline as host code, eliminating separate CUDA C++ kernel files or FFI wrapper layers. NVIDIA recommends reaching for Tile first for architecture-portable code, dropping to SIMT only when fine-grained control over memory and threads is needed. Both projects are early-stage but functional, with interoperability planned between the two tracks and the wider CUDA ecosystem.