NVIDIA brings Rust to GPU kernels with cuda-oxide and cutile-rs
NVIDIA announced CUDA Rust, two open-source projects that compile Rust GPU kernels natively and catch aliasing bugs at compile time.
NVIDIA has announced CUDA Rust, an effort to make Rust a first-class language for writing GPU kernels. Rust code could already launch CUDA kernels, but the kernel body usually had to live in another language. Two NVlabs open-source projects close that gap: cuda-oxide for the SIMT model and cutile-rs for the newer Tile model. Both compile Rust kernels natively and lean on Rust ownership rules to reject aliasing bugs at compile time.
Two programming models
SIMT is the model behind CUDA C++ and numba-cuda: you describe what one thread does and launch thousands of them. Tile describes what one tile of data does, leaving thread mapping and memory layout to the Tile IR compiler. NVIDIA recommends Tile first and SIMT when explicit thread and memory control is needed. Planned inter-language interoperability means picking Rust will not lock developers out of C++ or Python.
cuda-oxide is a custom rustc codegen backend. It routes #[kernel] functions through Rust MIR, the community Pliron IR framework and LLVM IR down to PTX, handing everything else to the standard backend. It requires Linux, a GPU with compute capability 8.0 or later, CUDA 12.x or newer, clang with libclang and a pinned nightly-2026-04-03 toolchain.
cutile-rs sits one level higher. Each tile block runs the kernel body once as a single logical thread over a sub-tensor, and the compiler decides how many real GPU threads back it. Requirements are lighter: compute capability 8.0 or later, CUDA 13.3, stable Rust 1.89 or newer and Linux, with no nightly and no custom LLVM.
The safety claim
On the SIMT side the output buffer is a DisjointSlice<f32>, giving each thread exclusive access to its own element, and out-of-bounds access returns an Option. Passing that output buffer back in as an input fails with E0502 under SIMT and E0382 under Tile. NVIDIA calls cutile-rs the stronger guarantee because ownership follows tensors across the launch boundary. Tile exposes no shared memory or thread indexing; shared memory in cuda-oxide still requires unsafe.
cutile-rs is published on crates.io and already used in Hugging Face's Grout inference engine and in mistral.rs. cuda-oxide is early alpha. NVIDIA says both projects are in alpha and not confirmed for production. Technical details are in the MarkTechPost report.
