Introducing CUDA Rust: Two Tracks for Writing GPU Kernels

Rust already shows up in NVIDIA’s own stack—Nova’s Linux driver, Dynamo’s core, NVTX bindings—but GPU kernels have mostly stayed in other languages even when launched from Rust. CUDA Rust aims to close that gap with kernels written in Rust and compiled natively to PTX, not wrapped foreign code.

The company mirrors CUDA’s own dual model: SIMT (thread-centric, familiar from CUDA C++ and numba-cuda) and Tile (data-tile-centric, with the compiler owning mapping). Guidance is explicit: prefer Tile first; drop to SIMT when you need manual control of threads and memory. Inter-language interop with CUDA C++ and Python is planned so a Rust choice does not strand teams.

This is less a finished product launch than a strategic stake in the ground. Both projects are early-stage and, by NVIDIA’s own admission, not production-ready. cuda-oxide is early alpha: a custom rustc codegen backend through Pliron and LLVM, pinned nightly, heavier toolchain. cutile-rs is further along—stable Rust 1.89+, CUDA 13.3, on crates.io, and already used outside NVIDIA in HuggingFace’s Grout inference engine and mistral.rs.

Signal Desk's read: NVIDIA is treating Rust kernel authorship as inevitable for the inference/serving/agent-runtime layer and is buying influence before community toolchains harden around it. The safety pitch is the differentiator worth watching: both tracks try to make classic buffer-aliasing mistakes fail at compile time (DisjointSlice and launch contracts on SIMT; partition ownership on Tile). That is a real systems argument, not just syntax sugar.

What not to overclaim: maturity. Shared memory on the SIMT track still needs unsafe today; APIs will move; coverage is incomplete. The understated risk for NVIDIA is fragmentation—another frontend that teams try and abandon if nightly pins and incomplete coverage persist. Success looks like Tile becoming the default Rust path and SIMT graduating off nightly without losing the safety story.

NVIDIA says it will grow and mature CUDA Rust into 2027 and beyond alongside mature CUDA C++ and Python toolchains. Community precursors (rust-cuda, Rust-GPU, CubeCL, and others) are acknowledged; NVIDIA says it has been working with rust-cuda maintainers. Melih Elibol is presenting “Fearless Concurrency on the GPU” at RustConf 2026 in Montréal.