In September 2026, Nvidia announced native GPU programming support in Rust, expanding beyond its mature CUDA C++ and CUDA Python toolchains. The move reflects a broader shift in AI infrastructure toward Rust, which catches entire classes of bugs at compile time without sacrificing performance.
Rust has become increasingly prevalent in Nvidia’s own systems. The Nova Linux driver is written in Rust, Nvidia Dynamo is built on a Rust core, and NVTX has Rust bindings. However, GPU kernels—the programs that run directly on graphics processors—traditionally had to be written in other languages, with Rust able only to launch them.
Nvidia CUDA Rust closes that gap, enabling kernels to be written in Rust and compiled natively to PTX (Parallel Thread Execution), rather than wrapping code written elsewhere.
The announcement includes two programming tracks, mirroring CUDA’s existing approaches. The SIMT track uses the Single-Instruction Multiple-Thread model already familiar to CUDA C++ and numba-cuda programmers, where developers specify what one thread does and launch thousands of copies. The Tile track represents a newer programming model, also available in C++ and Python, where developers specify what one tile of data does and let the compiler handle mapping to the hardware.
Nvidia recommends starting with Tile, since the compiler automatically decides how tiles map onto different GPU architectures, avoiding architecture-specific code. SIMT is recommended when developers need fine-grained control over memory and thread management.
For the SIMT track, Nvidia offers cuda-oxide, a custom Rust compiler backend that intercepts compilation and routes kernel functions through Rust’s MIR (Mid-level IR), the community Pliron IR framework, and LLVM IR down to PTX. Requirements include Linux, a GPU with compute capability 8.0 or later, CUDA toolkit 12.x or newer, and a pinned nightly Rust toolchain.
The Tile track uses cutile-rs, which operates at a higher abstraction level. It performs computations on tiles rather than individual values, with the compiler deciding how many actual GPU threads back each tile. This approach has lighter requirements: a GPU with compute capability 8.0 or later, CUDA 13.3, stable Rust 1.89 or newer, and Linux—no nightly toolchain or custom LLVM needed. Cutile is published on standard Rust package repositories, making setup simpler.
Nvidia plans to support inter-language interoperability, meaning developers won’t be locked into Rust if they choose it for their GPU code. The company plans to continue growing and maturing CUDA Rust into 2027 and beyond.
Key facts
- Nvidia announced native GPU kernel programming in Rust in September 2026
- Two programming tracks available: SIMT (thread-level) and Tile (data-tile level)
- Cuda-oxide requires Linux, GPU compute capability 8.0+, CUDA 12.x+, and nightly Rust
- Cutile-rs has lighter requirements: no nightly toolchain or custom LLVM needed
- Nvidia’s own infrastructure increasingly uses Rust, including Nova Linux driver and Nvidia Dynamo