NVIDIA announced CUDA Rust on September 8, 2026, as a way to write GPU kernels natively in Rust rather than wrapping code from another language. The initiative consists of two tracks that mirror CUDA's own programming models: SIMT (Single Instruction, Multiple Threads) and Tile.
The logo appears in the original NVIDIA developer blog post announcing CUDA Rust, confirming the source of the technology.
The SIMT track is realized by cuda-oxide, a custom rustc codegen backend that compiles #[kernel] functions through Rust MIR, the Pliron IR framework, and LLVM IR down to PTX. This track requires Linux, a GPU with compute capability 8.0 or later, a CUDA toolkit (12.x or newer), clang with libclang headers, and a pinned nightly Rust toolchain. It is currently in early alpha.
The Tile track is provided by cutile-rs, which operates on stable Rust 1.89 or newer with CUDA 13.3 and does not require a custom LLVM installation. cutile-rs is published on crates.io and is already used in production inference engines such as HuggingFace's Grout and mistral.rs.
This figure from the NVIDIA blog illustrates the Tile programming model that cutile-rs implements, helping readers visualize how threads collaborate on data tiles.
Both tracks enforce memory safety at compile time. cuda-oxide uses a DisjointSlice type to give each thread exclusive access to its own data and employs #[launch_contract] attributes to validate launch configurations. cutile-rs relies on tensor partitioning and ownership guarantees to ensure exclusive access without needing a special type.
NVIDIA states that it plans to support inter-language interoperability between CUDA Rust, CUDA C++, and CUDA Python so that the choice of frontend does not lock developers into a single ecosystem.