Nvidia announces two ways to write CUDA kernels in Rust, and Hugging Face is already using one in production

Nvidia announced on 8 September that developers can now write CUDA kernels natively in Rust, through two open projects called cuda-oxide and cutile-rs. The post surfaced on Hacker News on 16 September with 133 points. Nvidia says both projects are early-stage and neither is production-ready, and names Hugging Face's Grout inference engine and mistral.rs as teams already building on cutile-rs outside Nvidia.

What the two projects do

Cuda-oxide is a custom rustc codegen backend. Nvidia's post explains that it intercepts compilation and routes functions marked #[kernel] through Rust MIR and LLVM IR down to PTX, so the kernel stays in Rust until the standard LLVM backend takes over. Nvidia calls this the SIMT track: the developer keeps manual control over shared memory and thread indexing, at the cost of unsafe blocks. It requires Linux, a GPU with compute capability 8.0 or higher, CUDA 12.x or newer, clang with libclang headers, and a pinned nightly toolchain from 3 April 2026.

Cutile-rs sits at a higher level. Nvidia describes it as tile-based programming, where a #[cutile::module] macro embeds the kernel in the host binary and JIT-compiles through the CUDA Tile IR the first time it runs. The compiler decides how tiles map to each GPU architecture, so Nvidia writes that this "eliminates architecture-specific choices". Cutile-rs runs on stable Rust 1.89 or newer, needs CUDA 13.3, and is published on crates.io.

Which one to reach for

Nvidia's own guidance in the post is direct: "reach for Tile first, and drop to SIMT when you need that control." The team also says CUDA Rust, CUDA C++ and CUDA Python will interoperate, so a project can mix the three without locking itself out of any one of them.

The status is stated plainly on the page. Cuda-oxide is early alpha, coverage is incomplete, APIs will change. Cutile-rs is further along, "already used outside Nvidia", but Nvidia still labels it not production-ready. Adopters are asked to file bug reports and feature requests rather than trust the current release surface.

Two named users tell you the projects are more than a proof of concept: Hugging Face's Grout inference engine and mistral.rs. Nvidia does not publish performance numbers, integration details or a maintenance commitment for either, and there is nothing measured to compare against writing the same kernel in CUDA C++ today.

For a Rust-native inference stack, keeping the kernel in Rust removes the second toolchain, the separate build step and the boundary where types stop being checked. That is a real gain for a small team that already lives in Cargo, and a real risk on the alpha side, because a pinned nightly and a shifting API is not a place to build a production dependency yet. Cutile-rs on stable Rust is the softer entry point, and mistral.rs is the closest reference build for a public codebase.

Source

Nvidia Technical Blog: Introducing CUDA Rust: two tracks for writing GPU kernels by Sri Koundinyan, Melih Elibol and Jonathan Bentz, 8 September 2026.