NVIDIA Announces CUDA Rust: Compile-Time-Safe GPU Kernels

The systems layer of artificial intelligence infrastructure is increasingly adopting the Rust programming language, ranging from inference engines and drivers to agent runtimes. While components like NVIDIA's Nova Linux driver, the core of NVIDIA Dynamo, and NVTX bindings have already embraced Rust, the GPU kernel remained a notable exception until this latest major announcement aimed at bridging the gap.

The initiative introduces two distinct tracks that mirror CUDA's established programming models. The SIMT model allows developers to define the behavior of a single thread and launch thousands of them, similar to CUDA C++. Meanwhile, the newer Tile model enables developers to describe operations on a data tile, leaving the Tile IR compiler to handle thread mapping and memory layout optimization automatically.

On the implementation side, cuda-oxide operates as a custom rustc codegen backend, routing functions through Rust MIR, the Pliron IR framework, and LLVM IR down to PTX. It requires Linux, a GPU with compute capability 8.0 or higher, CUDA 12.x, and a pinned nightly toolchain. Conversely, cutile-rs operates at a higher abstraction level using macro-based AST embedding and JIT compilation via CUDA Tile IR.