{"slug": "nvidia-introduces-cuda-rust-for-writing-gpu-kernels", "title": "NVIDIA Introduces CUDA Rust for Writing GPU Kernels", "summary": "NVIDIA introduced CUDA Rust, two new approaches that let developers write GPU kernels directly in Rust and compile them to PTX, the intermediate representation used by NVIDIA GPUs. The two tracks are cuda-oxide, which uses the traditional SIMT model with a custom rustc code-generation backend routing Rust kernel code through Rust MIR, Pliron, LLVM IR, and finally PTX, and cutile-rs, which uses NVIDIA's CUDA Tile programming model with tensor partitioning and Rust ownership rules so each tile gets exclusive access to its writable portion of a tensor, rejecting some aliasing and race-condition errors at compile time. NVIDIA said the models are intended to complement each other, plans interoperability between CUDA Rust, CUDA C++, and CUDA Python, and warned that feature coverage is incomplete and APIs will change as development continues through 2027 and beyond.", "body_md": "NVIDIA is expanding CUDA with native Rust support, introducing two new approaches that let developers write GPU kernels directly in Rust instead of relying on kernels written in CUDA C++ or another language.\n\nUntil now, Rust applications could launch CUDA kernels, but the GPU-side kernel code was typically written elsewhere. NVIDIA’s new CUDA Rust effort closes that gap by compiling Rust GPU kernels directly to PTX, the intermediate representation used by NVIDIA GPUs.\n\nThe company is pursuing two separate tracks: [cuda-oxide](https://github.com/NVlabs/cuda-oxide) and [cutile-rs](https://github.com/NVlabs/cutile-rs). Both target Rust developers, but they follow different GPU programming models.\n\nThe first, cuda-oxide, uses the traditional SIMT model familiar from CUDA C++. Developers explicitly work with threads, blocks, indexing, and memory, while the project provides a custom rustc code-generation backend that routes Rust kernel code through Rust MIR, Pliron, LLVM IR, and finally to PTX.\n\nThe second approach, cutile-rs, works at a higher abstraction level using NVIDIA’s CUDA Tile programming model. Instead of describing what individual GPU threads should do, developers operate on tiles of data, while the compiler handles details such as how those tiles map onto physical GPU threads.\n\nMoreover, cutile-rs uses tensor partitioning and Rust ownership rules. Each tile gets exclusive access to its writable portion of a tensor, allowing some aliasing and race-condition errors to be rejected at compile time instead of during GPU execution.\n\nNVIDIA says the two models are intended to complement each other rather than compete. Tile is positioned as the preferred starting point when developers do not need direct control over thread mapping and memory behavior, while SIMT remains available for workloads where that lower-level control matters.\n\nThe company also plans interoperability between CUDA Rust, CUDA C++, and CUDA Python, meaning developers should eventually be able to mix the different CUDA frontends instead of committing an application entirely to one language.\n\nFinally, NVIDIA warns that feature coverage is incomplete and APIs will change as development continues. The company plans to expand and mature CUDA Rust through 2027 and beyond.\n\nFor additional details, [see NVIDIA’s announcements](https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/).", "url": "https://wpnews.pro/news/nvidia-introduces-cuda-rust-for-writing-gpu-kernels", "canonical_source": "https://linuxiac.com/nvidia-introduces-cuda-rust-for-writing-gpu-kernels/", "published_at": "2026-09-18 12:58:06+00:00", "updated_at": "2026-09-18 13:27:56.118850+00:00", "lang": "en", "topics": ["ai-infrastructure", "developer-tools", "ai-chips"], "entities": ["NVIDIA", "CUDA Rust", "cuda-oxide", "cutile-rs", "PTX", "Rust", "CUDA C++", "CUDA Python"], "alternates": {"html": "https://wpnews.pro/news/nvidia-introduces-cuda-rust-for-writing-gpu-kernels", "markdown": "https://wpnews.pro/news/nvidia-introduces-cuda-rust-for-writing-gpu-kernels.md", "text": "https://wpnews.pro/news/nvidia-introduces-cuda-rust-for-writing-gpu-kernels.txt", "jsonld": "https://wpnews.pro/news/nvidia-introduces-cuda-rust-for-writing-gpu-kernels.jsonld"}}