News › ai-infrastructure
By Zayden R., September 17, 2026
NVIDIA has launched CUDA Rust, enabling developers to write GPU kernels directly in Rust. This development offers two tracks, cuda-oxide and cutile-rs, bridging a gap in Rust's GPU ecosystem.
NVIDIA has unveiled CUDA Rust, a significant stride in GPU programming that allows developers to write GPU kernels natively in Rust. Announced on September 8, 2026, this initiative introduces two programming tracks, cuda-oxide and cutile-rs, designed to cater to both advanced and general-purpose GPU programming needs.
CUDA Rust directly addresses a longstanding limitation in Rust's GPU ecosystem. Historically, Rust was often used as a wrapper for kernels written in other languages like CUDA C++. With CUDA Rust, developers can now compile native Rust kernels directly to PTX, NVIDIA's low-level assembly language for GPUs. This aligns well with the industry's increasing preference for Rust in systems programming, thanks to its ability to catch bugs at compile time without compromising on performance.
The cuda-oxide track adopts a SIMT (single instruction, multiple threads) model, similar to traditional CUDA C++ workflows. It employs a custom rustc codegen backend that compiles SIMT-style GPU kernels written in Rust directly to PTX using the Pliron IR framework and LLVM. However, cuda-oxide requires a pinned nightly toolchain and LLVM, indicating it's still in early alpha stages.
On the other hand, cutile-rs offers a more stable option, running on stable Rust 1.89+ with CUDA 13.3, without the need for a custom LLVM. It enables tile-based GPU programming where the compiler manages thread mapping and memory layout through CUDA Tile IR JIT compilation. This track is already being utilized in projects like HuggingFace's Grout inference engine and mistral.rs, showcasing its practical applicability.
Both cuda-oxide and cutile-rs emphasize memory safety at compile time. Cuda-oxide uses DisjointSlice and launch contracts to prevent aliasing, while cutile-rs leverages tensor partitioning and ownership to ensure exclusive access.
NVIDIA's plan to support inter-language interoperability between CUDA Rust, CUDA C++, and CUDA Python suggests a flexible future where developers are not locked into a single ecosystem. This is a solid improvement for those invested in Rust and GPU programming, as it offers more robust and diverse tools for developing high-performance applications.