NVIDIA announced CUDA Rust, introducing two tracks for writing GPU kernels in Rust: cuda-oxide (a custom rustc backend targeting PTX via Pliron/LLVM) and cutile-rs (a stable-Rust tile-based programming framework). Both enforce compile-time memory safety, with cutile-rs already adopted by HuggingFace's Grout and mistral.rs. NVIDIA plans cross-ecosystem interoperability with CUDA C++ and CUDA Python.
Background
NVIDIA has been investing in expanding its GPU ecosystem beyond CUDA C/C++ and Python, seeking to attract systems-language developers who prioritize performance and memory safety.
- Source
- Lobsters
- Published
- Sep 9, 2026 at 09:21 PM
- Score
- 8.0 / 10