Nvidia announces native GPU programming in Rust
Points and comments are a snapshot, not live.
NVIDIA releases two Rust-based tracks for writing native GPU kernels.
NVIDIA announced CUDA Rust, with two tracks for writing GPU kernels in Rust. The SIMT track (cuda-oxide) uses a custom rustc codegen backend to compile `#[kernel]` functions to PTX, requiring nightly Rust and Linux. The Tile track (cutile-rs) works on stable Rust via JIT compilation through CUDA Tile IR, letting the compiler handle architecture mapping. Both provide compile-time memory safety for GPU kernels. Neither is production-ready; cuda-oxide is early alpha, while cutile-rs is published on crates.io and already used in HuggingFace's Grout and mistral.rs. NVIDIA plans to continue maturing both projects into 2027.
What commenters are saying
Commenters expressed cautious optimism about NVIDIA's Rust GPU push, noting current limitations. Several users praised cudarc for its 1-to-1 mapping with CUDA, calling the new dialect harder to rely on for debugging. One commenter noted cuda-oxide is Linux-only and requires async, preventing them from trying it. Some questioned the benefit, arguing kernels were "meant to be written in C" and that higher-level DSLs like Triton are preferable. Others countered that C is just a default, not a mandate, and that Rust's safety guarantees are valuable. A subthread flagged the article's prose as AI-generated, sparking debate about NVIDIA's use of LLMs for documentation.