Nvidia announces CUDA Rust: Write your GPU kernels in Rust!
Rust is not just a fad in GPU programming. Rust eliminates data races, which can be extremely difficult to debug when developing in CUDA, at compile time. While I don’t personally develop GPU kernels in my daily work with agents, having good inference infrastructure makes a big difference in terms of the performance and reliability of the APIs that I use.
Besides the stable Rust (1.89+) + tile solution adopted by Hugging Face (Grout inference engine) and mistral.rs, Nvidia also provides CUDA Rust, allowing you to write GPU kernels (code executed on the GPU) in either SIMT (using cuda-oxide, on nightly Rust) or Tile (using cutile-rs, on stable Rust 1.89+).
Tile kernels are statically compiled for all GPUs and have already been deployed!
Source: developer.nvidia.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles