NVIDIA’s engineers released CUDA Oxide 0.3, their experimental compiler for writing CUDA in Rust.
What’s new
- Runs on stable Rust (1.99.1 right now). Earlier versions needed nightly.
- The host runtime is unified with cutile-rs in the cuda-rust monorepo.
- An experimental cute-rs, a Rust take on CUTLASS CuTe, runs Qwen3-8B (BF16 + NVFP4) and Qwen3-TTS end to end on Blackwell GPUs.
- Full cuda-gdb DWARF debugging, in-kernel event tracing, register and spill diagnostics, and optimizations for Blackwell and Hopper.
My take
Dropping the nightly requirement is the big deal. Real models running end to end in Rust on CUDA is a good sign for local AI folks who’d rather not live in C++. It’s still experimental, so treat it that way.
God Bless, and Happy hacking!
