NVIDIA releases CUDA Oxide 0.3 with improved Rust and Blackwell support

NVIDIA software engineers released CUDA Oxide 0.3 in late October. This experimental Rust to CUDA compiler now uses the official stable Rust 1.99.1 version, whereas previous versions required Rust nightly builds.

Image: Phoronix

The update includes an experimental version of cute-rs, which is the CUTLASS CuTe in Rust. NVIDIA says cute-rs is able to run Qwen3-TTS and Qwen3-8B at BF16 + NVFP4. These models are being tested on NVIDIA Blackwell GPUs. The release also brings compiler and performance optimizations for Hopper and Blackwell GPUs.

For debugging, version 0.3 adds full cuda-gdb DWARF support, in-kernel event tracing, and diagnostics for registers and spills. Additionally, the host runtime is now combined with cutile-rs within the main repository.