A team evaluating CUDA Rust needs to distinguish the current repository from the September announcement. The documentation has evolved, so toolchain details should be checked against the chosen component and revision.
Start a CUDA Rust evaluation with current requirements
The September blog described a pinned nightly toolchain for cuda-oxide. The current combined repository instead documents a pinned stable toolchain managed by cargo oxide, while cutile-rs uses stable Rust. This change is a reason to follow the current component documentation rather than reuse a historical setup instruction.
https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/
https://github.com/NVIDIA/cuda-rust
Understand what the ownership rules cover
The repository explains that mutable buffers are divided into disjoint pieces and immutable buffers can be shared across a GPU launch. It distinguishes the direct thread control of SIMT from Tile’s implicit indexing and scheduling. These are documented mechanisms; they are not proof that every application written with the tools is free from all security or correctness problems.
https://github.com/NVIDIA/cuda-rust
Treat experiments as experiments
The current repository warns of incomplete functionality, bugs and changing APIs. Before considering a workload, a team should identify the exact component, its requirements and the behaviour it needs to validate. NVIDIA’s development plans do not establish a fixed date when every workload will be suitable for production.
https://github.com/NVIDIA/cuda-rust
Sources and further reading
https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/
https://github.com/NVIDIA/cuda-rust