NVIDIA publishes CUDA Rust with two tracks for writing GPU kernels

Sources: primary, discussion
Summary: NVIDIA published CUDA Rust, which compiles Rust GPU kernels natively to PTX rather than wrapping CUDA C++. The post sets out two tracks for writing those kernels, and states the compiler rejects at build time the case where a kernel output aliases its input. NVIDIA also states that both projects are early-stage and neither is production-ready, with cuda-oxide in early alpha, that the SIMT track needs a pinned nightly toolchain plus LLVM, and that the Tile track needs CUDA 13.3, stable Rust 1.89 or later, and compute capability 8.0 or later.
Why it matters: The Rust path to PTX is a real compilation target rather than a wrapper, but NVIDIA labels it not production-ready, so it is something to evaluate on a pinned toolchain rather than build on.