cuda-oxide

cuda-oxide lets you write GPU code in plain Rust and compiles it straight to CUDA. Normally, programming an NVIDIA GPU means CUDA C or C++ or a special DSL, this NVIDIA Labs project compiles standard Rust down to PTX, the low-level bytecode the GPU actually runs. You get generics, closures, GPU intrinsics, even async, in a language with real memory safety.

There's nothing to host, it's a compiler toolchain you install and build against. The ops burden is the real story: this is low-level GPU work, so you need an NVIDIA GPU, the CUDA toolkit, and the patience to deal with an early compiler. It is not a smooth on-ramp, it's a power tool for people who already know what PTX is.

Rust developers doing GPU compute who don't want to drop into C++ are the audience. If you're happy in CUDA C or using higher-level ML frameworks that hide the GPU entirely, you don't need this. The appeal is writing kernels in one safe language instead of context-switching into C++ every time you touch the hardware.

The catch: it's explicitly early-alpha. NVIDIA's own README warns of bugs, missing features, and breaking API changes. It's free and Apache-licensed and the idea is excellent, but this is a watch-and-experiment project, not something to put under production kernels yet. Treat it as a preview of where Rust-on-GPU is heading.