NVIDIA releases cuda-oxide, a Rust-to-GPU compiler

NVIDIA has open-sourced cuda-oxide, an experimental compiler that turns Rust functions directly into CUDA GPU code. Until now, developers of GPU applications had to choose between the convenience of Rust and the performance of CUDA.

How it works

cuda-oxide is a custom backend for the rustc compiler. It takes functions annotated with `#[kernel]` and runs them through a chain of transformations: Rust → Stable MIR → Pliron IR → LLVM IR → PTX (Parallel Thread Execution, the machine-level code for NVIDIA GPUs).

The key idea: the developer writes the code once. A single command, `cargo oxide build`, compiles both the CPU part (host) and the GPU part (device). Previously, developers had to mix Rust for the host with C/CUDA for the kernels.

Why it matters

Rust on the GPU is a holy grail for developers. CUDA has historically been tied to C/C++. Rust offers memory safety without a garbage collector, which is critical for high-performance computing.

• SIMT (Single-Instruction Multiple-Thread) programming becomes more accessible
• One toolchain ecosystem: Cargo handles all targets the same way
• Memory-safety guarantees even for GPU code
• Easier to integrate CUDA into existing Rust projects
• The gap between the host language and the kernel language narrows

cuda-oxide v0.1.0 is a first step. The tool is experimental, so production workloads may run into limitations and bugs.

What this means

NVIDIA is investing in the Rust ecosystem for GPU computing. That signals a shift away from C/C++ as a requirement for parallel computing. If cuda-oxide survives to a production release, it could become the standard for new CUDA projects in Rust and reshape the GPU development landscape.