cuda-oxide v0.1.0 compiles Rust SIMT GPU Kernels Directly to PTX

NVlabs has released cuda-oxide v0.1.0, an experimental rustc codegen backend that compiles #[kernel]-annotated Rust functions directly to PTX, enabling developers to write CUDA SIMT GPU kernels in standard Rust without requiring C/C++ code or foreign — function bindings. The device output is PTX, the assembly — like intermediate used to target NVIDIA GPUs. This could make it easier for developers targeting NVIDIA hardware to author SIMT kernels in safe Rust.

cuda-oxide works by intercepting rustc at CodegenBackend::codegen_crate() and running a device — only compilation pipeline. The toolchain transforms Rust source through the rustc frontend into rustc_public (Stable MIR), lowers that to Pliron IR, converts to LLVM IR, and finally emits PTX as the device artifact. Pliron is presented as a Rust-native, MLIR-like intermediate representation that lets the compiler and its supporting tools be built with cargo rather than a C++ toolchain. That design choice is intended to reduce external build dependencies and streamline development of the entire stack inside Rust's ecosystem.

The project also offers single — source host+device compilation: a single cargo oxide command can build host and device code from the same Rust sources. That single — command workflow aims to simplify mixed host/device projects by keeping both sides in one Rust build system invocation.

cuda-oxide sits alongside other Rust GPU efforts but pursues a distinct goal of ‘‘bringing CUDA into Rust.’’ By contrast, rust-gpu targets SPIR‑V for Vulkan, rust-cuda emphasizes Rust-first ergonomics and NVVM IR, and projects such as CubeCL or std.offload follow different cross — compile or embedded — DSL approaches. NVlabs says it coordinated with rust-cuda maintainers and views these projects as complementary.