<p>NVIDIA Labs has released <a href="https://github.com/NVlabs/cuda-oxide" target="_blank" rel="noopener nofollow">cuda-oxide 0.1</a>, an experimental open-source compiler that lets you write SIMT GPU kernels in standard Rust and compile them directly to PTX — no DSLs, no foreign language bindings.</p>
<h2>How it works</h2>
<p>cuda-oxide is a custom <code>rustc</code> codegen backend. Functions marked with <code>#[kernel]</code> go through a pipeline: Rust MIR → Pliron IR → LLVM IR → PTX. A <code>#[cuda_module]</code> macro embeds the resulting device artifacts directly into the host binary, so host and device code live in the same source file.</p>
<p>The compiler supports generics, closures with captures, and the full range of GPU intrinsics you’d expect: warp operations, shared memory, TMA, atomics, and thread block clusters. There’s also an async execution layer built around <code>DeviceOperation</code> graphs for GPU programming.</p>
<h2>Performance</h2>
<p>The project ships 46 examples, including a GEMM kernel that hits 868 TFLOPS on an NVIDIA B200 — about 58% of cuBLAS’s speed-of-light. Not parity, but respectable for an alpha.</p>
<h2>Requirements and status</h2>
<p>You’ll need Rust nightly, CUDA Toolkit 12.x or later, and LLVM 21+. The project is explicitly in alpha: bugs are expected, and the API will break. That said, it’s under active development from NVIDIA Labs, which gives it more institutional backing than most experimental GPU Rust projects.</p>
<p>For anyone who’s wrestled with CUDA C++ or tried to make existing Rust GPU crates work, this is worth watching. Writing kernels in idiomatic Rust with real ownership semantics and generics is a meaningful quality-of-life improvement — if the project matures.</p>


Collections

NVIDIA Labs has released cuda-oxide 0.1, an experimental open-source Rust-to-CUDA compiler that allows writing SIMT GPU kernels in standard Rust, compiling them to PTX without DSLs or foreign bindings. It works as a custom rustc codegen backend, routing Rust MIR through Pliron IR and LLVM IR to PTX. The compiler supports generics, closures, warp operations, shared memory, TMA, atomics, and thread block clusters. A sample GEMM kernel achieves 868 TFLOPS on an NVIDIA B200 (~58% of cuBLAS). The project requires Rust nightly, CUDA Toolkit 12.x+, and LLVM 21+, and is explicitly in alpha with an unstable API.

NVIDIA Labs releases cuda-oxide 0.1, an experimental Rust-to-CUDA compiler