NVIDIA routes Rust GPU kernels to PTX through a custom rustc codegen backend

`cargo oxide run` drives a build in which a custom rustc codegen backend intercepts compilation. Functions marked `#[kernel]` go through Rust MIR, then the community Pliron IR framework, then LLVM IR, and come out as PTX; everything else in the crate is handed to the standard backend ^[11]. The GPU dialects sitting on top of Pliron are NVIDIA's, and every dialect and transform stays in Rust until the standard LLVM backend takes over ^[12]. The first `cargo oxide run` builds that codegen backend, which NVIDIA says takes a while, and later runs reuse the cache ^[15]. A CI runner with a cold cache pays that build every time.

`cargo oxide doctor` is the gate. It checks for Linux, a GPU with compute capability 8.0 or later, a CUDA toolkit 12.x or newer, clang with its libclang headers, the pinned nightly toolchain, and optionally a system LLVM ^[13]. The install command pulls cargo-oxide from the NVlabs GitHub repository and names the pin in the invocation: `cargo +nightly-2026-04-03` ^[14]. The post is dated September 2026, which puts the required nightly roughly five months behind the announcement ^[22].

The scaffolded kernel shows where Rust's guarantees land in device code. The output parameter is a `DisjointSlice<f32>`, and the store happens only inside `if let Some(c_elem) = c.get_mut(idx)`, while the two inputs are read with the plain `usize` returned by `idx.get()` ^[18]. Above the function, `#[launch_bounds(256)]` caps threads per block so the compiler can budget registers, and `#[launch_contract(domain = 1, block = (256, 1, 1))]` declares one-dimensional indexing with 256-thread blocks ^[19].

On which track to build, NVIDIA's guidance is to reach for Tile first, because the compiler decides how tiles map onto each architecture and the source then does not encode architecture-specific choices; you drop to SIMT when you need that control or want to manage memory and threads yourself ^[9]. Tile is available on the C++ and Python frontends too, so that advice is not specific to Rust ^[8].

The evidence in the post is correctness. The worked example is elementwise addition over 1,024 floats, written on both tracks, both complete programs that run and print the same line ^[17], and the scaffolded version reports `PASSED: all 1024 elements correct` ^[16]. The post does not include performance numbers. Whether PTX from cuda-oxide costs anything against PTX from CUDA C++ on a kernel that matters is a measurement the adopting team makes on its own workload, on its own architecture, with its own occupancy limits.

NVIDIA also says it plans to support inter-language interop, so a team that picks the Rust frontend is not meant to be locked out of the others ^[10].