NVIDIA introduces CUDA Rust: cuda-oxide (SIMT) and cutile-rs (Tile)

NVIDIA’s Developer Blog (2026-09-08) introduces CUDA Rust as two tracks for writing GPU kernels in Rust that compile toward PTX / CUDA Tile IR, not wrappers around foreign kernel source. Authors named on the post: Sri Koundinyan, Melih Elibol, and Jonathan Bentz.

Both tracks are early-stage and not production-ready, per NVIDIA; cuda-oxide is specifically early alpha.

Two tracks

cuda-oxide (SIMT): Custom rustc codegen backend; #[kernel] via Rust MIR → Pliron IR → LLVM → PTX; host and device in one file. Install path in the post uses a pinned nightly (cargo +nightly-2026-04-03 install --git https://github.com/NVlabs/cuda-oxide.git cargo-oxide). Prerequisites as stated: Linux; GPU compute capability 8.0+; CUDA toolkit 12.x or newer; clang/libclang; optional system LLVM (cargo oxide doctor checks).

cutile-rs (Tile): #[cutile::module] embeds kernel AST; JIT through CUDA Tile IR; crate/cargo add name cutile. Prerequisites as stated in September: GPU CC 8.0+; CUDA 13.3; stable Rust 1.89+; Linux; no nightly and no own LLVM. (As of 0.5.0 the README sets a minimum Tile IR version per GPU instead, from 13.2 on Ampere, Ada and most Blackwell parts; see the update below.)

NVIDIA’s guidance in the post: prefer Tile first (architecture mapping in the compiler); drop to SIMT when you need thread/memory control. Language choice is separate from the programming model; inter-language interop with C++/Python is planned.

What this brief does not claim

• No GFLOPS / speedups. The post’s side-by-side vecadd over 1,024 floats (both print a correctness pass) is a correctness demo only.
• No production readiness.
• External use of cutile-rs in Hugging Face Grout and mistral.rs is NVIDIA’s claim in the post, not independent verification here. mistral.rs (Eric Buehler) is not affiliated with Mistral AI.

Licenses (from repo LICENSE files)

Checked on GitHub: cuda-oxide and cutile-rs both carry Apache License 2.0 (NVIDIA copyright headers). The license is taken from the repositories, not from the blog post.

Who should care

Rust GPU teams watching official NVIDIA frontends for SIMT vs Tile should read the Developer Blog post and the two NVlabs repos, and treat toolchain pins (nightly-2026-04-03, Rust 1.89+, CUDA 12.x vs 13.3) as post-dated requirements to recheck before depending on them.

Update, 5 November 2026: cuTile Rust 0.5.0 adds CUDA 13.5 tile-library calls, and the toolkit floor is per GPU

NVIDIA has released cuTile Rust 0.5.0, published on GitHub and as cutile 0.5.0 on crates.io on 5 November (the in-tree changelog dates it 4 November). Both tracks now live in one repository, NVIDIA/cuda-rust, still under Apache-2.0, and the old NVlabs/cutile-rs repository is archived. The project’s README still calls CUDA Rust “an early-stage research project” and warns of “bugs, incomplete features, and API breakage.”

What 0.5.0 adds, per the release notes:

• Tile libraries, in public preview for CUDA 13.5. Kernels can declare and call functions from compiled tile libraries through #[cutile::foreign] unsafe fn. These declarations and calls are unsafe, and pointer operands and shape or tuple constant parameters are not supported yet. The README says this needs Tile IR 13.5 and a tileiras that lists the feature, run with --preview=simt.
• Unsafe GPU communication in cutile::tileir: unicast and multicast pointers, multicast load-reduce, reduce and store, and system-scope signaling. The README limits unicast and multicast to sm_100, sm_103 and sm_107 GPUs on an NVLink SHARP-capable fabric, and TMA data-validity checks to sm_107.
• A shared host runtime for the Tile and SIMT tracks. This is a breaking upgrade from 0.4: update the Tile and host crates to 0.5.0 together, move cuda_core::simt imports to the crate root, switch to the unified cuda_async API and replace the old debug flags with CompileOptions::codegen. Some kernel input and output traits now need an unsafe impl with documented obligations.
• Likelihood hints are accepted but, per the notes, “discarded in 13.5 bytecode and have no backend performance effect.”

The prerequisite line above is out of date. The cuTile README no longer gives one CUDA version. It lists a minimum Tile IR version per GPU: 13.2 for Ampere and Ada (sm_80 to sm_89) and for Blackwell sm_100, sm_103, sm_110, sm_120 and sm_121; 13.3 for Hopper (sm_90); and 13.4 for sm_107. It still recommends CUDA 13.3, which FP4 packing and block-scaled MMA require, and stable Rust 1.89 or newer. CUDA 13.5 is needed only for the new 13.5 features, not for all cuTile work. On the SIMT side, the cuda-oxide README in the new repository now pins stable Rust 1.99.1 rather than the nightly named above, and asks for CUDA Toolkit 13.0 or newer.