NVIDIA introduces CUDA Rust with two tracks for writing GPU kernels

8 Sep 2026: NVIDIA published that CUDA Rust lets developers write GPU kernels in Rust compiled natively to PTX — two tracks, cuda-oxide (SIMT) and cutile-rs (Tile). Company technical blog is the primary. Both projects are early-stage; not production-ready.

8 Sep 2026: NVIDIA Technical Blog “Introducing CUDA Rust: Two Tracks for Writing GPU Kernels” (authors Sri Koundinyan, Melih Elibol, Jonathan Bentz) announces CUDA Rust — GPU kernels written in Rust and compiled natively to PTX, not merely launched from Rust wrappers. That dated company technical blog is the filing event.

Two tracks matching CUDA’s own models: (1) cuda-oxide — SIMT-style kernels via a custom rustc codegen backend (Rust MIR → Pliron IR → LLVM → PTX); early alpha; needs pinned nightly toolchain, Linux, GPU compute capability 8.0+, CUDA toolkit. (2) cutile-rs — Tile-based kernels on stable Rust 1.89+ with CUDA 13.3; published on crates.io; JIT through CUDA Tile IR.

NVIDIA says cutile-rs is already used outside NVIDIA in Hugging Face’s Grout inference engine and in mistral.rs; cuda-oxide remains early alpha. Both are early-stage and neither is production-ready — coverage incomplete, APIs will move (NVIDIA’s framing).

Safety pitch on the post: both tracks aim to catch memory-aliasing mistakes at compile time (DisjointSlice / launch contracts on SIMT; tensor partition ownership on Tile). This filing does not invent benchmark numbers or claim production readiness.

Context NVIDIA gives: Nova Linux driver in Rust, Dynamo Rust core, NVTX Rust bindings — kernels were the gap. Interop with CUDA C++/Python is planned so frontend choice does not lock ecosystems.

This filing does not invent release version numbers, pricing, Windows support, or performance versus CUDA C++ — those do not appear as hard claims on this primary. Distinct from the already-filed d-Matrix NVLink Fusion, Skild S1, and robotaxi-stack NVIDIA blog dumps.

A dated NVIDIA primary puts native Rust GPU kernels on the CUDA roadmap — systems/AI infra teams that already ship Rust finally get first-party kernel tracks, with the desk keeping the early-alpha / not-production-ready line clear.