Nvidia Opens CUDA to Rust With Two Competing Compilers

Nvidia introduced two separate ways to write GPU kernels in Rust rather than CUDA C++, publishing both as open-source projects around a Sept. 8, 2026 developer-blog post and a RustConf 2026 talk in Montreal. Neither compiles code written in another language; both take Rust source and produce PTX, the intermediate instruction format Nvidia's driver turns into machine code for the GPU, according to Nvidia's announcement.

Two tracks, two different bets on where safety comes from

The first track, cuda-oxide, mirrors traditional CUDA's SIMT model, where a programmer writes what one thread does and launches thousands of copies of it. It works through a custom rustc backend that routes Rust's mid-level intermediate representation through an internal format called Pliron and then LLVM to reach PTX, according to the project's GitHub repository. The second track, cutile-rs, uses a newer tile-based model in which a programmer describes what happens to one block, or tile, of data, and a separate compiler called Tile IR decides how to map that across threads and memory. Both require a GPU of compute capability 8.0 or higher, which covers Nvidia's Ampere generation and everything released since.

cutile-rs claims 7 TB/s on a B200, in Nvidia's own numbers

cutile-rs is the more finished of the two. It is published on the crates.io Rust package registry, runs on stable Rust 1.89 with no custom LLVM build required, and Nvidia's own benchmark numbers, which The Terminal could not independently verify, claim 7 terabytes a second on element-wise operations and 2 petaflops on a GEMM matrix multiplication, both run on a B200 GPU, according to the cutile-rs repository. The same source credits Melih Elibol, Jared Roesch, Isaac Gelado, Eric Buehler and Michael Garland, Nvidia's senior director of research, as the project's paper authors. Buehler is also the creator of mistral.rs, one of the two production users cutile-rs names for itself.

1. Qwen3-4B on RTX 5090171 tokens/sec
2. Qwen3-32B on B20082 tokens/sec

Source: cutile-rs GitHub README, accessed 2026-09-16

cuda-oxide is alpha, and Nvidia's own sources disagree on its CUDA floor

cuda-oxide is earlier-stage. Its GitHub README describes the project as being "in an early stage (alpha)" with "bugs, incomplete features, and API breakage" to expect, and it lists requirements of a pinned Rust nightly toolchain with the rust-src, rustc-dev and llvm-tools components, clang and libclang version 21, and CUDA Toolkit 13.0 or higher. That last figure does not match Nvidia's own developer blog post announcing the project, which lists cuda-oxide's CUDA requirement as version 12.x or higher, a full major version below the README. The Terminal found no note on either page acknowledging the mismatch as of Sept. 16.

Track	Rust toolchain	CUDA floor	Status
cuda-oxide (SIMT)	Nightly, pinned	12.x+ per Nvidia's blog, 13.0+ per its README	Alpha
cutile-rs (Tile)	Stable 1.89+	13.3 recommended, 13.2 minimum	On crates.io; used in Hugging Face's Grout and mistral.rs

Already running inference outside Nvidia

cutile-rs is not just a demo. Hugging Face's Grout inference engine and Buehler's mistral.rs both use it in production, according to the cutile-rs repository, and the project's own microbenchmarks put the runtime cost of its safety checks at under 0.3 percent variation in a persistent GEMM workload. A faster, memory-safer kernel layer underneath a Rust inference runtime is exactly the pitch cutile-rs is making, the same problem local-inference projects like Colibri, which runs a 744-billion-parameter model on 16GB of consumer memory, are trying to solve from the software side. cuda-oxide has no equivalent production users listed. Elibol is presenting the pair's underlying safety model, in a talk titled "Fearless Concurrency on the GPU," at RustConf 2026, which runs Sept. 8 through 11 in Montreal, according to Nvidia's announcement. It is not the only systems-language performance claim in September 2026 that needs its baseline checked: Swift's JavaScriptKit bridging library claimed up to 40 times faster WebAssembly calls in the same window, a gain specific to code still on the older bridging path rather than the whole runtime. Nvidia has not said when, or whether, cuda-oxide will reach the stability cutile-rs already claims.