CUDA Rust lets developers write GPU kernels in Rust instead of C++. NVIDIA released it on September 8, 2026, as two open source projects: cuda-oxide and cutile-rs. Both compile Rust straight to PTX, the assembly CUDA GPUs run. Neither wraps a C++ layer underneath. Teams that already build inference and data pipelines in Rust gain the most. CUDA Rust closes the last gap that kept their GPU kernels stuck in C++.
What CUDA Rust Actually Is
CUDA Rust is not one library. It is NVIDIA’s umbrella term for two separate projects. Each lets developers write GPU kernels natively in Rust. Both compile straight to PTX, the intermediate assembly CUDA GPUs execute. Neither project wraps CUDA C++ under the hood. Both use the same compiler toolchain NVIDIA relies on internally. That is why NVIDIA calls this a first class frontend, not a community binding.
NVIDIA ties the effort to one driver. Teams increasingly write the systems layer of AI infrastructure in Rust. That shift touches how AI is reshaping software development practices well beyond GPU code. GPU kernels stayed the one piece still locked to C++. Closing that gap lets a team write an entire inference pipeline in one language. It removes the need for a separate C++ kernel layer underneath a Rust service.
Two Tracks, Two Different Tradeoffs
NVIDIA split the effort into two tracks. Each track solves a different problem. Cuda-oxide implements the classic SIMT model. CUDA C++ developers already know this model, where each thread runs the kernel independently. Cuda-oxide works as a custom rustc codegen backend. It routes code through Rust’s MIR, the Pliron IR framework, and LLVM before emitting PTX. Cuda-oxide requires a pinned nightly Rust toolchain, CUDA 12.x or newer, and a GPU with compute capability 8.0 or higher. It gives developers the same low level control over thread indexing and memory layout that CUDA C++ offers. Rust’s type system backs that control.
Cutile-rs takes the opposite approach. It targets CUDA’s newer Tile model, where a kernel runs once over a whole sub-tensor instead of per thread. The compiler handles thread mapping and memory layout automatically. It runs on stable Rust 1.89 or newer and needs CUDA 13.3. It installs as a normal crates.io package with cargo add cutile. That maturity already shows in production. Cutile-rs runs inside Hugging Face’s Grout inference engine and in mistral.rs, as NVIDIA confirmed in its official CUDA Rust announcement.
Both tracks catch bugs at compile time instead of runtime. They catch memory aliasing bugs and data races before code ever executes. Rust’s ownership model makes that possible. Cuda-oxide uses a DisjointSlice type that enforces exclusive per-thread access. Cutile-rs uses tensor partitioning instead, so ownership of each tile transfers and two threads can never touch it at once. That is the real pitch: the compiler now catches bugs that used to surface only in production CUDA C++ code.
How To Try It Today
Getting started depends on your workload. Cutile-rs offers the lower friction path. It needs only stable Rust and a standard cargo add cutile command, which makes it a solid first experiment. Teams can test tile level kernels without touching a nightly toolchain. Cuda-oxide demands more setup. It requires a specific pinned nightly build (nightly-2026-04-03), plus clang with libclang headers installed. You install it manually, through the cargo-oxide subcommand on NVIDIA’s GitHub repository, not crates.io.
Both tracks currently run only on Linux. Both require a GPU with compute capability 8.0 or higher, which covers NVIDIA’s Ampere generation and newer. Teams on older GPUs, or on Windows, will need to wait or test elsewhere. NVIDIA says picking one frontend will not lock a team out of the other. Interoperability between cuda-oxide, cutile-rs, and existing CUDA C++ code is on the roadmap, though it isn’t fully built yet.
What’s Still Missing
NVIDIA is direct about where these projects stand. Neither one is production-ready as a full CUDA C++ replacement. NVIDIA’s own documentation calls both alpha software. Cuda-oxide’s pinned nightly toolchain creates a real adoption barrier, since many teams enforce strict toolchain policies. Its manual GitHub installation process reflects that early stage. Rust tooling still lacks the profiler integration and debugging support CUDA C++ developers expect.
Cutile-rs stands further along, but its abstractions have limits. Some workloads need the fine-grained control cuda-oxide and CUDA C++ both offer. No compiler can make that tradeoff for you automatically. Evaluate CUDA Rust the way you would evaluate any new addition to your team’s software development practices. Treat it as a way to remove a class of bugs. It is not yet a finished replacement for a decade of CUDA C++ tooling.
Key Takeaways
- NVIDIA released CUDA Rust on September 8, 2026, as two separate open source projects: cuda-oxide and cutile-rs.
- Cuda-oxide targets the SIMT model with explicit thread control and needs a pinned nightly Rust toolchain plus CUDA 12.x.
- Cutile-rs targets the newer Tile model, runs on stable Rust 1.89 and CUDA 13.3, and installs through crates.io.
- Both catch memory aliasing and data race bugs at compile time using Rust’s ownership model instead of surfacing them at runtime.
- Neither project is production-ready yet, though cutile-rs already runs in Hugging Face’s Grout engine and in mistral.rs.
Frequently Asked Questions About CUDA Rust
What is CUDA Rust?
CUDA Rust is NVIDIA’s umbrella term for two open source projects: cuda-oxide and cutile-rs. Both compile Rust code natively to PTX for NVIDIA GPUs, with no C++ layer underneath.
What is the difference between cuda-oxide and cutile-rs?
Cuda-oxide implements the per-thread SIMT model. It gives developers explicit control over thread indexing and memory, and needs a pinned nightly Rust toolchain. Cutile-rs implements the newer Tile model, where the compiler manages thread mapping automatically, and runs on stable Rust.
Is CUDA Rust ready for production use?
Not generally. NVIDIA describes both projects as alpha software. Cutile-rs already runs in production, inside Hugging Face’s Grout engine and mistral.rs. Cuda-oxide sits at an earlier stage and needs more manual setup.
Do I need to give up CUDA C++ to use CUDA Rust?
No. NVIDIA says choosing a Rust frontend will not lock a team out of CUDA C++ or CUDA Python. Full interoperability between the frontends is still on the roadmap.
What hardware and software do I need to try it?
Both tracks currently require Linux and a GPU with compute capability 8.0 or higher. Cuda-oxide needs a pinned nightly Rust toolchain and CUDA 12.x or newer. Cutile-rs needs stable Rust 1.89 or newer and CUDA 13.3.
Why did NVIDIA build this instead of relying on existing community Rust GPU projects?
Rust already runs much of the systems layer under modern AI infrastructure, including inference servers and data loaders. GPU kernels stayed the one piece locked to C++. NVIDIA’s own compiler support closes that gap directly.
Final Thoughts
CUDA Rust will not replace CUDA C++ this year, and NVIDIA is not claiming otherwise. It gives Rust-heavy teams a real, vendor backed path to write GPU kernels without a C++ detour. Start with cutile-rs on a noncritical kernel if your stack already leans on Rust. Watch NVIDIA’s release notes for the interoperability and stability updates that alpha status still implies.
Photo by Mariia Berezovsky: Unsplash
Johannah Lopez is a versatile professional who seamlessly navigates two worlds. By day, she excels as a SaaS freelance writer, crafting informative and persuasive content for tech companies. By night, she showcases her vibrant personality and customer service skills as a part-time bartender. Johannah's ability to blend her writing expertise with her social finesse makes her a well-rounded and engaging storyteller in any setting.
