In a watershed development for systems programming and accelerated computing, NVIDIA on September 16, 2026 officially announced native GPU programming in Rust. The announcement marks the culmination of a multi-year effort to modernize the accelerated computing stack, providing developers with memory-safe, ergonomic, and zero-cost abstractions for authoring custom kernels directly on Nvidia GPUs.
For over eighteen years, CUDA C and C++ have served as the undisputed foundation of GPU computing. However, as frontier foundation models have grown to hundreds of billions of parameters, the software infrastructure surrounding inference and distributed training has grown overwhelmingly complex. Runtimes such as vLLM, TensorRT-LLM, and Triton continually battle buffer overflows, out-of-bounds shared memory accesses, and race conditions across thread warps—defects that are notoriously difficult to reproduce and debug at scale.
To resolve this structural challenge without degrading hardware throughput, NVIDIA introduced two complementary development tracks: cuda-oxide and cutile-rs. The first, cuda-oxide, is designed for low-level Single Instruction, Multiple Threads (SIMT) programming. It implements a custom rustc codegen backend leveraging the Pliron intermediate representation framework and LLVM to lower Rust functions directly into Parallel Thread Execution (PTX) assembly. To uphold Rust's strict safety guarantees on massively parallel hardware, cuda-oxide introduces 'DisjointSlice' abstractions and explicit launch contracts that mathematically prevent memory aliasing across concurrent threads.
The second track, cutile-rs, targets high-level tile-based GPU computing. Designed to run on stable Rust 1.89+ alongside standard CUDA 13.3 drivers, cutile-rs allows developers to write tensor computations in terms of multi-dimensional tiles rather than managing individual thread indexes and warp shuffle instructions manually. The compiler's JIT backend handles thread mapping, register allocation, and asynchronous tensor copy pipelines automatically, leveraging Rust's ownership and move semantics to ensure that only one warp can mutate a given shared memory tile at any point in time.
Performance benchmarks presented by NVIDIA systems engineers Melih Elibol and Jonathan Bentz demonstrate that memory safety does not require sacrificing performance. In microbenchmarks evaluating FP8 matrix multiplications (GEMM) and causal attention kernels across Nvidia H100 and B200 GPUs, kernels compiled with cutile-rs achieved 99.2% of the raw FLOP utilization of highly hand-tuned CUDA C++ kernels, while cutting debugging and verification time by more than half.
The open-source ecosystem has rapidly mobilized around the release. Hugging Face announced that its new Grout modular inference engine has already transitioned its custom attention heads to cutile-rs, while the popular local inference runtime mistral.rs has merged CUDA Rust support into its mainline repository. Furthermore, NVIDIA stressed that CUDA Rust is designed with seamless cross-language Foreign Function Interface (FFI) bindings, allowing engineering teams to drop Rust-authored kernels directly into existing PyTorch and CUDA C++ pipelines without rewriting their entire infrastructure.
By bringing Rust's legendary type system, affine typing, and compile-time verification to GPU microarchitectures, NVIDIA has addressed one of the most glaring vulnerabilities in the AI infrastructure landscape, setting a new standard for robust, high-performance foundation model deployment.