NVIDIA/cuda-rust

CUDA Rust is NVIDIA's early-stage platform for programming NVIDIA GPUs in Rust. It combines a shared host runtime for driver bindings, contexts, streams, buffers, events and async work with two ways to write device kernels, so the main fit is Rust developers doing GPU-accelerated numerical or systems work rather than an application end user.

Two kernel models

The SIMT path (cuda-oxide) is a custom rustc codegen backend plus cargo helper and device crates for the traditional CUDA threading model; programmers control thread and block indices, shared memory and synchronization, and code is compiled ahead of time from Rust to PTX. Its toolchain is a pinned stable Rust 1.99.1 managed by cargo oxide, with rustc-dev and llvm-tools CUDA Rust README.

The Tile path (cutile-rs) is a proc-macro plugin and crate set where tensors are partitioned into sub-tensors, kernels compute on whole tiles, and the compiler decides tile mapping and tunes per GPU; it JIT-compiles on first launch from Rust through Tile IR to cubin and works on stable Rust 1.89 or newer CUDA Rust README. Tile removes explicit thread indexing and synchronization within a kernel; SIMT is recommended for sparse or pointer-chasing workloads, while Tile suits dense numerical tensor math. Both can run on the same stream over shared buffers.

What the runtime provides

The host side consists of cuda-bindings for CUDA driver API bindings, cuda-core (with cuda-core-derive) for contexts, streams, memory, modules, events and launches, and cuda-async for lazy operations, futures, streams and graphs CUDA Rust README. Both kernel models extend Rust ownership across the launch boundary: mutable buffers are split into disjoint pieces, immutable buffers are shared, and launchers keep borrows alive while GPU work is in flight.

Requirements and current status

Using either model requires a compatible NVIDIA GPU and CUDA driver; the SIMT path additionally needs its pinned toolchain setup, while Tile requires only stable Rust 1.89+. Installation and reference material are in the separate cuda-oxide and cutile-rs books, and the runtime crates are documented in the cutile-rs book CUDA Rust docs.

The project describes itself as early-stage research: bugs, incomplete features and API breakage are expected, so it is not a stability commitment for production work CUDA Rust README. The whole repository, including both kernel models and shared host crates, is Apache-2.0 licensed LICENSE.

HotRepo's take

A single daily Trending observation on 2026-10-29 recorded 3,866 stars and +32 for that day; the supplied board history is too short to support any trend claim.