NVIDIA宣布推出原生Rust GPU编程工具

NVIDIA于2026年9月宣布推出两个Rust项目，使开发者能够直接用Rust编写GPU内核代码 1。其中cuda-oxide采用自定义rustc代码生成后端，实现SIMT编程模型，而cutile-rs则在更高层级提供Tile编程模型 1。两个项目都将代码编译为PTX格式 1。

cuda-oxide目前处于早期alpha阶段，要求开发环境包括Linux操作系统、GPU计算能力8.0或更高版本、CUDA工具包12.x或更新版本，以及pinned nightly Rust工具链 1。相比之下，cutile-rs的要求相对较轻，需要GPU计算能力8.0或更高、CUDA 13.3、稳定版Rust 1.89或更新版本，且已发布至crates.io官方包管理库 1。cutile-rs已在实际项目中得到应用，包括HuggingFace的Grout推理引擎和mistral.rs 1。Melih Elibol计划在2026年9月8-11日举办的蒙特利尔RustConf 2026大会上，发表题为"Fearless Concurrency on the GPU"的演讲 1。

Nvidia has unveiled native GPU programming capabilities in Rust, enabling developers to write GPU kernels directly in the language 1. The announcement, made in September 2026, introduces two projects that compile to PTX code: cuda-oxide, which implements a custom rustc code generation backend for SIMT programming, and cutile-rs, a higher-level tile programming model 1.

The cuda-oxide project is in early alpha stage and requires Linux, GPU compute capability 8.0 or higher, CUDA Toolkit 12.x or newer, and a pinned nightly Rust toolchain 1. Cutile-rs, already released to crates.io, operates at a higher abstraction level with more relaxed requirements: GPU compute capability 8.0 or higher, CUDA 13.3, stable Rust 1.89 or newer, and Linux 1. Both projects represent early-stage efforts, yet cutile-rs has already gained adoption among production systems, including HuggingFace's Grout inference engine and mistral.rs 1.

Melih Elibol is scheduled to present "Fearless Concurrency on the GPU" at RustConf 2026 in Montreal on September 8-11, further highlighting the initiative 1.