NVlabs/cutile-rs

cuTile Rust provides a safe, tile-based kernel programming DSL for the Rust programming language. It features a safe host-side API for passing tensors to asynchronously executed kernel functions.

What it solves

Writing GPU kernels is traditionally error-prone, often leading to memory safety issues and data races. cuTile Rust provides a system for writing these kernels in Rust, extending Rust's strict ownership and borrowing rules to the GPU to ensure that kernels are memory-safe and free of data races by design.

How it works

It uses a tile-based approach where mutable outputs are split into disjoint pieces (tiles), while read-only inputs are shared. The system employs a #[cutile::module] macro to transform Rust functions into GPU kernels and generates a host-side launcher. These kernels are then JIT-compiled through CUDA Tile IR, allowing them to be executed synchronously, via async/await, or as CUDA graph replays.

Who it’s for

Developers who want to write high-performance GPU kernels using Rust's safety guarantees instead of traditional CUDA C++, as well as researchers building ML inference engines (like Grout or Candle) that require custom, optimized GPU operations.

Highlights

  • Memory Safety: Extends Rust's ownership rules to the device to prevent data races.
  • JIT Compilation: Compiles kernels on the fly using CUDA Tile IR.
  • Flexible Execution: Supports synchronous, asynchronous, and CUDA graph replay execution modes.
  • High Performance: Capable of reaching near-cuBLAS performance for GEMM operations and high throughput for LLM inference.

Related

  • Dispatch
    cuTile Rust 0.2.0 Enables Safe, Data-Race-Free GPU Kernels in Rust

    cuTile Rust 0.2.0 provides a tile‑based DSL that extends Rust's ownership model to GPU kernels, delivering memory‑safe, data‑race‑free execution with performance comparable to low‑level CUDA libraries.

  • Dispatch
    NVIDIA Announces CUDA Rust: Native GPU Kernels in Rust via SIMT and Tile Tracks

    In September 2026 NVIDIA released CUDA Rust, enabling native Rust GPU kernels through two tracks—SIMT (cuda-oxide) and Tile (cutile‑rs)—which bring compile‑time safety and Rust‑first ergonomics to CUDA programming.

  • Project
    tracel-ai/cubecl

    A Rust-based JIT compiler and language extension for writing high-performance compute kernels that run across CUDA, HIP, Metal, WebGPU, and CPU from a single codebase.

  • Project
    NVIDIA/cutile-python

    A Python programming language for NVIDIA GPUs that allows developers to write and launch parallel kernels using Tile IR without needing low-level CUDA C++.

  • Dispatch
    cuda-oxide: Bringing Rust's Safety and Expressiveness to CUDA

    NVIDIA introduces cuda-oxide, an experimental Rust-to-CUDA compiler that allows developers to write GPU kernels in idiomatic Rust without relying on DSLs or foreign language bindings.