CUDA Rust offers two ways to write GPU kernels

NVIDIA is developing two paths for writing GPU kernels in Rust. cuda-oxide gives programmers explicit control over threads and memory, compiling kernel functions to PTX. cutile-rs works with tiles of data and lets the compiler map them onto GPU threads. Both connect Rust ownership to GPU access patterns, but their toolchains and maturity differ.

NVIDIA · Sep 8, 2026

Why it matters

A concrete bridge between Rust systems work and GPU inference infrastructure.

More context

Start by comparing the vector-add examples. The SIMT path uses a custom rustc backend and pinned nightly toolchain; the tile path uses stable Rust 1.89 or newer with CUDA 13.3. Both require Linux and supported NVIDIA hardware. Try a small isolated kernel before considering integration.

Keep in mind

cuda-oxide is early alpha. These are evolving NVIDIA-specific tools; 1min has not compiled or benchmarked the examples.

Original read