CUDA binaries now run on AMD GPUs in Windows

You can now run unmodified CUDA software directly on AMD Radeon cards under Windows.

A new open-source project by developer Speedstu combines ZLUDA and AMD's HIP SDK to translate CUDA calls on Windows. The repository provides PowerShell scripts that download required packages, set up the runtime, and launch target applications. The AMD Radeon RX 9060 XT is currently the only fully validated GPU, but built-in diagnostics let users scan other cards.

Why it matters: Most AI tools depend on Nvidia's CUDA ecosystem, making non-Nvidia GPUs hard to use for local builds. This project gives Windows developers a simple path to run CUDA-facing frameworks like LibTorch on AMD hardware without modifying their code.

Here's the gist: clone the repository, run .\scripts\install.ps1, and launch your application using .\scripts\run-zluda.ps1. Core drivers and math libraries like cuBLAS, cuSPARSE, and cuFFT pass integration checks, though cuDNN isn't available in the stable Windows HIP SDK yet.

Nvidia's software monopoly on desktop AI just picked up another leak.