ZLUDA Enables CUDA-Targeted Windows Applications on AMD GPUs

CUDA Compatibility on AMD Hardware

A new reproducible setup has been released that allows CUDA-targeted Windows applications to run on AMD GPUs. This is achieved through the use of ZLUDA, a compatibility layer, in conjunction with AMD's ROCm/HIP platform. The setup is specifically designed for CUDA-facing compute applications, including those that utilize CUDA-enabled LibTorch.

Validated Hardware and Software Stack

The current validated hardware for this setup is the AMD Radeon RX 9060 XT (gfx1200). While other AMD GPUs may work, they are not guaranteed. The public, upstream-only path has been tested using ZLUDA v6-preview.69, AMD HIP SDK 6.4, and LibTorch 2.3.0 + cu118. Key CUDA libraries such as cuBLAS, cuBLASLt, cuSPARSE, and cuFFT have all passed validation checks.

Integration and Performance

An integration test involved a 2,216,347-parameter PPO network, which successfully completed forward/inference, PPO learning, and optimizer work on the CUDA-facing device. A clean validation iteration also completed 65,536 timesteps using the runtime produced by this repository. This demonstrates the functional capability of the compatibility layer for specific workloads, though full CUDA API/library coverage is workload-dependent.

Installation Process

Installation involves cloning the repository and running a PowerShell script. The script detects the AMD GPU, verifies the AMD driver and HIP SDK, downloads ZLUDA and LibTorch, and then runs ZLUDA's cuda_check.exe against the installed AMD stack. Users can opt to skip LibTorch download if not needed. A separate script is provided to launch applications using the ZLUDA compatibility DLLs.