New Open-Source Setup Enables Running CUDA Windows Applications on AMD GPUs Using ZLUDA and ROCm/HIP

A new open-source project provides a reproducible software environment to run CUDA-targeted Windows applications on AMD GPUs. The solution combines ZLUDA, an open-source CUDA compatibility layer, with AMD's ROCm/HIP platform, enabling CUDA compute workloads to execute on AMD hardware under Windows. This approach is particularly relevant for CUDA-facing applications such as those using LibTorch for AI model training and inference.

Currently, the setup has been validated only on the AMD Radeon RX 9060 XT GPU (gfx1200 architecture). Other AMD GPUs may work but are unverified, and users are encouraged to report compatibility results. The stack uses publicly available components including ZLUDA v6-preview.69, AMD HIP SDK 6.4, and LibTorch 2.3.0 with CUDA 11.8 support.

The project demonstrates successful execution of CUDA libraries such as cuBLAS, cuSPARSE, and cuFFT through their ROCm/HIP equivalents, passing internal compatibility checks. A complex AI workload involving a 2.2 million-parameter PPO network was run end-to-end, including forward inference and optimizer steps, validating the practical usability of the stack.

Installation involves running provided PowerShell scripts that detect the AMD GPU, verify driver and SDK versions, download required components, and configure the runtime environment. The launcher stages necessary ZLUDA DLLs alongside the target application and sets environment variables for HIP/ROCm.

While the setup does not cover the entire CUDA API surface and some CUDA extensions or libraries like cuDNN and TensorRT are unsupported in the stable HIP SDK, it offers a functional path for many CUDA compute workloads on AMD GPUs under Windows. Performance tests show the public upstream stack performs comparably to previous custom overlays.

This development matters because it provides a pathway for Windows users with AMD GPUs to run CUDA-dependent applications without switching hardware or operating systems. It also fosters broader hardware compatibility for CUDA software, which has traditionally been limited to NVIDIA GPUs.

The project is open source under the MIT license, with detailed documentation, validation reports, and scripts available on GitHub. Users are encouraged to contribute compatibility reports and test results to help expand supported hardware and workloads.