CUDA for AMD on Windows

A Windows-based application built on ZLUDA + AMD HIP/ROCm enables CUDA-targeted Windows applications to run on AMD GPUs. This setup is designed for compute-intensive tasks, including those using CUDA-enabled LibTorch. Currently, AMD Radeon RX 9060 XT (gfx1200) is the only validated hardware, but other AMD GPUs may also work. The project tested this setup without any private/recovered DLLs, using the same CUDA-facing LibTorch workload that inspired the project.

The documentation outlines the validation process and the importance of installing the latest AMD GPU driver and AMD HIP SDK for Windows, including HIP Libraries. The HIP SDK version 6.4 is confirmed to work, while newer versions require verification. The launcher stages the necessary ZLUDA compatibility DLLs and adjusts HIP/ROCm runtime paths for the application.

The GPU scanner logs the model, GFX architecture, driver version, and HIP information, but does not collect personal data. While this setup supports AMD GPUs, it may not work with all CUDA programs or AI models, as CUDA API/library coverage is workload-dependent. The AMD Windows HIP SDK guide is available at the provided link. A benchmark test on AMD Radeon RX 9060 XT PPO workload showed that the public upstream path achieved 13,278 median overall SPS, compared to 12,876 for the custom overlay, indicating the upstream method is currently the faster option.

The original development environment also explored using a custom cuBLAS/cuBLASLt/HIP overlay, but it is not required for the validated public path. The project's scripts and documentation are licensed under the MIT license, and third-party components retain their individual licenses.