CUDA Compatibility for AMD GPUs on Windows
Run CUDA applications on AMD hardware using ZLUDA and ROCm/HIP.
1Introduction to CUDA Compatibility
CUDA, traditionally exclusive to NVIDIA GPUs, can now run on AMD hardware through ZLUDA and ROCm/HIP. This setup allows Windows users to execute CUDA-targeted applications on AMD GPUs, expanding hardware options for developers and researchers. The project focuses on reproducibility and compatibility, ensuring that CUDA-enabled workloads can function seamlessly on AMD GPUs.
2Validated Hardware and Prerequisites
The AMD Radeon RX 9060 XT (gfx1200) is the only validated GPU for this project. Users must install the latest AMD GPU driver and the AMD HIP SDK for Windows, including HIP Libraries. While newer versions of the HIP SDK may work, they remain unverified until tested. This ensures that the setup is stable and reliable for CUDA applications.
3Installation and Setup Process
To install, users need to clone the CUDA-for-AMD-Windows repository and run the installation script. The script stages the required ZLUDA compatibility DLLs and sets the HIP/ROCm runtime paths. This process ensures that CUDA-targeted applications can run without modifications, providing a seamless experience for users transitioning from NVIDIA to AMD hardware.
4Performance Benchmarks and Testing
A controlled A/B test ran 10 iterations per runtime on the same RX 9060 XT PPO workload. The public upstream path achieved a median overall SPS of 13,278, outperforming the custom overlay by 3.03%. Historical tuned runs reached up to 109k steps/s, demonstrating the potential for high-performance CUDA applications on AMD GPUs.
5Limitations and Known Issues
Only the RX 9060 XT is currently validated, and ZLUDA is not a complete CUDA implementation. Windows exposes only a subset of the full ROCm ecosystem, and cuDNN/MIOpen is unavailable in the stable HIP SDK path. NCCL, TensorRT, and some custom CUDA extensions may fail, limiting the scope of compatible applications.
6GPU Compatibility Reporting
Users can run GPU scanner scripts to diagnose their machines and generate compatibility reports. These scripts record GPU model, architecture, driver, and HIP information without collecting personal data. Failed tests are useful for improving compatibility and expanding the range of supported GPUs in future updates.
7Project Structure and Documentation
The project includes scripts for installation, diagnostics, scanner, staging, and launcher functions. Documentation covers validation, architecture, benchmarks, and troubleshooting. Examples provide integration and reference snippets, while manifests store pinned versions, hashes, and GPU architecture metadata.
8Conclusion and Future Directions
This project opens new possibilities for running CUDA applications on AMD GPUs, offering a viable alternative to NVIDIA hardware. While current limitations exist, ongoing development and user feedback will drive improvements. Future updates aim to expand GPU compatibility and enhance performance, making CUDA applications more accessible on AMD platforms.
Original article: Hacker News — CUDA Compatibility for AMD GPUs on Windows