Open-Source Project Enables Unmodified NVIDIA CUDA Libraries to Run on AMD GPUs in Windows
Developer Speedstu released CUDA-for-AMD-Windows, an open-source automated setup combining ZLUDA v6-preview.69 and AMD HIP/ROCm to run unmodified NVIDIA CUDA libraries on AMD GPUs under Windows. The project demonstrated end-to-end 2.2 million-parameter PPO reinforcement learning network training on an AMD Radeon RX 9060 XT graphics card without modifying CUDA libraries. NVIDIA's proprietary CUDA software stack forms a powerful ecosystem moat, locking developers into its hardware for AI and GPU-accelerated workloads. Enabling seamless CUDA application compatibility on AMD GPUs in Windows broadens hardware choices for developers and reduces reliance on NVIDIA infrastructure. The automated PowerShell configuration intercepts CUDA Driver APIs and maps core libraries like cuBLAS, cuSPARSE, and cuFFT to AMD's HIP SDK 6.4 equivalents. Benchmarks show the official upstream configuration achieved a median training throughput of 13,278.46 steps per second (SPS), slightly outperforming legacy ZLUDA custom override paths.
## BACKGROUND
CUDA is NVIDIA's proprietary parallel computing platform and runtime environment that heavily dominates artificial intelligence and high-performance computing development. ZLUDA is an open-source drop-in translation layer that intercepts CUDA API calls and translates them to run on non-NVIDIA GPUs through AMD's ROCm and HIP frameworks.