Developer's script lets Windows AMD GPUs execute CUDA-only software via ZLUDA bridge
A solo developer has released an automated PowerShell toolkit that pairs the ZLUDA translation layer with AMD's native HIP/ROCm SDK on Windows, allowing CUDA-specific libraries to run on Radeon hardware. The setup successfully mapped cuBLAS, cuSPARSE, and cuFFT to AMD equivalents, and even trained a small reinforcement-learning model on an RX 9060 XT without virtualization. cuDNN support remains missing, and only the RX 9060 XT is officially supported at this time.
The toolkit automates GPU architecture detection and pins ZLUDA to a specific preview build, pairing it with AMD's official Windows HIP SDK. Benchmark results show the official ZLUDA path achieving 13,278 steps per second versus 12,876 for a salvaged legacy overlay, a roughly 3% gap. AMD's recent ROCm expansion to Windows consumer GPUs laid groundwork for this project, though ZLUDA itself lost its AMD funding with the v6 release.
The developer's proof-of-concept trained a 2.2-million-parameter reinforcement learning network end-to-end on an RX 9060 XT without virtualization. Missing cuDNN support remains a notable limitation, and the RX 9060 XT is the sole officially supported GPU, restricting broader applicability for now.
This project could meaningfully lower barriers for developers and researchers who rely on CUDA-only tools but use AMD hardware, potentially broadening access to AI experimentation on consumer GPUs. It may also pressure NVIDIA's ecosystem lock-in, though missing cuDNN support and single-GPU validation limit immediate impact. Individual developers and small teams could benefit most, while larger organizations may wait for broader hardware support and stability before adopting this approach.