THE CRUNCH
A solo developer has automated a workflow to run CUDA workloads on a Radeon RX 9060 XT in Windows, bridging ZLUDA with AMD's HIP SDK. The project, dubbed CUDA-for-AMD-Windows, uses a PowerShell setup to map CUDA libraries like cuBLAS and cuFFT to their AMD equivalents, allowing a 2.2-million-parameter PPO reinforcement learning model to train on the gaming GPU without virtualisation or dual-booting. While cuDNN and a
While cuDNN and other critical libraries remain unsupported, the toolkit offers a promising path for developers to run legacy or niche CUDA-only tools on their daily Windows machines. Benchmarks show the clean official setup achieves a median throughput of 13,278 steps per second, with a custom overlay running roughly 3% slower. The author notes that performance is still a hit compared to native CUDA, and the project relies on ZLUDA, a hobby project that lost commercial backing.
This development highlights that the barrier to running CUDA-exclusive software on AMD hardware is largely a tooling problem rather than an insurmountable hardware flaw. However, the solution remains a tinkerer's tool rather than a production-ready strategy, as it depends on a solo developer and lacks support for key libraries like TensorRT and NCCL.