A solo developer has created a PowerShell-based toolkit that bridges the ZLUDA translation layer with AMD's native HIP/ROCm SDK for Windows. This project allows users to run CUDA-exclusive libraries, including cuBLAS, cuSPARSE, and cuFFT, directly on AMD hardware without virtualization or dual-booting.
The developer successfully trained a 2.2-million-parameter reinforcement-learning network on a Radeon RX 9060 XT using unmodified CUDA libraries. This GPU is currently the only model officially supported by the toolkit. The setup uses ZLUDA version v6-preview.69 and AMD's HIP SDK 6.4.
Benchmark tests showed the official ZLUDA and HIP SDK stack achieved a median throughput of 13,278 steps per second. A version using salvaged legacy binaries reached 12,876 steps per second, which is approximately 3% slower. The author noted that performance remains dependent on the specific software stack.
The toolkit is an open-source project and not an enterprise-grade solution. It currently lacks support for critical machine learning libraries such as cuDNN, TensorRT, and NCCL. ZLUDA is maintained as a hobby project following the loss of its commercial backing, and the developer advises against using this pipeline for production workloads.
