GitHub project documents reproducible CUDA compatibility setup for AMD GPUs on Windows
Speedstu’s project combines ZLUDA with AMD ROCm/HIP and has been validated on the Radeon RX 9060 XT, while broader hardware and application support remains unverified.
A GitHub project from Speedstu documents a reproducible Windows setup designed to run some CUDA-targeted applications on AMD GPUs through ZLUDA and AMD ROCm/HIP. The project is aimed at CUDA-facing compute workloads, including applications using CUDA-enabled LibTorch.
The documented validation currently covers only AMD’s Radeon RX 9060 XT, based on the gfx1200 architecture. Other AMD GPUs are listed as candidates rather than confirmed supported devices, and the project asks users to report both successful and failed tests.
The setup includes scripts that stage ZLUDA compatibility DLLs beside the target application and configure HIP and ROCm runtime paths. Its reference configuration uses AMD’s HIP SDK 6.4; newer versions are described as potentially compatible but remain unverified.
In a controlled 10-iteration comparison on the RX 9060 XT, using a PPO workload and excluding the first iteration as warm-up, the public upstream path recorded median performance of 13,278 overall steps per second. A recovered custom overlay reached 12,876, making it about 3.03 per cent slower in that test.
The project says the validated PPO workload completed without cuDNN. It cautions, however, that CUDA API and library coverage depends on the workload, while convolution-heavy applications may require newer or additional HIP/ROCm components. The claims are based on the project’s own documentation and do not establish universal compatibility or independent verification.
