GoKawiilZLUDA project enables CUDA workloads to run on AMD GPUs via Windows
Developers behind the ZLUDA project have released a reproducible stack combining ZLUDA v6-preview.69 with AMD's HIP SDK 6.4, allowing CUDA-targeted Windows applications, including LibTorch-based machine learning workloads, to run on AMD hardware. The setup has been validated specifically on the Radeon RX 9060 XT (gfx1200), where core CUDA libraries like cuBLAS and cuFFT passed checks and a PPO reinforcement learning network completed training over 65,536 timesteps.
GoKawiil's interpretation of the reporting above, not reported fact.
CUDA's dominance has long locked developers into Nvidia hardware for GPU-accelerated computing and AI workloads. A working, reproducible translation layer that lets AMD cards handle CUDA-facing software could lower costs and expand hardware options for machine learning practitioners, though the current validation covers only one GPU model and success remains workload-dependent.
- The stack pairs ZLUDA with AMD's HIP/ROCm libraries to translate CUDA calls for AMD GPUs on Windows.
- Validation so far is limited to the AMD Radeon RX 9060 XT; other cards are untested candidates.
- A real PPO neural network training workload using CUDA-facing LibTorch completed successfully as proof of concept.
AMD Radeon RX 9060 XT Graphics Card — This is the exact GPU validated in the ZLUDA + ROCm/HIP stack, so it's the surest way to try running CUDA-targeted Windows apps on AMD hardware. If you want to reproduce the LibTorch and PPO training tests described, the RX 9060 XT (gfx1200) is the card the project confirms working.
See AMD Radeon RX 9060 XT Graphics Card on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-09-13
Published there as: “CUDA for AMD on Windows”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.