Why This Matters
A community project has published a reproducible Windows setup that lets CUDA-targeted applications run on AMD GPUs via ZLUDA plus AMD's HIP/ROCm stack, validated end-to-end on a real PPO training workload with LibTorch. If such compatibility layers mature, they chip away at CUDA's lock-in, which is the core of Nvidia's competitive moat. For now the scope is narrow, but it shows upstream-only tooling can work without proprietary shims.
Key Takeaways
- ZLUDA v6-preview.69 + AMD HIP SDK 6.4 + LibTorch 2.3.0/cu118 passed checks for nvcuda, cuBLAS, cuBLASLt, cuSPARSE and cuFFT.
- Validation ran a 2.2M-parameter PPO network through inference, learning and optimizer steps over 65,536 timesteps.
- Only the Radeon RX 9060 XT (gfx1200) is verified; other AMD cards and other CUDA workloads are unproven, with coverage being workload-dependent.
CUDA for AMD on Windows
WORKING REPRODUCIBLE STACK IS NOW UPLOADED.
Run CUDA-targeted Windows applications on AMD GPUs through ZLUDA + ROCm/HIP.
A reproducible Windows CUDA compatibility setup built around ZLUDA + AMD HIP/ROCm. It is intended for CUDA-facing compute applications, including workloads that use CUDA-enabled LibTorch.
Important Validated hardware is currently AMD Radeon RX 9060 XT ( gfx1200 ) only. Other AMD GPUs are candidates, not guaranteed working devices. If you test another card, please open a GPU compatibility report, whether it works or fails.
Verified today
The public, upstream-only path has been tested without any private/recovered DLLs:
ZLUDA v6-preview.69 from the official ZLUDA release
from the official ZLUDA release AMD HIP SDK 6.4
LibTorch 2.3.0 + cu118
... continue reading
GoKawiil