{"slug": "cuda-for-amd-on-windows", "title": "CUDA for AMD on Windows", "summary": "A reproducible Windows CUDA compatibility stack built on ZLUDA v6-preview.69 and AMD HIP SDK 6.4 has been published on GitHub, validated only on the AMD Radeon RX 9060 XT (gfx1200). The setup passes cuda_check for nvcuda, cuBLAS, cuBLASLt, cuSPARSE and cuFFT, and ran a 2,216,347-parameter PPO network through forward/inference, learning and optimizer work, completing one validation iteration of 65,536 timesteps with LibTorch 2.3.0 + cu118. The project's author cautions that CUDA API and library coverage is workload-dependent and that other AMD GPUs are unverified candidates rather than supported devices.", "body_md": "**WORKING REPRODUCIBLE STACK IS NOW UPLOADED.**\n\nRun CUDA-targeted Windows applications on AMD GPUs through ZLUDA + ROCm/HIP.\n\nA 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.\n\nImportant\n\n**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](https://github.com/Speedstu/CUDA-for-AMD-Windows/issues/new?template=gpu-compatibility.yml), whether it works or fails.\n\nThe public, upstream-only path has been tested without any private/recovered DLLs:\n\n- ZLUDA `v6-preview.69` from the official ZLUDA release\n- AMD HIP SDK `6.4`\n- LibTorch `2.3.0 + cu118`\n- RX 9060 XT / `gfx1200`\n- `nvcuda` , cuBLAS, cuBLASLt, cuSPARSE and cuFFT all pass`cuda_check`\n- a real **2,216,347-parameter PPO network completed forward/inference, PPO learning and optimizer work on the CUDA-facing device**\n- one clean validation iteration completed **65,536 timesteps** using the runtime produced by this repository\n\nThat integration test used the same CUDA-facing LibTorch training workload that originally motivated this project. See [`docs/VALIDATION.md`](/Speedstu/CUDA-for-AMD-Windows/blob/main/docs/VALIDATION.md).\n\nThis does **not** mean every CUDA program or AI model works. CUDA API/library coverage is workload-dependent.\n\n```\nCUDA-targeted Windows application\n              |\n            ZLUDA\n              |\n cuBLAS / cuSPARSE / cuFFT compatibility\n              |\n rocBLAS / hipBLASLt / rocSPARSE / HIP\n              |\n           AMD GPU\n```\n\nInstall a current AMD GPU driver and the **AMD HIP SDK for Windows including HIP Libraries**.\n\nThe validated reference uses HIP SDK 6.4. Newer versions may work but should be treated as unverified until reported.\n\nAMD Windows HIP SDK guide:\n[https://rocm.docs.amd.com/projects/install-on-windows/en/docs-6.4.2/index.html](https://rocm.docs.amd.com/projects/install-on-windows/en/docs-6.4.2/index.html)\n\n```\ngit clone https://github.com/Speedstu/CUDA-for-AMD-Windows.git\ncd CUDA-for-AMD-Windows\npowershell -ExecutionPolicy Bypass -File .\\scripts\\install.ps1\n```\n\n`install.ps1` will:\n\n1. detect the AMD GPU and native `gfxXXXX` target;\n2. verify the AMD driver/HIP SDK and required math libraries;\n3. download the pinned official ZLUDA Windows build;\n4. download LibTorch `2.3.0+cu118` (about 2.66 GB);\n5. verify the downloaded SHA-256 hashes;\n6. generate `.runtime\\runtime-config.json` and`.runtime\\gpu-report.json` ;\n7. run ZLUDA's `cuda_check.exe` against the installed AMD stack.\n\nIf you do not need LibTorch:\n\n```\n.\\scripts\\install.ps1 -SkipLibTorch\n.\\scripts\\run-zluda.ps1 -Program C:\\path\\to\\app.exe\n```\n\nThe launcher stages the required ZLUDA compatibility DLLs beside the target application and sets the HIP/ROCm runtime paths for that run.\n\nYou can also stage without launching:\n\n```\n.\\scripts\\stage-runtime.ps1 -TargetDir C:\\path\\to\\your-app\n.\\scripts\\doctor.ps1\n.\\scripts\\gpu-scan.ps1\n.\\scripts\\test-runtime.ps1\n```\n\nThe GPU scanner records the model, `gfx` architecture, driver and HIP information. It does not intentionally collect usernames, tokens or user files.\n\nExample on the validated machine:\n\n``` php\nAMD Radeon RX 9060 XT -> gfx1200 -> RDNA4 -> validated-reference\n```\n\n| GPU | Target | Project status | \n|---|---|---|\n| Radeon RX 9060 XT | `gfx1200` | ✅ validated reference | \n\nThe scanner recognizes other Windows HIP architecture families and marks them as **unverified candidates** rather than claiming support. Detection is not proof that a workload runs.\n\nAMD's current Windows hardware table:\n[https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html](https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html)\n\nCurrent upstream runtime check:\n\n| CUDA-facing component | Result | \n|---|---|\n| CUDA driver / `nvcuda` | ✅ | \n| cuBLAS | ✅ via rocBLAS | \n| cuBLASLt | ✅ via hipBLASLt | \n| cuSPARSE | ✅ via rocSPARSE | \n| cuFFT | ✅ | \n| cuDNN |  | \n\nThe stable Windows HIP SDK does not ship the full ROCm AI-library stack such as MIOpen, so convolution-heavy software that requires cuDNN can need a newer/nightly HIP stack or additional work. Dense/GEMM-heavy LibTorch training does not necessarily require cuDNN; the validated PPO workload completed without it.\n\nHistorical tuned runs of the ZLUDA + LibTorch path were roughly **70k–109k overall steps/s** on the reference RX 9060 XT. The fresh reproducibility validation was intentionally a short correctness run, not a tuned benchmark. See [`docs/BENCHMARKS.md`](/Speedstu/CUDA-for-AMD-Windows/blob/main/docs/BENCHMARKS.md).\n\nThe original development environment also experimented with a custom cuBLAS/cuBLASLt/HIP overlay. Those recovered binaries are **not required by the current validated public path** and are not distributed by this repository.\n\nThey remain fingerprinted in `manifests/recovered-artifacts.sha256` for archival/research purposes. The normal `install.ps1` path uses upstream ZLUDA plus the installed AMD HIP SDK only.\n\nPlease publish an issue. Failed tests are useful too.\n\n```\n.\\scripts\\gpu-scan.ps1 -OutputPath .\\gpu-report.json\n.\\scripts\\test-runtime.ps1\n```\n\nThen open a [GPU compatibility report](https://github.com/Speedstu/CUDA-for-AMD-Windows/issues/new?template=gpu-compatibility.yml) and include the application, result and first useful error/output.\n\n```\nscripts/              install, diagnostics, scanner, staging and launcher\nmanifests/            pinned versions, hashes and GPU architecture metadata\ndocs/                 validation, architecture, benchmarks and troubleshooting\nexamples/             integration/reference snippets\n.runtime/             generated dependencies and reports; ignored by Git\nlocal-artifacts/      local archival files; ignored by Git\n```\n\n- Only RX 9060 XT / `gfx1200` is currently validated by this project.\n- ZLUDA is not a complete CUDA implementation.\n- Windows exposes only a subset of the full ROCm ecosystem.\n- cuDNN/MIOpen is not available in the validated stable HIP SDK path.\n- NCCL, TensorRT, unsupported PTX behavior and some custom CUDA extensions may fail.\n- `ZLUDA_CC=8.6` is a CUDA-facing compatibility value, not the AMD GPU architecture.\n\nProject-owned scripts and documentation are MIT licensed. ZLUDA, AMD ROCm/HIP, NVIDIA CUDA components and PyTorch/LibTorch retain their own upstream licenses. See [`THIRD_PARTY_NOTICES.md`](/Speedstu/CUDA-for-AMD-Windows/blob/main/THIRD_PARTY_NOTICES.md).", "url": "https://wpnews.pro/news/cuda-for-amd-on-windows", "canonical_source": "https://github.com/Speedstu/CUDA-for-AMD-Windows", "published_at": "2026-09-13 14:25:13+00:00", "updated_at": "2026-09-13 15:45:14.985499+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-tools", "developer-tools", "ai-chips"], "entities": ["ZLUDA", "AMD", "ROCm", "HIP SDK", "LibTorch", "Radeon RX 9060 XT", "GitHub", "CUDA"], "alternates": {"html": "https://wpnews.pro/news/cuda-for-amd-on-windows", "markdown": "https://wpnews.pro/news/cuda-for-amd-on-windows.md", "text": "https://wpnews.pro/news/cuda-for-amd-on-windows.txt", "jsonld": "https://wpnews.pro/news/cuda-for-amd-on-windows.jsonld"}}