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.

The public, upstream-only path has been tested without any private/recovered DLLs:

• ZLUDA v6-preview.69 from the official ZLUDA release
• AMD HIP SDK 6.4
• LibTorch 2.3.0 + cu118
• RX 9060 XT / gfx1200
• nvcuda , cuBLAS, cuBLASLt, cuSPARSE and cuFFT all passcuda_check
• a real 2,216,347-parameter PPO network completed forward/inference, PPO learning and optimizer work on the CUDA-facing device
• one clean validation iteration completed 65,536 timesteps using the runtime produced by this repository

That integration test used the same CUDA-facing LibTorch training workload that originally motivated this project. See docs/VALIDATION.md.

This does not mean every CUDA program or AI model works. CUDA API/library coverage is workload-dependent.

CUDA-targeted Windows application
              |
            ZLUDA
              |
 cuBLAS / cuSPARSE / cuFFT compatibility
              |
 rocBLAS / hipBLASLt / rocSPARSE / HIP
              |
           AMD GPU

Install a current AMD GPU driver and the AMD HIP SDK for Windows including HIP Libraries.

The validated reference uses HIP SDK 6.4. Newer versions may work but should be treated as unverified until reported.

AMD Windows HIP SDK guide: https://rocm.docs.amd.com/projects/install-on-windows/en/docs-6.4.2/index.html

git clone https://github.com/Speedstu/CUDA-for-AMD-Windows.git
cd CUDA-for-AMD-Windows
powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1

install.ps1 will:

1. detect the AMD GPU and native gfxXXXX target;
2. verify the AMD driver/HIP SDK and required math libraries;
3. download the pinned official ZLUDA Windows build;
4. download LibTorch 2.3.0+cu118 (about 2.66 GB);
5. verify the downloaded SHA-256 hashes;
6. generate .runtime\runtime-config.json and.runtime\gpu-report.json ;
7. run ZLUDA's cuda_check.exe against the installed AMD stack.

If you do not need LibTorch:

.\scripts\install.ps1 -SkipLibTorch
.\scripts\run-zluda.ps1 -Program C:\path\to\app.exe

The launcher stages the required ZLUDA compatibility DLLs beside the target application and sets the HIP/ROCm runtime paths for that run.

You can also stage without launching:

.\scripts\stage-runtime.ps1 -TargetDir C:\path\to\your-app
.\scripts\doctor.ps1
.\scripts\gpu-scan.ps1
.\scripts\test-runtime.ps1

The GPU scanner records the model, gfx architecture, driver and HIP information. It does not intentionally collect usernames, tokens or user files.

Example on the validated machine:

AMD Radeon RX 9060 XT -> gfx1200 -> RDNA4 -> validated-reference

The 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.

AMD's current Windows hardware table: https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html

Current upstream runtime check:

The 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.

Historical 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.

The 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.

They 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.

Please publish an issue. Failed tests are useful too.

.\scripts\gpu-scan.ps1 -OutputPath .\gpu-report.json
.\scripts\test-runtime.ps1

Then open a GPU compatibility report and include the application, result and first useful error/output.

scripts/              install, diagnostics, scanner, staging and launcher
manifests/            pinned versions, hashes and GPU architecture metadata
docs/                 validation, architecture, benchmarks and troubleshooting
examples/             integration/reference snippets
.runtime/             generated dependencies and reports; ignored by Git
local-artifacts/      local archival files; ignored by Git

• Only RX 9060 XT / gfx1200 is currently validated by this project.
• ZLUDA is not a complete CUDA implementation.
• Windows exposes only a subset of the full ROCm ecosystem.
• cuDNN/MIOpen is not available in the validated stable HIP SDK path.
• NCCL, TensorRT, unsupported PTX behavior and some custom CUDA extensions may fail.
• ZLUDA_CC=8.6 is a CUDA-facing compatibility value, not the AMD GPU architecture.

Project-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.