CUDA for AMD on Windows

This project provides a method to execute Windows applications designed for NVIDIA CUDA on AMD hardware by leveraging ZLUDA in conjunction with ROCm/HIP. It is specifically tailored for compute-heavy workloads, including those utilizing CUDA-enabled LibTorch.

[!IMPORTANT] Hardware Compatibility: Currently, the only fully validated hardware is the AMD Radeon RX 9060 XT (architecture gfx1200). While other AMD GPUs may work, they are considered "candidates" and are not guaranteed. Please submit a GPU compatibility report if you test other cards.

🛠️ Technical Validation

As of the latest update, the public upstream path has been verified using the following stack:

• ZLUDA: v6-preview.69 (Official release)
• AMD HIP SDK: 6.4
• LibTorch: 2.3.0 + cu118
• Hardware: RX 9060 XT / gfx1200

Integration Results

The following components passed the cuda_check utility: nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT.

Furthermore, a real-world PPO network consisting of $\approx 2,216,347$ parameters successfully completed:

1. Forward pass/inference.
2. PPO learning and optimizer functions.
3. A full validation iteration of $65,536$ timesteps.

⚙️ How it Works

The system acts as a translation layer, redirecting CUDA calls to AMD-compatible libraries.

🚀 Installation Guide

Step 1: Prerequisites

• Install the latest AMD GPU Driver.
• Install the AMD HIP SDK for Windows (including HIP Libraries).
  • Note: Versions newer than 6.4 may work but are unverified.
  • Official AMD Windows HIP SDK Guide

Step 2: Setup the Environment

Clone the repository and execute the installation script via PowerShell:

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

The install.ps1 script performs the following:

• Detects the native gfxXXXX target and AMD GPU.
• Validates the HIP SDK and necessary math libraries.
• Fetches the pinned official ZLUDA build.
• Downloads LibTorch 2.3.0+cu118 ($\approx 2.66\text{ GB}$).
• Verifies files via SHA-256 hashes.
• Creates .runtime\runtime-config.json and .runtime\gpu-report.json.
• Runs cuda_check.exe to verify the stack.

Tip: If LibTorch is not required for your application, use: .\scripts\install.ps1 -SkipLibTorch

🏃 Running Applications

To launch a CUDA-targeted program:

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

This launcher places the necessary ZLUDA DLLs alongside the executable and configures the HIP/ROCm runtime paths.

Alternative Options:

• Manual Staging: .\scripts\stage-runtime.ps1 -TargetDir "C:\path\to\your-app"
• Diagnostics:
  • .\scripts\doctor.ps1
  • .\scripts\gpu-scan.ps1 (Logs GPU model, architecture, and driver info; does not collect personal data).
  • .\scripts\test-runtime.ps1

📊 Compatibility & Performance

Hardware Status

Runtime Coverage

Note: Software requiring heavy convolution (cuDNN) may need a nightly HIP stack. However, GEMM-heavy LibTorch training (like the PPO workload) functions without it.

Performance Benchmarks

A controlled A/B test was conducted on 2026-09-13 using the RX 9060 XT (10 iterations, first iteration discarded as warmup).

The public upstream path was approximately $3.03\%$ faster than the custom overlay. Historical tuned runs have reached between $70\text{k}$ and $109\text{k}$ steps/s.

📜 Additional Notes

Custom Overlay: The project previously experimented with a custom cuBLAS/HIP overlay. This is not required for the validated path and is slower for the reference workload. These artifacts are fingerprinted in manifests/recovered-artifacts.sha256 but are not provided as binaries due to incomplete provenance and the inclusion of third-party AMD binaries.

Repository Structure:

• scripts/: Contains install.ps1, run-zluda.ps1, etc.
• manifests/: SHA-256 fingerprints.
• docs/, examples/, benchmarks/: Supporting materials.