'CUDA for AMD on Windows' allows CUDA-compatible Windows applications to run on AMD GPUs.
As the name suggests, ' CUDA for AMD on Windows ' is a project aimed at running CUDA-compatible Windows applications on AMD GPUs.
Speedstu/CUDA-for-AMD-Windows: Run CUDA-targeted Windows applications on AMD GPUs with ZLUDA + ROCm/HIP.
https://github.com/Speedstu/CUDA-for-AMD-Windows
CUDA for AMD on Windows aims to create a reproducible Windows CUDA compatible environment for running CUDA applications by combining ZLUDA , which acts as a compatibility layer for running CUDA applications on AMD GPUs in a Windows environment, with AMD's ROCm/HIP stack.
[code]
CUDA-targeted Windows application
|
ZLUDA
|
cuBLAS / cuSPARSE / cuFFT compatibility
|
rocBLAS / hipBLASLt / rocSPARSE / HIP
|
AMD GPU
[code]
At the time of writing, the system tested was the 'AMD Radeon RX 9060 XT,' and other AMD GPUs were considered as potential compatible options. The software used for testing is as follows:
• ZLUDA v6-preview.69 (Official Release)
- AMD HIP SDK 6.4
LibTorch 2.3.0 + cu118
・nvcuda・cuBLAS・cuBLASLt・cuSPARSE・cuFFT (passed cuda_check)
As of the time of writing this article, the operational verification status and compatibility status are as follows:
| GPU | Target ID | situation | Note |
|---|---|---|---|
| Radeon RX 9060 XT | gfx1200 | Verified | Project integration work completed. |
| Radeon RX 9070 XT | gfx1201 | Externally verified | reference |
| Radeon 890M | gfx1150 | Partial support from the community | HIP 7.2/GEMM is working. There have been reports of conv2d hanging and abnormal output from SDPA. |
| Other AMD GPUs | (Architecture-dependent) | Unconfirmed |
To install CUDA for AMD on Windows, you need to follow these steps:
- AMD prerequisite installation: Install the latest AMD GPU drivers and the appropriate AMD HIP SDK for Windows, including the HIP library (the correct version depends on the GPU architecture).
- Run the installer: Clone the repository from the official GitHub and run the 'install.ps1' script.
Please note that the following limitations apply as of the time of writing this article.
- Only the RX 9060 XT/GFX1200 is recognized as a fully verified project reference.
ZLUDA is not a complete CUDA implementation.
- Even if the result of cuda_check is 'pass,' the correctness is not guaranteed.
Only a small fraction of the ROCm ecosystem's features are available on Windows.
- cuDNN/MIOpen is not available in the verified stable HIP SDK path.
NCCL, TensorRT, unsupported PTX, and some custom CUDA extensions may not work.
The CUDA for AMD on Windows project is an important initiative that expands the range of applications for AMD GPUs and strengthens collaboration with the CUDA ecosystem, so it will be worth keeping an eye on its future developments.
Related Posts:
