CUDA for AMD on Windows Targets RX 9060 XT With ZLUDA and ROCm
For AMD GPU owners, the usable range of CUDA-only Windows software still depends on ZLUDA's incomplete coverage, a narrow verified GPU list, and the absence of cuDNN/MIOpen, NCCL, and TensorRT on the stable path.
Reporting from 1 source: GIGAZINE.
A project called CUDA for AMD on Windows pairs ZLUDA, a CUDA compatibility layer, with AMD's ROCm/HIP stack to run CUDA-targeted Windows applications on AMD GPUs. At the time of writing it lists the Radeon RX 9060 XT (gfx1200) as its verified target, with the RX 9070 XT externally verified, the Radeon 890M partially supported by the community, and other AMD GPUs unconfirmed. Installation needs an AMD GPU driver, a matching HIP SDK for Windows, and an install.ps1 script run from the repository.
The verification target is narrow. The project recognizes only the Radeon RX 9060 XT (gfx1200) as a fully verified reference, with integration work complete there. The Radeon RX 9070 XT (gfx1201) is listed as externally verified and used as a reference, while the Radeon 890M (gfx1150) has partial community support, with reports of conv2d hangs and abnormal SDPA output. Other AMD GPUs remain unconfirmed and depend on architecture.
Setup requires the latest AMD GPU driver and a suitable AMD HIP SDK for Windows including HIP libraries, with the version depending on the GPU architecture. After that, users clone the repository and run the install.ps1 script. The verified software set includes ZLUDA v6-preview.69 (official release), AMD HIP SDK 6.4, and LibTorch 2.3.0 + cu118, with nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT passing cuda_check.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.