Running CUDA on AMD GPUs Demonstrated on RX 9060 XT

CUDA-for-AMD-Windows, published by Speedstu, is a configuration package for running CUDA-specific workloads on AMD GPUs in Windows. In reporting by Zak Killian of Tom’s Hardware, it is emphasized that this effort uses a method without virtualization or dual-booting. Through updates to ROCm, AMD has officially brought PyTorch and the HIP development kit to Windows for consumer GPUs. The targets are the Radeon RX 7000 series and RX 9000 series. Within the supported framework, the AMD environment on Windows has reached a practical level. However, certain proprietary applications, old repositories, and specialized AI tools still contain CUDA-dependent implementations. This configuration package is positioned as a bridge to fill that gap.

running rigidly CUDA-exclusive workloads on AMD hardware in Windows is possible without virtualization or dual-booting

The above is the key point presented in the original article. This article is based on Tom’s Hardware (All Rights Reserved). It organizes the facts in reliance on fair quotation under Article 32 of the Japanese Copyright Act.

The CUDA Wall That Remains Despite ROCm

Improvements

AMD’s Windows support has established a foundation for running training and inference through official channels. The ability to run PyTorch on consumer GPUs is significant. The HIP development kit is also now available on Windows. Still, assets that accept nothing but CUDA remain in the field. Applications dependent on a specific company’s products have not been ported. Old repositories are no longer maintained, with no prospect of revision. Experimental AI tools are written on the assumption that NVIDIA GPUs are present. Such assets will not launch simply because the user’s hardware is from AMD. A divide remains that expanded official support alone cannot resolve. This configuration package is a practical answer to that divide.

How ZLUDA and HIP Are Connected

This configuration package is not a new runtime environment. At its core is automation that links ZLUDA with AMD’s HIP and ROCm development kits. ZLUDA is known as a translation layer that replaces CUDA calls with other methods. It has a history of having once received funding from AMD. A series of PowerShell procedures detects the user’s GPU architecture. It fetches a pinned version, ZLUDA v6-preview.69. It maps them to the ROCm math libraries present on Windows. In theory, users do not need to handle complex build work. Its distinctive feature is that it is organized as reproducible steps. It is designed not for manual replacement, but for automatic connection of pathways.

Training Demonstrated on RX 9060 XT

The demonstration used a Radeon RX 9060 XT. At present, that model is the only officially supported GPU. The developer intercepted CUDA driver APIs and mapped them to the AMD side. The targets were cuBLAS, cuSPARSE, and cuFFT. All are CUDA libraries heavily used in numerical computing. As a demonstration, a 2.2-million-parameter PPO reinforcement learning network was trained. Importantly, unmodified CUDA libraries were used. The workload ran end-to-end on the AMD GPU. It demonstrated that specific training workloads can work via the translation layer. It is a small but meaningful bridge.

Supported Libraries and Current Limitations

The supported scope is limited. In addition to the driver API, connections for cuBLAS, cuSPARSE, and cuFFT have been confirmed. These cover matrix operations, sparse matrix processing, and fast Fourier transforms. The reinforcement learning demonstration succeeded within this scope. Meanwhile, cuDNN does not work yet. In deep learning, dependence on cuDNN is strong for convolution and normalization. Large-scale image recognition and generative workloads therefore remain out of scope. Support being limited to the RX 9060 XT is also a constraint. The entire RX 7000 and RX 9000 lines are not covered. Checking the supported libraries and hardware is essential for practical use.

Impact on Developers and Windows Environments

The benchmark test included in the documentation also draws attention. It compared the 2.2-million-parameter reinforcement learning workload on the RX 9060 XT. The public upstream path is a combination of official ZLUDA and HIP SDK 6.4. It recorded a median throughput of 13,278 steps per second. The figure remains a measurement result under specific conditions. Still, it shows that translation-layer overhead is within a practical range. The impact is significant for engineers who use Windows as their daily development machine. Issues associated with processing through WSL2 can be avoided. CUDA-dependent assets can be tested without the hassle of dual-booting. A path has opened to proceed to verification without waiting for porting. It could function as a way to avoid dependency hell.

Editorial Opinion

Regarding short-term impact. In the next 3 to 6 months, we expect to see increased attempts to try CUDA-dependent experimental repositories in AMD environments on Windows. We assess that the translation layer will be used for verification before formal porting to HIP. Due to lack of cuDNN support, production adoption will be limited. We expect trials on development machines and bug reports to accumulate.

Regarding the long-term view. In 1 to 3 years, we expect AMD support for CUDA-dependent assets to split into two layers. We assess that official porting and translation-layer use will coexist. We expect the base for training verification on consumer GPUs to broaden. The lock-in by a specific company seems likely to ease. We assess that sustainability will depend on establishing a maintenance system.

Questions from the editorial team. We need to ask whether use of the translation layer will replace porting efforts. We see it as an open question who will verify performance and accuracy. We need to ask whether support for more GPUs will expand. We also assess whether usage terms and rights issues will be clarified as a point of debate. We expect sharing of actual measurements to inform future decisions.

References

• “ Solo dev enables running CUDA on AMD hardware in Windows, getting multiple CUDA libraries running on a gaming Radeon RX 9060 XT GPU in Windows — CUDA-exclusive workloads on AMD hardware in Windows possible without virtualization or dual-booting “, by Zak Killian — Tom’s Hardware, 2026-09-14T14:23:12.000Z (ARR)
• Source URL: https://www.tomshardware.com/pc-components/gpu-drivers/solo-developer-wires-zluda-to-amds-hip-getting-multiple-cuda-libraries-running-on-a-radeon-rx-9060-xt-in-windows-cuda-exclusive-workloads-on-amd-hardware-in-windows-is-possible-without-virtualization-or-dual-booting