CUDA for AMD on Windows Runs a Real Workload on Radeon RX 9060 XT

A community project called CUDA for AMD on Windows has assembled a reproducible path for running CUDA-targeted Windows applications on AMD graphics hardware. Rather than reimplementing the whole software stack, it combines the open-source ZLUDA CUDA compatibility layer with AMD’s HIP SDK and ROCm libraries, then wraps the setup in scripts that detect the GPU, fetch pinned components and test the resulting runtime.

The project currently validates one reference GPU: the Radeon RX 9060 XT, identified as gfx1200. On that machine, its published checks report working CUDA-facing nvcuda, cuBLAS, cuBLASLt, cuSPARSE and cuFFT paths. More importantly than a library probe, the developer reports completing inference, PPO learning and optimizer work with a 2,216,347-parameter LibTorch network, including a 65,536-timestep validation iteration.

A CUDA-targeted application still presents CUDA calls at the top of the stack. ZLUDA provides the compatibility layer, while CUDA-facing math and sparse operations can ultimately be served by AMD-side libraries such as rocBLAS, hipBLASLt and rocSPARSE through HIP. The practical appeal is that a compatible prebuilt Windows application can potentially run without being rewritten around a different compute API.

That is materially different from saying AMD has implemented NVIDIA CUDA. CUDA remains NVIDIA’s platform, ZLUDA is an independent compatibility project, and AMD supplies the underlying HIP/ROCm pieces used on the Radeon side. The CUDA for AMD on Windows repository mainly makes those moving parts easier to reproduce and validate together.

AMD’s own documentation helps explain why the compatibility ceiling matters. The current Windows HIP SDK exposes the HIP runtime plus math and primitive libraries, but AMD still lists communication libraries and AI libraries such as MIOpen and MIGraphX as unavailable in the Windows HIP SDK. That leaves important gaps for CUDA software that expects a broader NVIDIA library stack.

The project consequently warns that cuDNN-dependent software, NCCL, TensorRT, unsupported PTX behavior and some custom CUDA extensions may fail. A dense or GEMM-heavy workload that maps cleanly onto the available libraries is a much easier target than assuming an arbitrary CUDA AI application will behave like it does on a GeForce card.

AMD officially supports the RX 9060 XT in its Windows HIP SDK hardware matrix, so the underlying Radeon compute target is not the experimental part. The experiment is the compatibility bridge that lets software built for CUDA reach that AMD stack. ZLUDA itself describes its software as alpha quality, which is an important boundary for anyone considering production use.

For Windows Radeon owners, the useful development is therefore narrower but still meaningful: a documented, repeatable setup has moved beyond a hello-world demonstration and completed a real CUDA-facing training workload. Broader GPU reports, more library coverage and application-by-application validation will determine whether that becomes a practical compatibility option rather than an interesting developer stack.