New Reproducible Stack Enables CUDA Apps on AMD GPUs
AI-generated image · Bay Street Wire
A compatibility layer using ZLUDA and ROCm allows CUDA-targeted Windows applications to run on AMD hardware.
A reproducible Windows CUDA compatibility setup has been released, allowing users to run CUDA-targeted applications on AMD GPUs via ZLUDA and ROCm/HIP, as first reported on Hacker News.
The stack is specifically intended for CUDA-facing compute applications, including those utilizing CUDA-enabled LibTorch. According to the project documentation, the only currently validated hardware is the AMD Radeon RX 9060 XT (gfx1200). While other AMD GPUs are considered candidates, the developer notes they are not guaranteed to work.
Technical validation performed on September 13, 2026, confirmed that nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT all passed the "cuda_check" test. The project successfully ran a 2,216,347-parameter PPO network, completing forward inference, PPO learning, and optimizer tasks. The validated reference setup utilizes ZLUDA v6-preview.69, AMD HIP SDK 6.4, and LibTorch 2.3.0 + cu118.
Despite these gains, the project notes significant limitations. ZLUDA is not a complete CUDA implementation, and API coverage varies by workload. Furthermore, cuDNN is not available in the validated stable Windows HIP SDK because the full ROCm AI-library stack, including MIOpen, is not included. This means convolution-heavy software requiring cuDNN may require additional work or a nightly HIP stack, though the validated PPO workload was able to complete without it.