Citizen-built stack runs Windows CUDA apps on AMD GPUs: only the RX 9060 XT is validated

A community GitHub project now runs Windows applications written for NVIDIA's CUDA programming layer on an AMD Radeon RX 9060 XT, the only graphics card it has been validated against. The stack, Speedstu/CUDA-for-AMD-Windows, combines the ZLUDA v6 compatibility shim, AMD's HIP SDK 6.4, and a pinned LibTorch 2.3.0+cu118 build. The author describes the toolchain as "public upstream-only," with no private or recovered DLLs, per the project's own README.

CUDA is NVIDIA's proprietary GPU programming layer. ZLUDA is a third-party shim that translates CUDA calls to AMD's HIP path, and HIP is AMD's own translation layer that targets its Radeon GPUs.

Five CUDA libraries and one workload class pass on the RX 9060 XT. The author reports that nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT clear ZLUDA's cuda_check, and a 2,216,347-parameter PPO reinforcement-learning network completed forward pass, learning, and optimizer steps, running 65,536 timesteps in one clean iteration. The project's own README warns: "This does not mean every CUDA program or AI model works."

ZLUDA v6 lost its commercial sponsor earlier this year and is now community-driven. The project is soliciting compatibility reports from RX 6000, 7000, and 9000 series owners on Hacker News. Broader coverage hinges on those tests.