A ZLUDA + ROCm kit runs CUDA apps on AMD Radeon, on Windows
Validated Sept 13 on one Radeon (RX 9060 XT): cuBLAS, cuFFT and LibTorch training run unmodified — but no cuDNN, NCCL or TensorRT.
Copy markdown
What runs today, unchanged
The kit pins ZLUDA v6-preview.69 over AMD's HIP SDK 6.4 and translates CUDA calls to ROCm at runtime: cuBLAS→rocBLAS, cuBLASLt→hipBLASLt, cuSPARSE→rocSPARSE, plus cuFFT. Your existing CUDA or LibTorch binary runs as-is — point a PowerShell script at the .exe, no recompile.
The catch: no cuDNN, NCCL or TensorRT
The stable Windows HIP path ships no MIOpen backend for cuDNN, so convolution-heavy training and most PyTorch vision pipelines won't run; multi-GPU (NCCL) and TensorRT are out too. Only the Radeon RX 9060 XT (gfx1200) is validated — other cards are "candidates, not guaranteed."
Build this weekend
If your workload leans on cuBLAS/cuFFT — RL loops, custom LibTorch training, GEMM-heavy inference — you can train on a Radeon without touching your CUDA code or buying an NVIDIA card. The dev proved it with a 2.2M-param PPO network running 65,536 timesteps on the AMD device.