RTX 5060 runs a model on Linux over USB4 on one M1 Max
Joshua Warren said an RTX 5060 on an M1 Max MacBook Pro, Linux, over USB4, runs “Qwen3.8-27B” fully on the card at 25 tps on CUDA after a 10-line fix for 16K memory pages.
NVIDIA's open driver had no case for the Mac's 16K memory pages, so setup failed before that change. The rate is his measurement on that machine, on CUDA.
Warren said he sent the fix to NVIDIA.