“Apple extremely restricted eGPUs on Apple Silicon. Until now.” That’s how Joshua Warren of the Omarchy M team opened his post on X today, and yeah, that got my attention. He plugged an AMD Radeon RX 7600 into a 14-inch M2 Max MacBook Pro running Omarchy and got llama.cpp running Qwen2.5-7B on the card over Vulkan.
What works
- The chain is an RX 7600 in a USB4 enclosure through a dock. amdgpu loads with no ring errors and RADV sees the card.
- His numbers on that one machine: roughly 946 to 968 tokens/sec on prompt processing and 29 to 32 tokens/sec on generation with Qwen2.5-7B Q4_K_M.
- Generation is GPU or driver limited, not CPU. It stayed flat from 1 to 12 threads, and he’s still digging into why.
- The repo also lists an Intel Arc Pro B60 measured on an M1 Max. An RTX 5060 Ti build compiles but hasn’t been run yet.
The catch
This isn’t “install a package” territory. It took a one-line device-tree fix (the USB4 tunnel BAR window was mapped to the wrong address, so every VRAM read came back as all ones) plus kernel patches, including a fix for a SIGBUS in the Vulkan driver. None of it is upstream. The pull request is open and Joshua says there’s no rush to merge it. The card has to be attached at boot, hot-plug and display output are untested, and a stock kernel won’t do this yet.
My take
Real hardware hacking, the kind I grew up on. If you’ve got an Apple Silicon Mac and a spare AMD card, the repo has a read-only doctor script that checks each layer without installing anything or touching your boot entries. Start there, and don’t go patching kernels on a machine you need Monday morning.
Have a blessed day! And Happy Hacking!
