Why you don't look at an used external PCIe chassis with a PCIe expander, adapt it to use external PSUs. There are thread about this, people did this successfully in the past for rendering rigs.
Btw, for your intended usage you probably would spend less with a dedicated PC with a mining motherboard. Seems you don't want to run macOS, no?
A dedicated PC with a mining motherboard is probably smarter, and maybe I will save up for one. I really like Mac Pros, though, and I do better work when I have an aesthetic or emotional connection with my tools. Also, 12-cores of xeon and 96 gb of ram is extremely helpful for multicore but not GPU-accelerated data analysis on large datasets.
Is an external PCIe chassis with a PCIe expander different from an eGPU because it connects directly to a PCIe slot rather than going through a Thunderbolt card? I haven't been able to find an affordable one except for this dubious number for ~ € 400 on ebay sold as-is, untested:
I do have a flashed Thunderbolt 3 card on the way and am looking at eGPU though, despite the performance hit.
Why external GPUs instead of pixla's mod? I don't like how hot the GPUs make the entire interior of the Mac Pro during machine learning, and when I'm not doing a training run, there's no reason for them to be on at all, but I don't want to have to connect and disconnect them internally every time I switch between macOS and linux. I may try editing the NVIDIA GPU BIOS to underclock, limit max power, and use a more aggressive fan policy, and that might turn out to have less performance cost than an eGPU.
I am planning to do pixla's mod for my internal RX 580, though, which has better macOS support and is the GPU I use for actual graphics and monitor plug. My hopefully sufficient relevant experience consists of soldering the power socket back to the logic board on my PowerBook G3 in 1999, replacing broken iphone components, and crimping ethernet connectors. I hope that's sufficient. Until I get around to doing that, I'm being careful not to push the RX 580, and it has stayed quiet and cool, hopefully with no high power draws.
Just a bit of brainstorming now: what about connecting one or two CUDA GPUs internally to the PCIe slots, but connecting an external power supply to their power sockets, one with its own power switch, and then only switching it on before and off after booting into and shutting down linux? I (probably) have an extra Mac Pro 4,1 power supply I could use. I could in principle stick that in the optical drive area and run the power cord through an unused PCI bracket hole.
The addresses the power problem, but not the internal heat problem. Could I address the heat problem by just setting a much more aggressive fan policy on the Mac Pro? E.g., normal/quiet during normal usage, but fans at max >35 C. Alternatively, maybe I shouldn't worry about the ambient interior temperature of the Mac Pro because the affected components have their own fan control and overheating shutdown protection.
I unfortunately don't know enough about electrical engineering to know if these ideas are completely stupid and/or dangerous to me or to the hardware, so feedback is appreciated
