One could stick Mac mini inside of a supermicro jbod chassis and hot glue cables to enclosure ![]()
Definitely. Plenty.
The driving force here, however, is pure power savings, for users in high cost of living areas. I pay $0.52/kWh here, and if I could save 40 watts (~200/year) off of my Supermicro server by swapping an Intel C262-based MLB+CPU+RAM with a Mac mini – I would jump into taking risk for any possible shenanigans head first.
This is correct.
Sleep/wake is a universally hard problem to solve and I"d just disable sleep, ASPM, and anything else that increases risk of glitches all to save two cents in power.
There would be additional latency --thunderbolt will add 2-5 microseconds each way vs direct PCIE connection, so double that for roundrip.
Internal crossbars behaved just like regular PCIE switches, so no extra risk here, PCI switches are a solved problem.
Any retimers in the cable itself will add under 1 microsecond, and I’d argue a short cable without retimiers shall be used.
On the software side – thunderbolt controllers often limit maximum payload size to under 512B, so some performance loss is expected, but irrelevant with rotational drives. Thunderbolt also does not allow peer to peer DMA – but that is not used unless you use NVME disks.
The only issue that I can think of that may become an issue with old HBAs is hot plug: thunderbolt initialization emulates PCIE hot plug. Old HBA have no idea how to deal with this. So, use new ones, which have modern pcie frontend.
Having written this, I very much feel this is a horrible idea. Let’s scrap it, and wait for an actual low power ARM hardware.
This appears to be a dead end: parallels does not seem to allow pass through of a random pcie device: KB Parallels: Firewire/Thunderbolt device is not working in Parallels Desktop
