I picked up a fairly cheap 1U Supermicro server off eBay and turned it into a dedicated Chia Gigahorse GPU farming node. I’d been piggybacking on my existing KVM hypervisor with a Tesla P4 passed through to a VM, with 100+ drives mounted on the host, merged with mergerfs, and shared into the guest over NFSv4. It worked, but I hit the wall at around 1050TB: lookup times started creeping past the ~22 second threshold, stale shares were running about 4–5%, and I even had to take one of my JBOD shelves offline. This build is a Supermicro CSE-815TQ with an X11SSW-F board, E3-1230 v5, 32GB ECC UDIMM, an RTX A4000, an LSI 9200-8e for external SAS, and an Intel X550 10GbE card. The “catch” is PCIe lanes and power in a 1U chassis, so I used the riser’s lane split to run the A4000 at Gen3 x8 (fine for farming) and fed GPU power using an adapter cable off the spare CPU 4-pin connectors. Once racked and connected to the disk shelves, lookup times dropped into the 2–6 second range even on heavy filter passes, power averaged about 250W during lookups, and the whole thing is now simpler, expandable, and easy to isolate on its own VLAN.