One Workload Shape, Two Wrong Ways to Serve It
Render farms dispatch work frame by frame, batch inference splits traffic request by request, cloud desktops assign one machine per person — the common shape of these businesses is a large number of mutually independent small tasks, each needing one CPU and at most one GPU card. Serving them with an 8-10 GPU server wastes the expensive multi-GPU density and heavy airflow entirely; scattering dozens of tower PCs loses control of power, cooling and asset management. The PC-Farm series answers in between: 4U4H packs four pluggable PC nodes into a 4U chassis, 6U6H packs six. Nodes run fully independently while the chassis handles power delivery and rack mounting — bringing “a pile of small machines” into the management regime of a standard machine room.
Hot Swap Plus Fully Independent Power Control: Maintenance at Node Granularity
PC-Farm nodes support hot swap and fully independent power on/off, and together these set the granularity of every maintenance action. When a node's OS crashes or its hardware fails, the response is to power down that node alone, pull it, and slide in a spare — every other node in the chassis stays online throughout. Planned work such as OS reinstalls or driver upgrades likewise needs no chassis-wide downtime window. Compared with virtualization-based consolidation, nodes share no host kernel, so a software fault on one node cannot reach its neighbors — the failure domain's boundary is the node's physical boundary. This is especially friendly to unattended hours: the on-call action is simply to power off and isolate the problem node from the cluster, and the repair can wait for daylight.
The 750W Backplane and Four CRPS 1100W Supplies: Power in Two Layers
Power delivery reads in two layers. At the node layer, the backplane provides up to 750W peak per node — a budget that bounds the CPU, GPU card and peripherals inside each node, and means fully loaded nodes never contend for power: the budget is fixed per node, not pooled across the chassis and left to chance. At the chassis layer, four redundant CRPS 1100W supplies feed the whole system; if one fails, the others carry the load and the failed unit is replaced online. CRPS is the common server redundant-supply standard, so spares interchange with other equipment in the room. Centralized supply replaces one PSU per PC, cutting failure points — and power redundancy itself is something scattered desktop machines simply never have.
One GPU Card per Node: Workloads That Scale by Count
Each 6U6H node supports one GPU card, and the matching workload profile scales horizontally by node count: a render farm splits frames or shots across nodes rendering independently, with zero inter-node communication; an inference service deploys one model replica per node and routes requests across them, so a single node failure costs only one replica's throughput; a cloud graphics terminal serves a group of users from one node. There is only one deciding test: do tasks need GPU interconnect between them? Training and large-model inference that need multi-GPU parallelism or a single task's large memory footprint belong to the 8-10 GPU platforms — PC-Farm scales the node-count dimension, not per-machine compute.
Between 4U4H and 6U6H
The two models share the same node hot swap, independent power control, backplane budget and CRPS redundancy; the difference is density and form — four nodes in 4U versus six in 6U. Where rack units are scarce and node count comes first, 6U6H is the pick, and its one-GPU-per-node support also brings it closer to inference and render-farm duty; 4U4H starts at a smaller chassis granularity, suiting rack-by-rack expansion in step with business batches. In the end, PC-Farm answers a plain question: when a business consists of dozens of unrelated small tasks, the more reasonable thing to put in the rack is dozens of small machines that can each be handled alone — not one big machine that must never stop.