Clusters, explained honestly

A cluster is multiple machines wired together to act as one bigger computer.

✅ A real datacenter cluster

GPUs are joined by very fast links — NVLink / NVSwitch inside a box, InfiniBand between boxes.
Memory effectively pools: many GPUs behave like one huge GPU, so a single model far bigger than one card can run across them efficiently.
This is how a 671B model runs — spread over many server GPUs that talk to each other at hundreds of gigabytes per second.

⚠️ A pile of desktop cards

Cards talk over ordinary PCIe slots and consumer networking — far slower than NVLink or InfiniBand.
Memory does not pool cleanly. You can't just add up the GB on four gaming cards and expect one giant model to run across them smoothly.
Great for experiments and smaller models; not a substitute for real server GPUs when you need to run one enormous model well.

So what should you take away?

When a build on this site combines several machines, we always show you the combined memory, combined power, and combined price — and we flag it as a cluster. If you're running a truly large model for real users, spend on server GPUs with proper interconnects. We won't pretend a stack of desktop cards is the same thing, because it isn't.