What is a cluster?

A cluster is a group of separate machines wired together with a fast enough network that they can share one job and act like a single, bigger computer — including pooling their memory.

The honest difference

Real datacenter clustering

H100 SXM / DGX H100 / GB200 NVL72

These connect GPUs with NVLink and NVSwitch — dedicated links running at 900 GB/s or faster. A model split across many GPUs can pass data between them almost as fast as if it were talking to itself. The combined memory is usable, not just added up on paper.

A pile of desktop cards

Many RTX 4090s over PCIe/Ethernet

Consumer cards don't have NVLink between each other. They talk over PCIe or a regular network — 10 to 50 times slower. You technically get more total memory, but the GPUs spend most of their time waiting for data instead of computing, so a large model becomes painfully slow rather than merely expensive.

What "combined" means on this site

Whenever a recommended build uses more than one machine, we add up:

Combined GPU memory — the total the model can actually use once wired together.
Combined power — every machine's watts, added up, because your power bill and cooling see all of them.
Combined price — the total cost to own every unit in the build.

See this in action on any model's build page.