Bektursun's AI Hardware Store

What's a cluster, really?

Any time a recommended build uses more than one GPU or machine, we call it a cluster. Here's what that word actually means — and doesn't.

Plain-English definition: a cluster is a group of computers (or GPUs) wired together and coordinated by software so they can work on one job as if they were a single, larger machine.

The part that actually makes a cluster work

The word doing all the work in that definition is wired together. Real datacenter GPU clusters use dedicated high-speed interconnects — NVIDIA's NVLink between GPUs inside a machine (up to 130 TB/s across a full NVL72 rack), NVSwitch fabric tying 8 GPUs in a DGX system into one coherent pool, and InfiniBand networking linking whole racks together. That hardware is what lets one model's memory and computation span multiple GPUs without the GPUs constantly waiting on each other over a slow link. The software stack (schedulers, parallelism libraries) is built assuming that fast interconnect exists.

What a cluster is not: putting several unrelated desktop PCs next to each other, or several add-in cards sitting in the same tower with nothing but standard PCIe between them, is not equivalent to a datacenter GPU cluster. On paper you can add up their memory and get a big number — and for simply fitting a model's weights somewhere, that number is real. But without high-speed interconnect and coordinated infrastructure, those GPUs are far slower at working together, and some workloads won't run across them at all. We show you both numbers on model pages precisely so you can see this tradeoff instead of us hiding it.

How our builds add up

When a build in this store spans multiple units, we always show the same three combined numbers:

  • Combined GPU memory — the sum of every unit's memory, which is what determines whether the model fits.
  • Combined power — the sum of every unit's power draw, in watts, with the everyday comparison alongside it.
  • Combined price — the sum of every unit's price, so you see the real total cost of the build.

We do not make speed or performance claims from these totals — memory, power, and price add up arithmetically; real-world throughput depends on the interconnect and software, which is exactly the point above.

Back to Open-Source Models