If you walked in not knowing what GPU memory, a cluster, or a watt meant — this page is for you. No jargon, no assumptions.
Think of it as the model's workspace — the desk it has to spread everything out on while it's running. Before a model can answer a single question, the whole thing has to be loaded onto that desk. If the desk (memory) is too small, the model doesn't run slowly — it simply doesn't fit, the same way a 6-foot table can't hold an 8-foot blueprint no matter how you fold it. That's why every product and model page on this site leads with the memory number: it's the one spec that decides whether a purchase can even work, before price or speed matter at all.
A cluster is a group of GPUs wired together with a very fast, dedicated connection (NVIDIA's NVLink inside one machine, or InfiniBand between machines) so their separate desks get pushed together and bolted into one much bigger desk. A single card might have an 80GB desk; 8 of them clustered together in a DGX H100 have one 640GB desk. The key word is real: the wiring has to be fast enough that the GPUs can share work and memory as if they were one chip. See our full "Understanding clusters" page for why a pile of ordinary desktop cards in separate computers is NOT a cluster, even if their memory adds up to a big number on paper.
Two reasons: your electric bill, and your building. A GPU pulling 700 watts, running 24 hours a day, costs real money every single month — and datacenter systems pull thousands of watts continuously, not just at peak. The bigger issue for serious buyers is infrastructure: a rack drawing 120,000 watts needs its own dedicated power feed and industrial cooling, the same way you can't run a car-charging station off a household outlet. We translate every wattage number into average-homes-worth of continuous power (1 home ≈ 1,200W) and EV-battery-charges-per-day (1 EV battery ≈ 90kWh) so the number means something you can actually picture, not just a spec-sheet digit.
The B stands for billion — it's the model's parameter count, roughly "how much the model has learned and has to remember." Llama 3.1 405B has 405 billion parameters. The industry rule of thumb is simple: each billion parameters needs about 1GB of GPU memory to hold, plus about 20% extra room for the live conversation. So parameter count isn't just a bragging-rights number — run it through that formula and it tells you, in gigabytes, the smallest desk you're legally allowed to buy.
Start from the model, not the hardware. Work out its minimum GB (see above), then look for the cheapest real hardware that clears that number: a desktop card (24-48GB) for hobby-sized models, a handful of datacenter GPUs wired together (up to a few hundred GB) for most serious open models, and a packaged system or rack (640GB and up) only once a single model genuinely needs that much shared memory. Buying a rack you don't need is exactly as wrong as buying a card that's too small — this store's advisor pages exist so you never have to guess.
Four things drive the price up as you move up the lineup: (1) memory — more GB per card costs more to manufacture; (2) reliability — datacenter parts are validated for years of 24/7 use, desktop parts aren't; (3) interconnect — the NVLink/NVSwitch wiring that lets multiple GPUs act as one machine is expensive engineering, which is most of why a DGX H100 costs more than 8 separate H100s would; (4) scale — a rack bundles the power delivery, liquid cooling, and networking for 72 GPUs into one order. None of it is arbitrary markup — each jump in price buys a specific, checkable capability, which is exactly why we show memory, power, and price side by side on every product.