Hardware Wiki

Everything this store sells, explained in plain words: the animated core ideas, how cooling works, and — at the bottom — a dictionary of all the jargon. Every number on this page comes from the same real products and formulas used everywhere else on this site.

GPU memory — the shelf space for a model

An AI model is basically a huge list of numbers, called parameters. To run the model, every one of those numbers has to be loaded into the GPU's own memory — a fast workspace built into the card itself. If the model doesn't fit in that workspace, it doesn't run slower. It simply doesn't run.

required GB = parameters in billions × 1 GB, plus 20% working room

Example — GLM-5.2 has 753 billion parameters: 753 × 1 GB × 1.2 = 903.6 GB needed.

GLM-5.2 on one NVIDIA DGX B200

The model needs 903.6 GB
This server has 1,440 GB

✓ Fits — with about 536 GB to spare.

GLM-5.2 on one NVIDIA GeForce RTX 5090

The model needs 903.6 GB
This card has 32 GB

✗ Doesn't fit — the model is about 28× bigger than the card's memory.

Bars in each panel are drawn to the same scale, so you can see the real size difference. A brilliant desktop card is still the wrong tool for a 753-billion-parameter model — see our model pages for what does fit.

Watts — what it costs to keep it running

A watt measures how fast something uses electricity. Every product page on this site shows a watts number. Here's how to feel what that number means, using one NVIDIA DGX B200 server, which draws 14,300 watts while working.

0.0 homes

An average home draws about 1,200 watts around the clock — this one server draws as much as that many homes, all day, every day.

0.00 car batteries

A typical electric-car battery holds about 90 kWh of energy. Running this server for one full day uses that many batteries' worth.

Clusters — many machines, one job

A cluster is just several separate computers wired together and made to work as a team on one job.

A real cluster: 3 NVIDIA DGX B200 servers joined by NVLink and InfiniBand — links so fast their memory genuinely pools into one workspace of 4,320 GB.

A pile of desktop cards: each card is stuck with its own 32 GB. Ordinary connections (PCIe slots, regular networking) are far too slow to share one model between cards.

The difference is the wiring. Datacenter machines are connected with NVLink and InfiniBand — links so fast that the machines can genuinely act like one giant GPU. Desktop cards only have ordinary connections, hundreds of times slower, so a pile of them can't efficiently hold one big model together — the cards would spend their time waiting instead of working.

CPU, RAM, GPU, GPU memory — who does what

Cooling & maintenance — keeping it alive

Every watt in becomes heat out. That's physics, not a design flaw. A 575 W desktop card is a small space heater. A 14,300 W server is a wall of them. A 120,000 W rack makes as much heat as 100 homes' worth of constant power use — all of it in one cabinet.

Chips protect themselves: too hot, and they slow down (called thermal throttling — you paid for speed you're not getting); hotter still, and they shut off. Good cooling isn't a luxury, it's how you get the performance you bought. Watch the two ways it's done:

Air cooling

A fan pushes cool air (blue) through metal fins bolted to the hot chip; the air carries the heat away (orange). This is your desktop card, and — scaled up to a wall of screaming fans — our air-cooled DGX servers.

Liquid cooling

Liquid flows across a cold plate strapped to the chip, picks up the heat (red), and dumps it at a radiator (turning blue) before coming back around. A car engine survives the same way. Rack-scale systems like the GB200 NVL72 are liquid-cooled because air physically can't move that much heat.

Every product page lists a rough "ideal cooling" for that hardware, and we stock matching cooling gear on our Cooling page.

Best practices — the honest checklist

The dictionary — jargon, translated

Every term you'll meet on a spec sheet, each with an analogy to something you already know. Words like these are quietly linked to this dictionary wherever they appear across the store.

Parameter

One tiny learned setting inside a model — like one of billions of little screws that each got turned exactly right during training. The parameter count (7 billion, 753 billion…) is the model's size, and it's what decides how much GPU memory you need.

Token

A small chunk of text — roughly a short word, or a piece of a longer one. Models read and write one token at a time, like someone reading syllable by syllable, except millions of times faster.

GPU memory / VRAM

The workspace built into the graphics card where the whole model must sit while it runs. Like a workbench: if the project doesn't fit on the bench, you can't work on it at all. See it animated ↑

HBM

"High Bandwidth Memory" — the extra-fast type of GPU memory on datacenter chips. Like having your tools strapped to your arm instead of in a toolbox across the room: same tools, dramatically less time reaching for them.

Precision (FP16 / FP8 / FP4)

How many digits each parameter is stored with. Rounding to fewer digits — like writing prices in whole dollars instead of cents — makes the model take less memory and run faster, at a small cost in exactness.

InfiniBand

Ultra-fast networking between whole servers in a cluster — a private highway between buildings. Ordinary office networking is a city street with traffic lights by comparison.

PCIe

The standard slot a graphics card plugs into inside a PC — the wall outlet of computer parts. Perfectly fine for one card; far too slow to make a pile of cards behave like one big GPU.

TDP (watts)

The most power a part is designed to draw. Every watt that goes in comes back out as heat, so this number is both your electricity bill (see it animated ↑) and your cooling problem ↑.

Node

One computer in a cluster. Like one musician in an orchestra — useful on its own, but the point is what they play together.

Rack

The standardized metal cabinet that servers bolt into, like a bookshelf for computers. Our largest product, the GB200 NVL72, is one entire rack engineered to act as a single machine.

Training vs. inference

Training is teaching the model — like writing the cookbook, done once at enormous cost. Inference is using it — cooking from the book. The hardware in this store is sized for inference: running models, not creating them.