Nami Tech Hardware

Understanding clusters

What is a cluster?

A cluster is several GPUs wired together so tightly that they act as one giant GPU. Think of it as bolting several desks together into one massive workspace.

Real cluster vs. a pile of desktop cards

Both can add up to a big "GB" number on paper. Only one of them can actually run a single huge model.

At a glance Real cluster — DGX H100 Pile of desktop cards — 6x RTX 6000 Ada
Total memory 640GB 288GB
Interconnect speed ~900GB/s per GPU (NVLink) Standard Ethernet/Wi-Fi — roughly 1,000x slower
Can run a single large model? Yes No
Support One NVIDIA support contract 6 individual card warranties
Approximate price ~$200,000 ~$40,800
Real NVLink-connected datacenter GPU system

A real cluster — e.g. DGX H100

  • 8 H100 GPUs sit inside one chassis, wired by NVLink at up to ~900GB/s per GPU.
  • Their 80GB memories act as one 640GB pool — a single model can be split cleanly across all 8.
  • Built, tested, and sold by NVIDIA as one product, with one support contract.
Several separate desktop-class GPU cards

A pile of desktop cards — e.g. 6x RTX 6000 Ada

  • Each card lives in its own regular computer, connected only by ordinary ethernet or Wi-Fi — thousands of times slower than NVLink.
  • Each card's 48GB of memory stays its own separate island; there's no hardware pooling them into one 288GB pool.
  • A single huge model can't realistically split itself across them the way it can across a real cluster — at best you run several small, separate jobs.
Real cluster — NVLink ≈900GB/s per GPU — wired together as one machine H100 H100 H100 H100 H100 H100 H100 H100 640GB unified pool → one shared pool
Pile of desktop cards Ethernet / Wi-Fi only (slow) — no shared memory between cards RTX 6000 Ada 48GB RTX 6000 Ada 48GB RTX 6000 Ada 48GB RTX 6000 Ada 48GB RTX 6000 Ada 48GB RTX 6000 Ada 48GB → six separate islands

The honest takeaway: 6 RTX 6000 Ada cards add up to 288GB on a spec sheet, which sounds like it beats a single H100. But that number is not honestly comparable to a real cluster's shared memory — the desktop pile can't act as one machine, so it can't run one model that's bigger than any single card in the pile.

Why memory pooling actually matters

A big model has to load as one piece in memory — you can't hand it half a brain. Qwen3 235B, one of the models this store advises on, needs about 282GB to load. Our pile of 6 RTX 6000 Ada cards adds up to 288GB on a spec sheet — technically more than enough. But that memory sits in 6 separate 48GB islands, and no single piece of it is bigger than 48GB. It's like having six half-full gas tanks instead of one full one: the total fuel is there, but no car can run on six disconnected tanks at once. A real cluster's NVLink wiring is what turns 8 separate 80GB tanks into one true 640GB tank — that's the only way the "GB" number on the spec sheet becomes a number you can actually use.

Combining real clusters

When a build combines multiple systems, we always show the combined memory, power, and price together.

Example: 2 × DGX H100 wired together
1 × DGX H100 (640GB, 10,200W)$200,000
1 × DGX H100 (640GB, 10,200W)$200,000
Combined memory
1,280GB
Double one system — enough for DeepSeek-V3 671B (805GB minimum).
Combined power
20,400W
about 17 average homes running non-stop.
Combined price
$400,000
Two complete systems, no extra hardware needed.

At the very top end, a single GB200 NVL72 rack is 72 GPUs wired into one machine: 13,500GB shared memory, 120,000W, ~$6.5M. That's about 100 average homes running non-stop — for one rack.