RTX 4090
NVIDIA's flagship desktop graphics card. Built for gamers and creators, but strong enough to run small open AI models on a single home PC.
The price of one card, before tax — the cost of a high-end desktop part, not a business machine.
GPU memory (VRAM) is like the model's workspace — the desk it has to lay everything out on while it works. 24GB is a decent-sized desk: fine for small open models (roughly up to ~18B parameters), but nowhere near big enough for the 100B+ models this store advises on.
450 watts is what the card pulls when working flat-out — similar to a hair dryer on high. That's about 38% of one average home's continuous draw.
RTX 6000 Ada
A professional workstation card used in studios, labs, and small offices. Same GPU family as the 4090, doubled memory, built for being left on 24/7.
Over 4x the desktop card, mostly for double the memory and a professional-grade design rated for continuous use.
48GB workspace — double the RTX 4090's desk space. Still far short of the 280GB+ that this store's big open models need from a single card, which is why datacenter cards exist.
Less power than the desktop card despite more memory, because it runs at efficiency-tuned clocks instead of gaming clocks. That's about 25% of one average home's continuous draw.
A100 80GB
The GPU that powered the first wave of large AI models. Lives in a server, not a desktop, and is designed to be wired to other A100s over NVLink so several cards share one memory pool.
Server-grade hardware sold to businesses, not individuals — support, validation, and NVLink wiring account for most of the jump over a desktop card.
80GB workspace per card — over 3x a desktop card's desk space. Datacenter cards like this are normally bought 4 or 8 at a time and wired together so their workspaces merge into one big desk instead of staying four separate small ones.
One card alone draws as much as a full-size microwave running continuously. That's about 33% of one average home's continuous draw.
H100 80GB
The current-generation datacenter GPU behind most of today's largest AI models. Faster than the A100 at the same memory size — and hungrier for power.
Roughly double the A100 for a large jump in speed. This is the GPU most frontier AI labs buy today.
80GB workspace — the same desk size as the A100. Speed keeps improving generation to generation, but the workspace-per-card ceiling barely moves, which is exactly why huge models need many cards wired together, not just one faster card.
One H100 alone draws more continuous power than half an average home. That's about 58% of one average home's continuous draw.
DGX H100
A complete, pre-built machine holding 8 H100 GPUs wired together with NVIDIA's NVLink so they share memory and act as one giant GPU. This is what a real cluster looks like in a single box.
The price of the whole machine — 8 GPUs plus the internal NVLink wiring, cooling, and support that let them behave as one unit instead of eight separate cards.
640GB combined workspace (8 x 80GB desks pushed together and welded into one desk by NVLink). A model too big for one card can spread across all 8 as if they were a single card — the core idea behind a real cluster.
As much continuous power as 8-9 average homes combined, running every hour of every day. That's about 8.5 average homes running non-stop.
GB200 NVL72
A full datacenter rack: 72 of NVIDIA's newest Blackwell GPUs, wired into one enormous shared-memory machine. This is the scale companies buy to train and run the biggest models in the world.
The price of a small building's worth of custom silicon, liquid cooling, and networking — not something you click "buy" on, but what frontier AI labs order by the rack.
13,500GB (13.5TB) — one shared workspace across all 72 GPUs, enough to run every model this store advises on at once, with enormous room to spare.
About as much continuous power as 100 average homes — a rack this size needs its own dedicated power feed. That's about 100 average homes running non-stop.