The glowing square is the die — it does the thinking. The 8 chips around it are VRAM, where the model lives. On this tier: 32 GB of room. Every model must fit in these chips — or it doesn't run.
PART 2 / 3
The vapor chamber
The copper plate lands exactly on the die. Pipes carry the heat into the fins. Cooling isn't decoration — it's what lets the card hold its speed minute after minute.
PART 3 / 3
The shroud & fans
The face you know. These fans work harder as the VRAM fills — in the calculator, you'll watch them spin up with your card and your model.
Assembled. So — what can it run?
RTX 5090 class · 32 GB · 33 of 51 models at recommended quality.
VRAM is your graphics card's own memory. A model has to fit in it — weights, plus room for the conversation — to run at full speed. Quantization (Q4, Q8…) squeezes a model into fewer bits so it fits in less memory. Q4 is smaller and slightly less sharp; Q8 is near-original quality but roughly twice the size. Tokens per second is how fast the model writes — roughly ¾ of a word per token. 15 tok/s feels like reading pace; 40+ feels instant. Unified memory (Apple silicon, some AMD chips) means the GPU shares your main RAM, so big models can fit without the usual slowdown penalty. Offload means part of the model spills into system RAM. It lets bigger models run, at a real speed cost — the calculator labels those results honestly.
