TeardownHow it works · Stop 1 of 9

What is it made of?

A datacenter GPU is not one chip but a package: a processor die, towers of memory, a silicon baseplate wiring them together, and a power-and-cooling system built to remove a kilowatt of heat. Follow a single multiply through it, then open any layer.

After this page you can answer
  1. What are the parts of an AI accelerator, and what does each one do?
  2. Why does the memory sit on the package instead of on the motherboard?
  3. Where does a kilowatt of heat go?
Data last revised November 2026
Follow the data

Follow one multiply through the package.

A model’s answer is billions of multiply-adds. Each one makes the same six-stage trip, and each stage is a part you can see. Step through it and the model lights up the part doing the work.

ASSEMBLEDEXPLODED
drag to rotate · pull the slider to explode
Stage 1 of 6 · HBM Memory Stacks

The weights wait in memory

Before anything is multiplied, the model’s numbers (its weights) sit in the HBM stacks beside the processor: 80 GB of them on an H100. No arithmetic happens here.

How HBM is built →

Every layer, one at a time.

Open any part for what it does, why it matters, and where it’s headed.

Figure 1.1An accelerator package taken apart layer by layer. Pick a part to light it up and follow a number from the HBM stacks, across the interposer, into the compute die and back out as heat.
Try this

Step through all six stages and count how many do arithmetic. One does. The other five move bytes or heat.

Then this

Pull the slider to EXPLODED and rotate the package. The memory stacks sit beside the die, not on top of it: every weight has to cross the baseplate between them.

Why it matters

Stage 2 is the slow one. The cores can multiply far faster than the interposer can deliver weights, which is why the next stop is about memory.

Recap

The short version

  1. An accelerator is a package: a compute die, HBM stacks and an interposer wiring them together.
  2. The weights live in HBM, not on the die. Every multiply starts with a trip across the interposer.
  3. Tensor cores do the arithmetic; most of the rest of the package exists to feed and cool them.
Check yourself

Why are the HBM stacks bonded a few millimetres from the die, instead of sitting in sockets on the board like ordinary RAM?

Pick one. The reasoning shows either way.