› The AI Hardware Atlas
AI isn't a cloud. It's a place.
Sand, copper, water and electricity, from a switch smaller than a virus to buildings that draw as much power as a city. Spin every piece 360°, zoom in, X-ray it, blow it apart, and turn on Signal to watch data, power and heat move.
Every exhibit ends with questions to talk over together. Printable question sheet (PDF)
Real: What it looks like.
- Electric power
- Data
- Heat
- Cool liquid
~15 cm
The Accelerator Module
Die + memory + power + cold plate = one GPU module.
Open the full exhibit →Chip
Where the math physically happens.
The Transistor
A switch with no moving parts. AI is billions of them flipping.
Per big AI chip: tens to hundreds of billions
Explore in 3D →
The GPU Die
Thousands of small cores doing the same math at the same time.
Area: up to ~800 mm² per die
Explore in 3D →
HBM: Stacked Memory
Memory chips stacked like pancakes and drilled through with copper.
Bandwidth per stack: roughly 1+ TB/s
Explore in 3D →
The Accelerator Module
GPU, memory, power delivery and a cold plate on one board.
Power per module: ~700–1,200+ W
Explore in 3D →
The CPU
The general-purpose brain that runs everything around the GPUs.
Cores per server CPU: dozens to ~100+
Explore in 3D →
Systolic Array (TPU-style)
A grid where numbers pulse through like a heartbeat.
Grid size: typically ~128×128 cells or more
Explore in 3D →
Board
Memory, storage, network and power around the chip.
DRAM: System Memory
The server's big, fast scratchpad. Forgets everything when the power goes off.
Per stick: ~32–256 GB
Explore in 3D →
NVMe SSD
Flash storage: no moving parts, keeps data when the power's off.
Speed: several GB/s
Explore in 3D →
Hard Disk Drive
Spinning mirrors and a flying needle: the cheapest way to keep a lot of data.
Capacity: ~20–30+ TB per drive
Explore in 3D →
Network Card + Optics
Turns electricity into flashes of light and back again.
Speed per port: ~400–800 Gb/s (2026)
Explore in 3D →
Network Switch
The traffic cop: dozens of ports, one giant routing chip.
Ports: ~32–64 at 400–800 Gb/s
Explore in 3D →
Power Supply
Turns wall power into the steady DC the chips want.
Efficiency: ~94–96%+ (Titanium grade)
Explore in 3D →
Liquid Cooling Loop
Water carries heat away far better than air can.
Water vs. air: ~3,500× more heat per volume
Explore in 3D →
Server + rack
Eight accelerators in a box, boxes in a cabinet.
Facility
Buildings full of racks, wired to a power plant.
Data
Where the training data lives and gets cleaned.
How it all works
Training, answering, and the software in between.
A Training Run
Thousands of GPUs learning one model together for weeks.
GPUs: thousands to 100k+ (frontier)
Explore in 3D →
Your Prompt's Round Trip
What happens in the second after you hit send.
First token: typically well under a second
Explore in 3D →
The Software Stack
Eight layers between 'hi' and a transistor flipping.
Layers shown: 8
Explore in 3D →
How to use the Atlas
- Real · Blueprint · X-ray · Exploded: four ways to look at the same thing.
- Signal: color-coded flows of data, light, power, heat and coolant.
- Numbered dots: tap one (or press T for a guided tour) and the camera flies in to explain it.
- Concept labs under each exhibit: small simulations you can poke at.
- Talk about it: two or three questions per exhibit, with no right answers, plus a printable sheet of all of them.