Skip to content
Hi, Bot

Network Card + Optics

Turns electricity into flashes of light and back again.

How big?A card about the length of a pencil; the fiber is thinner than a hair inside its jacket.

Real: What it looks like.

Keys: arrows rotate · + / − zoom · 0 reset · 1–4 views · S signal · T tour · L labels · Space spin. Models are stylised and built from code: proportions are honest, details are simplified.

What it is

The network interface card (NIC) moves data between the server and the network. A dedicated chip handles packets so the CPU doesn't have to.

The transceiver plugged into its cage has a tiny laser. It turns electrical signals into pulses of light that travel down optical fiber, and a receiver at the other end turns them back.

AI NICs use RDMA (remote direct memory access): one GPU's memory can write straight into another server's GPU memory without bothering either CPU.

Why AI needs it

Training splits one model across thousands of GPUs that must constantly swap results. Most AI servers have one fast NIC per GPU, so the network is as important as the chips.

Every labeled part

  1. 1

    NIC chip

    Packet processor under the heat sink. Explode the view to lift the heat sink off.

  2. 2

    Port cage

    Where a transceiver or cable plugs in.

  3. 3

    Optical transceiver

    A laser and a light sensor in a module the size of a pack of gum.

  4. 4

    Optical fiber

    Glass thinner than a hair. Light bounces down it for hundreds of metres. Signal mode shows the pulses.

  5. 5

    PCIe edge

    Plugs into the server, often right next to a GPU.

  6. 6

    Bracket

    Holds the card at the back of the server.

Try it · concept lab

Chop it into packets

Send a message across the network.

Packet size

#1#2#3#4#5#6#7#8#9#10#11#12#13#14#15#16
switch

Packets

16

Wire time @ 800 Gb/s

655 ns

Arrived in order?

—

Each packet carries an address and a number. Paths can reorder them; the receiving NIC puts them back in order and writes straight into GPU memory (RDMA).

Big idea: Networks move data in small numbered packets. AI NICs reassemble them straight into GPU memory, skipping the CPU.

Swap it: other ways to do the same job

  • InfiniBand vs. Ethernet

    InfiniBand was built for supercomputers and has very low delay. AI-tuned Ethernet (RoCE) is catching up and is everywhere else.

  • Copper cable (DAC)

    Cheaper and uses less power than optics, but only works for a couple of metres.

  • Co-packaged optics

    Puts the lasers right next to the switch chip to save power. Newer and harder to repair.

Talk about it

  1. Q1

    This card turns electricity into flashes of light. Why do you think light is used for long trips instead of plain wires?

  2. Q2

    Thousands of chips must keep sharing results while they learn. What happens in a group project when people stop sharing?

For grown-ups: there are no right answers here. Ask a question, then ask "why do you think that?" The reasons matter more than the answer.

Printable question sheet (PDF)

Connected exhibits