← Course contents · Part 2

· Visual lesson · Session 17 of 56


How does PCIe connect everything?

PCIe connects processors with devices such as GPUs, network cards, and SSDs. Follow the whole connection, not just the slot label.

Picture the idea

Two roads share one bridge

  1. Device A + device B

    Each requests 20 GB/s: 40 GB/s together

  2. Shared upstream connection

    Assumed sustained capacity: 30 GB/s

  3. CPU-side destination

    Cannot receive the full combined demand through that path

Illustrative bandwidth budget, not a specific PCIe generation. Actual sharing need not be evenly split.

Several devices can share one path

A PCIe link has a generation and a lane width. x16 means 16 lanes, not 16 GPUs. The actual connection may use fewer lanes than the physical slot suggests.

Devices behind a switch can share its upstream path. In the diagram’s invented example, two devices want 40 GB/s together but the shared path carries 30 GB/s.

Inspect without changing

On an authorized NVIDIA Linux host with driver tools:

nvidia-smi topo -m

Read its legend to understand device relationships. It is not a bandwidth benchmark. Use the paper exercise on a Mac with an Apple GPU.

Do not change security or isolation settings to chase a faster test without the required review.

Try it yourself

Draw two GPUs and a NIC behind one switch. Label each link with an assumed capacity.

Check your answer

Find the narrowest shared part of the path. Faster endpoints cannot make that connection carry more than its capacity. Actual sharing depends on traffic and arbitration.

Sources and further reading

← All sessions