Interconnect & Networking

Explainer

Interconnect & Networking explained

Interconnect is the set of links that move data between chips, between boards and between racks. It includes wires on a package, cables inside a server, and the network switches and optical fibre across a data center. A processor is only as useful as the speed at which it can get data in and out.

Common standards include PCIe, which connects a CPU to devices, NVLink, which connects GPUs to each other, Ethernet and InfiniBand for server networks, and CXL for sharing memory. A newer standard called UCIe defines how chiplets inside one package talk to each other.

How it differs from the usual approach

Compute chips are measured in operations per second. Interconnect is measured in bandwidth (how much data per second) and latency (how long one message takes). In large AI clusters, chips spend a real share of their time waiting for data from other chips, so a faster network can raise the useful output of the whole system without changing the processors.

Where it is used

• Linking GPUs together so that they work as one large accelerator during AI training.
• Connecting thousands of servers in cloud and AI data centers with Ethernet or InfiniBand.
• Joining chiplets inside one processor package.
• Connecting CPUs to memory expanders and storage using CXL and PCIe.
• Optical transceivers that carry traffic between racks and buildings.

What to watch

• Optical links are replacing copper over ever shorter distances. Co-packaged optics puts the optical parts next to the switch chip to cut power use.
• Ethernet is being adapted for AI clusters, The Ultra Ethernet Consortium published its 1.0 specification for that purpose in June 2025.
• Chiplet standards such as UCIe aim to let dies from different makers sit in one package.
• Network speeds in data centers have moved from 100 Gb/s per port toward 800 Gb/s and beyond.

Frequently asked questions