Networking & Interconnect Foundations
Glossary
- Back-end fabric
- The dense, specialized network that connects accelerators directly to one another within a cluster so they can exchange data during training. It is distinct from the front-end network that links the data center to the outside world and most directly determines usable performance.
- Bandwidth
- The amount of data a link can move per unit of time, analogous to the width of a pipe. It governs how quickly bulk data, such as large transfers between accelerators, can be shifted across the network.
- Collective operation
- A coordinated exchange in which every participating accelerator must both send and receive data before any can proceed, used to synchronize chips after a training step. Its performance is paced by the slowest link, making consistent low latency essential.
- Dark fiber
- Installed optical fiber that has no transmission equipment attached and therefore carries no data, though it is physically capable of doing so. Controlling dark fiber lets an operator supply its own electronics and capture capacity upgrades over time without new construction.
- Latency
- The time a single piece of data takes to travel from one end of a link to the other, analogous to the length of a pipe. It has an unavoidable floor set by distance and the speed of light in fiber, plus an added portion from switches, conversions, and queuing.
- Optical transceiver
- A device at each end of a fast link that converts electrical signals to light for transmission and back to electrical on receipt. Consumed by the million in a large cluster, its supply, cost, and power draw are first-order constraints on how large and fast a fabric can grow.
- Scale-up versus scale-out
- The two main tiers of the AI back-end network. Scale-up connects a small group of nearby accelerators at very high bandwidth so they act as one unit; scale-out uses switched networking to join those groups across racks and rows into a cluster of thousands.
- Subsea cable system
- An undersea fiber-optic cable, with periodic repeaters powered from shore, that carries the overwhelming bulk of intercontinental data. Its capacity, latency floor, and coastal landing stations shape where AI compute can be sited relative to its data and users.