The Conveyor: Fabric, Interconnect, and Data Movement
The network is not plumbing between computers; it is part of the computer. A frontier model lives across thousands of chips that must act as one, and they spend enormous time reconciling their work through collective operations — the conveyor belt between stations. A slow belt starves every machine tool, and because training runs in lockstep, the slowest straggler sets the pace for all. Beneath the fabric sits the deeper truth: moving data, not doing arithmetic, is what the factory fights. Each step outward in the memory hierarchy costs an order of magnitude more time and energy, so locality — keep data close, reuse it fully, talk to neighbors — is where throughput and watts are actually won.
Knowledge check
In synchronized large-model training, why can a single slow or failing link drag down the throughput of an entire cluster?
Show the answer
Because training advances in lockstep through collective operations: every chip computes a piece of each step, then all must reconcile their results before any can proceed. The exchange completes only when the slowest participant finishes, so one congested link, throttling node, or failing cable makes the whole cluster wait. In a lockstep line the tail — the slowest station — governs throughput, not the average, which is why fabric speed, topology, and straggler mitigation matter as much as arithmetic.