Compute interconnects include links within one server and networks across many servers. Their bandwidth and latency determine how efficiently distributed AI workloads can exchange model states, activations and results.
As clusters grow, communication can consume more time than calculation. Hardware topology, network protocols and workload partitioning must therefore be designed together so more accelerators produce useful scaling rather than additional waiting.
ELI5
A compute interconnect is the set of fast connections that lets chips and computers exchange data. It is like the road network between factories that must share materials to build one product.
For example, many accelerators training one model need to exchange updates repeatedly. If the connections are slow, adding more chips can create traffic jams instead of finishing the work faster.
