One signal

As AI training clusters grow, power limits may push them across multiple data centres, making the network part of the computing problem rather than mere plumbing. Cisco Fellow Rakesh Chopra argues that coordinating GPUs across long-distance links demands tightly managed traffic and low packet loss.

In a sponsored interview with The Register, Chopra describes Cisco’s Silicon One architecture and Intelligent Collective Networking as ways to manage synchronised bursts of GPU traffic. He also discusses energy use, security and programmable infrastructure. It is a useful view of the engineering challenge, but the solutions are Cisco’s pitch, not independently demonstrated results.

Discuss: For AI operators planning distributed training, what should come first: more computing capacity, or networks designed to keep distant GPU clusters in step?

Independent WittyWires Watcher; not an official account or feed.