Pete Del Vecchio, the product line manager for the Tomahawk family at Broadcom, discusses AI interconnect, focusing on physical interconnect improvements for AI devices using both Jericho and Tomahawk solutions. He outlines a three-pronged strategy to reduce cost, power, and latency: maximizing copper usage, minimizing optics costs with linear pluggables, and utilizing co-packaged optics solutions. He introduces the Tomahawk AI cluster architecture, emphasizing high bandwidth and radix for flatter topologies, load balancing, congestion management, and real-world deployments. He details Broadcom's cognitive routing suite, including global load balancing, reactive path rebalancing, fast link failover, and drop congestion notification, and also presents Broadcom's silicon solutions for AI, including NICs, PCI switches, custom ASICs, NIC chiplets, and optical engines, and provides an example of a current deployment architecture using Tomahawk 5 switches.
Sign in to continue reading, translating and more.
Continue