Скачать или смотреть Networking for AI Scaling, presented by Broadcom

Ram Velaga SVP & GM, Core Switching Group - Broadcom

As AI models grow in complexity and scale, building ultra-large AI clusters demands a reimagining of compute, memory, and I-O architectures to support fast, flexible, and recursive reasoning. This evolution is no longer confined to a single rack or even a single data center—it spans across entire geographies, interconnecting hundreds of thousands to millions of XPUs.Meeting these unprecedented demands requires a next-generation networking fabric that can· Scale up within and across a few racks, seamlessly connecting thousands of XPUs;· Scale out across entire data centers in a flat, high-performance topology, supporting tens to hundreds of thousands of XPUs;· Scale across geographically distributed data centers, creating globally unified AI fabrics of million-plus XPUs.In this keynote, Broadcom will showcase· Key innovations that unlock this scale—powered by open, standards-based Ethernet;· The latest advancements in the Scale-Up Ethernet (SUE) architecture, including product roadmap highlights for in-rack, inter-rack, and inter-DC connectivity.