Ram Huggahalli – Microsoft

Title: AI-Centric Datacenters and their Diverse Network Requirements

Bio: Ram Huggahalli is a Principal Hardware Engineer in Microsoft’s Azure Hardware and Systems Infrastructure organization. As part of the Strategic Planning and Architecture function within this organization, he leads the Optical Technology Pathfinding team that is evaluating emerging optical technologies for AI and Cloud DC applications. Prior to this role, Ram was part of the Microsoft Cobalt processor team contributing to its I/O architecture. Ram also worked for Intel on many projects spanning memory, server networking and large scale HPC fabrics with contributions to Intel Xeon processors and Omni-Path Architecture. Ram holds 20 patents and a Masters degrees in Electrical Engineering and Engineering Management from the University of Missouri.

Abstract: Datacenters equipped with specialized hardware (or accelerators) have enabled AI capabilities to be infused into a multitude of applications. As accelerators become increasingly cost efficient, we can anticipate even more exciting, richer and broader AI usages. The role of networking and the importance of network efficiency is now acknowledged as critical aspects of AI systems. However, short term directions indicate a divergence in network requirements leading to potentially alternative ecosystems. In this talk, we share some apparently diverging requirements and discuss the potential for Ethernet, once again, to emerge as a solution that provides a converged solution while also not compromising on efficiency.