Cisco Knowledge Network

Service Provider Webinar Series

Evolved Networking – the AI Challenge

Watch Webex On-Demand

Unlike traditional front-end networks, high-performance computing clusters (HPCC), artificial intelligence (AI), and machine learning (ML) require deterministic high bandwidth connections between many processing nodes. Historically these have been built with dedicated and proprietary technologies like InfiniBand (IB).

These networks have unique requirements and use unique interconnect technologies which have forced data centers to deploy specialized tools and networks. In this webinar, we’ll cover some of the unique characteristics of HPCC, ML, and AI networks and what that means to the network topologies and technologies used to connect these machines. We’ll cover some of the advantages and challenges of using fully generic ethernet for an interconnect and propose some alternative solutions. We will show how Cisco can provide customers with unique flexibility as their requirements evolve.

In this session, we’ll discuss the following:

  • Why operators are moving away from traditional HPCC interconnect to ethernet-based interconnect.
  • What is the difference between traditional HPCC and AI/ML?
  • How the traffic patterns of HPCC and AI/ML change the relevance of metrics like throughput, flow completion time, and job completion time.
  • Difference in traffic patterns for traditional front end networks vs HPCC vs AI.
  • The pros and cons of using generic ethernet with ECMP and fully scheduled fabrics.
  • How various topologies scale.
  • How Cisco Silicon One can help them solve difficult challenges.

Evolved Networking – the AI Challenge – Date originally recorded: December 1, 2022