Overview
Learn about AI networking fabric design and infrastructure in this 52-minute conference talk from Networking Field Day 33. Explore the challenges of interconnecting AI elements as generative AI and machine learning workloads continue to grow exponentially. Dive into current proprietary technologies' limitations and discover Intel's vision for an open, standards-based approach to networking infrastructure. Examine recent developments in network protocol stacks for reliable transport, including Intel's contributions and the Ultra Ethernet consortium's initiatives. Follow along as Principal Engineer Naru Sundar breaks down key concepts including topology, frontend networks, scale-up networks, power delivery considerations, custom protocols, AI fabrics, and congestion control. Understand how these elements come together in creating efficient, scalable Ethernet solutions for modern AI infrastructure needs.
Syllabus
Introduction
Topology
frontend network
scaleup network
tradeoffs
power delivery
balancing power delivery
HPC and AI
Power and cost
Custom protocols
AI Fabrics
Alra Ethernet
Google Falcon
Congestion Control
Ethernet for Scale Out
Taught by
Tech Field Day