AI Data Center Networks - Design and Infrastructure for Modern Training and Inference Workloads

Explore modern AI data center network infrastructure design in this 48-minute NANOG conference talk presented by Juniper Networks experts Michal Styszynski and Mahesh Subramaniam. Gain insights into GPU-to-GPU connect training cluster configurations and inference fabric design options essential for contemporary AI workloads. Learn about critical aspects of fabric-level load balancing, ROCEv2/DCQCN implementation, and data center fabric traffic engineering strategies. Benefit from the extensive industry experience of the speakers, who bring over a decade of expertise in data center networking, cloud architecture, and telecommunications infrastructure development. Discover practical approaches to addressing the unique networking requirements of AI-driven data centers while understanding best practices for optimizing network performance in AI computing environments.