SwiNOG

Swiss Network Operators Group

Networking for AI

Training Large Language Models (LLMs) or running large inference workloads using XPU clusters introduces distinct challenges for the underlying network. This session explores the core concepts of AI networking, emphasizing open, Ethernet-based architectures. We will examine standard AI workloads and contrast them with traditional data center traffic. Additionally, we will analyze the resulting design principles and implementation strategies, concluding with an overview of emerging industry trends.