During the presentation, Méndez analyzed the network traffic characteristics of Machine Learning (ML) training and inference, and the requirements to build network fabrics to interconnect xPUs at speeds of 200/400/800 Gbps. In addition, the session included a comparison between InfiniBand and Ethernet technologies, as well as current techniques for optimizing RDMA (RoCEv2) traffic in IP/Ethernet network structures and the future of UltraEthernet.