Networking for AI and Ultra Ethernet#

Massimo Magnani

• Why: AI cluster networking is the biggest forcing function on datacenter fabric design today — it is reshaping Ethernet itself (congestion management, telemetry, load balancing at job granularity). Even a regional-NOG take is worth the PM's time as a pulse check on how far Ultra Ethernet thinking has spread beyond the hyperscalers driving it.
• What: The program publishes no abstract and no extractable materials were available, so content cannot be described beyond the title — the session covers networking for AI/HPC workloads and the Ultra Ethernet effort. Treat this as a topic-signal pick, not a validated-results pick. (from abstract only — actually, no abstract at all; title and speaker only)
• How: Unknown from the program text. If it follows the shape of similar 2025 NOG talks, expect UEC's profile (telemetry, congestion control, RDMA-over-Ethernet semantics) — but that is inference, not a claim about this deck.
• Where applicable: Relevant to anyone tracking where lossless-Ethernet/RDMA fabric requirements are heading; depth and neutrality unknown.