Skip to content

Demystifying AI Infrastructure for a Network Engineer - BRKDCN-2971#

Cisco Live Europe 2026 (Amsterdam) digest Program Core pick Amsterdam · 9–13 February 2026

Abstract

Paresh Gupta (Cisco)

  • Why: RDMA/lossless Ethernet is squarely this PM's core territory, and this session packages the current AI-fabric networking stack — RoCE, PFC, DC QCN, collective libraries, Spectrum-X, Ultra Ethernet — in one place. Framed as intro, but the content map is the industry's Ethernet-for-AI state of play.
  • What: How GPUs communicate during distributed AI jobs and how that traffic differs from ordinary DC workloads; covers RoCE, PFC, DC QCN; collective operations via NCCL/RCCL/HCCL; NVIDIA Spectrum-X's problem statement and solution; and where Cisco Nexus, Hyperfabric for AI, and Ultra Ethernet fit (from abstract only).
  • How: Fundamentals-first build-up from GPU communication patterns to lossless-fabric mechanisms and vendor approaches.
  • Where applicable: AI/ML cluster fabric design generally. The Spectrum-X comparison is inevitably vendor-slanted; the RoCE/PFC/QCN and NCCL material is standard. Skip if you already operate RoCE fabrics day-to-day.