Skip to content

Real World Deployments: Architectural lessons learnt from a 100+ Node AI Factory Deployment - BRKXAR-2039#

Cisco Live US 2026 (Las Vegas) digest Program Core pick Las Vegas · 31 May – 4 June 2026

Abstract

Dave Lucas (Cisco), Nandu Shankar (Cisco)

  • Why: The closest thing in this program to an end-user case study with failure discussion — a real integration bring-up, not a clean-room reference architecture.
  • What: Lessons from standing up a 100+ node "AI factory" spanning compute, network, and storage, with dual GPU vendors, virtual and containerized workloads, RDMA-accelerated storage with GPUDirect, multi-tenant security, and a scalable RoCE front and back-end fabric (from abstract only).
  • How: Explicitly covers the failure modes that don't appear in config guides: multi-OS GPU driver inconsistencies, RoCE fabric optimization, cross-vendor GPU networking quirks, shared-services dependencies, cabling/topology missteps, and typical bottlenecks.
  • Where applicable: Deployment-specific (one environment, Cisco-built), but the cross-vendor integration pitfalls are broadly instructive for anyone running mixed-GPU AI infrastructure.