Skip to content

Benchmarking AIML Infrastructure at Scale: AIML Architectures, AI Workloads and Reproducibility for High Performance Data Centers - BRKDCN-2651#

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

Abstract

Rakesh Kumar (Cisco), Ram Gandikota (Cisco)

  • Why: The hardest open problem in AI fabrics is trustworthy, reproducible benchmarking — NCCL collective performance is dominated by network behavior and tuning is opaque. This is one of the few sessions engaging with methodology rather than just results.
  • What: The benchmarking methodologies Cisco used across large-scale cluster engagements: alignment with open standards, reproducibility strategies, application-level KPIs like NCCL collectives, network-tuning challenges, and how packet trimming and UEC-based (Ultra Ethernet Consortium) feature alignment are reshaping Ethernet standards for AIML clusters (from abstract only).
  • How: Case-study-driven methodology work — reducing test drift between benchmark and production, making network behavior a first-class measured variable.
  • Where applicable: AI training/inference clusters; the UEC-alignment content matters to anyone planning Ethernet-based AI fabrics. Lessons come from Cisco cluster engagements, so some method choices will be Cisco-toolchain-specific.