Ethernet Based Fabric for AI Cluster - Powered by Silicon One & SONiC, an Ultra High Performance, Scalable & Non-Blocking Ethernet Fabric - BRKCOC-3005#
Cisco Live Europe 2025 (Amsterdam) digest Program Core pick Amsterdam · 9–14 February 2025
Speakers: Lieven Colman (Cisco), John Banner (Cisco)
- Why: AI-cluster fabric build-out is the hottest physical-DC topic right now, and this is a practitioner's account of actually standing up non-blocking Ethernet fabrics for 256- and 512-GPU training clusters — scale, config choices, and failure discussion, not just a roadmap.
- What: A Cisco-on-Cisco internal deployment story covering the technology stack (Silicon One switching running SONiC, Nexus-based networks), the key configuration decisions, the challenges hit, and lessons learned while building production AI-cluster fabrics. (from abstract only)
- How: Builds non-blocking (oversubscription-free) Ethernet fabrics for GPU compute using commodity-OS (SONiC) on Silicon One silicon, integrated with Nexus-based underlay/management networks; the abstract promises concrete design trade-offs rather than architecture diagrams alone.
- Where applicable: Cisco-specific silicon and NOS stack, but the design questions (buffering, congestion management, non-blocking topology at 256–512 GPU scale) apply to any Ethernet-for-AI effort. Watch for whether they discuss RoCEv2/PFC vs. alternative lossless approaches — the abstract doesn't say which they chose.