Live Q&A with Speakers – James Zeng
summary
Live Q&A gives concrete operational answers: 7:1 (16:1 feasible) super‑spine over‑subscription because most traffic stays inside pods or racks; 400G copper DAC in Gen2 for intra‑rack cost and power reasons; DPR is destination‑index based (not per‑packet 5‑tuple hashing); controller uses a streaming protocol to push forwarding state and falls back to BGP then reconfigures in ~30s, with modest transient latency and no queue buildup.
AI-written summary · about
chapters (19)
- 0:00Welcome
- 0:24Network oversubscription rationale
- 1:48IB vs Rocky performance
- 2:30Copper DAC for Gen2 links
- 3:35Backend vs frontend topology
- 4:37Converged AI networking prospects
- 5:41DPR vs packet hashing
- 7:50Traffic engineering adaptation
- 9:57Meta’s AI workload optimizations
- 11:44Controller-node protocols
- 12:28Virtualization for AI workloads
- 13:13AI data center lessons
- 15:00Link failures in TE
- 16:07Service job provisioning
- 17:54AI cluster scalability
- 19:03Key AI fabric property
- 20:50Network partitioning in TE
- 21:55GenAI training challenges
- 22:37Closing
