Virgo: Scale-out Data Center Network for AI
summary
Virgo is a dedicated accelerator-to-accelerator fabric: a flat two-layer non-blocking topology built from high-radix switches and skinny links (e.g., 200G 30s lanes) to collapse hops and speed deployment. Multi-planar fabrics give N× scaling and tolerate a plane failure (eight planes → 87.5% bandwidth remains for a 1.6T NIC example). It ships SRv6 source routing and switch-driven adaptive routing, supports C‑SIG in-band telemetry that aggregates per-hop metrics without header growth, and surfaces sub-millisecond bursts that 30s averages miss; deployed for Google Cloud A5x (NVIDIA Vera Rubin NVL72) and next-gen TPUs.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:08AI era demands
- 3:16Virgo design principles
- 6:26Virgo architecture
- 8:59Multi-plane design
- 13:13High-resolution telemetry
- 15:40Demo: sub-millisecond insights
- 17:26Virgo recap and scaling
- 18:50Closing
related talks
17mMeta’s Network Journey to Enable AI
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
15mAquila: A unified, low-latency fabric for datacenter networks
29mLiving on the Edge with DCN Networking
22mScaling agentic AI: From infrastructure to enterprise outcomes – Duncan Ng
24mSilicon innovation and the modern network
19mBluebird: High-performance SDN for Bare-metal Cloud Services
19mScaling AI Network with DSF
