Virgo: Scale-out Data Center Network for AI
summary
Virgo is a dedicated accelerator-to-accelerator fabric that collapses to a flat two-layer non‑blocking topology using high‑radix switches and 200Gbps 30C “skinny” links to maximize switch ports. A GPU requiring 1.6Tb/s uses eight skinny links so a single plane failure leaves 87.5% bandwidth; multi‑plane design plus adaptive routing and SRv6 source routing provide resilience. Built with high‑resolution telemetry and C‑SIG in‑band signaling, Virgo is running with NVIDIA Vera Rubin NVL72 and next‑gen TPUs across multi‑building campuses.
AI-written summary · about
chapters (9)
- 0:00Virgo and AI networking
- 2:09Core design principles
- 4:38Virgo topology and architecture
- 7:49Multi-plane design and skinny links
- 10:36Multipath routing and load balancing
- 11:42High-resolution telemetry
- 13:51C-SIG congestion signaling
- 16:20Virgo deployment and scaling
- 17:43Industry collaboration
related talks
17mMeta’s Network Journey to Enable AI
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
22mScaling agentic AI: From infrastructure to enterprise outcomes – Duncan Ng
25mKeynote
15mAquila: A unified, low-latency fabric for datacenter networks
52mKeynote Address: Insights Gained from Delivering Two Generations of AI Supercomputers...
05mFrom Open Source to Real-World Scale
14mArchitecting Cloud Storage for AI Native Applications