Meta’s Network Journey to Enable AI
summary
Describes concrete network choices: moves from a 144‑GPU single‑BTSW star (single‑point‑of‑failure) to a two‑stage CLOS with RTSW/CTSW rack and cluster training switches, upgrades from 100G→200G→400G links to match V100→A100→H100 GPUs, and adopts RoCE with PFC and ECN for lossless RDMA. Builds undersubscribed leaf‑spine fabrics, reuses existing fiber plant, and stitches four AI zones into >10,000 GPUs per building via a meshed DC‑Scale plane; design meets scale and latency goals in production.
AI-written summary · about
chapters (5)
related talks
21mDC Networks for Generative AI
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
19mScaling AI Network with DSF
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
20mAnatomy of @Scale Machine Learning Network Operations
18mKeynote: Enterprise Data Center Migration to Hyperscale – Zaid Ali Kahn
25mKeynote
22mResilient AI Supercomputer Networking using MRC and SRv6