Scaling AI Network with DSF
summary
A credit based disaggregated schedule fabric converts packets into fixed-size cells, sprays them across fabric cards, and reassembles at egress to avoid hashing skew and out-of-order RDMA; a single L1 AI zone supports 4.6k 800G GPUs, four zones unify to ~18k GPUs per building, and multi‑building pods stitch to mega clusters. Performance hits theoretical rooflines on alltoall (128/256/512) and real GenAI jobs show similar words/sec intra‑ and inter‑zone. Failure handling uses input balance mode to rebalance reachability and avoid spine congestion.
AI-written summary · about
chapters (6)
- 0:00The impact of generative AI
- 1:37Challenges of legacy networks
- 5:10The Disaggregated Scheduled Fabric
- 8:02Architecting for GenAI clusters
- 13:20Fault handling and balance
- 16:45Performance and future work
related talks
17mMeta’s Network Journey to Enable AI
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
09mAI for NI
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
28mScaling agentic AI: From infrastructure to enterprise outcomes
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale
22mGenAI Training In Production: Software, Hardware & Network Considerations
23mNetworking for GenAI Training and Inference Clusters