Facebook Engineering


Meta's Backend Aggregation (BAG) layer is a centralized Ethernet-based super spine network that interconnects multiple data center fabrics across regions to build gigawatt-scale AI clusters. BAG connects tens of thousands of GPUs using modular Jericho3 ASIC chassis with up to 432x800G ports, supporting petabit-range bandwidth between regions. The architecture uses planar or spread connection topologies for inter-BAG connectivity, connects to L2 fabrics (DSF and NSF) with managed oversubscription ratios around 4.5:1, and employs eBGP with UCMP for load balancing. Deep buffer switches handle long cable distances while maintaining lossless congestion control, enabling Meta's Prometheus cluster to deliver 1-gigawatt capacity across multiple data center buildings.

Building Prometheus: How Backend Aggregation Enables Gigawatt-Scale AI Clusters

How BAG Is Helping Us Build Gigawatt-Scale AI Clusters