Meta's Backend Aggregation Enables Gigawatt-Scale AI Clusters

Meta's implementation of backend aggregation (BAG) represents a significant advancement in scaling AI infrastructure. By seamlessly connecting thousands of GPUs across multiple data centers and regions, BAG enables the creation of gigawatt-scale AI clusters like Prometheus. This infrastructure relies on interconnecting two different network fabrics: Disaggregated Schedule Fabric (DSF) and Non-Scheduled Fabric (NSF). The BAG layer serves as a centralized Ethernet-based super spine network, supporting immense bandwidth needs with inter-BAG capacities reaching the petabit range. Meta's distributed BAG layers strategically serve subsets of L2 fabrics, adhering to distance, buffer, and latency constraints. The choice between planar and spread connection topologies depends on site size and fiber availability, with the latter enhancing path diversity and resilience. The use of modular chassis equipped with Jericho3 (J3) ASIC line cards provides high-capacity ports for efficient data transfer. Careful management of oversubscription ratios is crucial for balancing scale and performance. As AI clusters continue to grow, BAG is expected to play an increasingly important role in meeting future demands and driving innovation across Meta's global network. The ability to efficiently interconnect and manage vast numbers of GPUs is essential for training and deploying increasingly complex AI models.