East-West Traffic Fabric

Category: infrastructure

The network architectural data flow model describing horizontal server-to-server communications within a data center cluster, as opposed to vertical client-to-server traffic.

Traditional cloud infrastructure is optimized for North-South traffic (users requesting web data from a server). AI data centers, however, are dominated by East-West traffic, where thousands of parallel GPUs continuously exchange neural network gradients, weights, and replication packets during distributed training. Managing this requires flat, non-blocking InfiniBand or RoCE (RDMA over Converged Ethernet) network topologies featuring dedicated network interface cards (NICs) per compute node.

Common Examples

• We deployed an ultra-low-latency InfiniBand spine-and-leaf fabric to handle the massive East-West traffic generated by our multi-agent simulation routines.
• A bottleneck in the East-West traffic pipeline will cause high GPU wait times, stalling our training iterations and inflating our compute budget.