AI-Ready Data Center

AI data centre network designs differ from general-purpose enterprise data centre designs in one fundamental way: the dominant traffic pattern is east-west rather than north-south. In a general-purpose data centre, the majority of traffic flows north-south, from servers to clients through a border router or load balancer. In a GPU training cluster, the dominant traffic is east-west, between GPUs during collective communication operations, where every GPU exchanges gradient data with every other GPU simultaneously. This traffic pattern requires a non-blocking, uniform-latency fabric where every east-west path has the same bandwidth and the same hop count, so that no GPU in a collective operation is consistently slower than the others due to a network bottleneck. The Arista AI data centre reference architecture addresses this requirement through a spine-leaf topology built from switches that are explicitly designed for uniform, lossless, high-bandwidth forwarding at scale.

The reference architecture spans the full port-density and throughput range of the Arista portfolio: 7010X for power-efficient 1/10/25G server access at under 0.3W per Gbps (296 Gbps, 220 Mpps); 7260X3 for high-capacity leaf-spine at up to 12.8 Tbps in 2RU with 64 MB dynamic buffer, scaling to 27,648 10/25G servers in a two-tier design at 3:1 oversubscription; 7700R4 Distributed Etherlink for 800G cross-row spine interconnect scaling to 4,608 non-blocking 800G hosts; and 7800R4 for modular 800G spine in the largest clusters. Latency across the fabric operates in the 550 ns to 1,500 ns range in non-blocking configurations. EOS runs a single binary across all platform tiers, implementing consistent PFC and ECN profiles and providing hitless In-Service Software Updates that allow EOS version upgrades without rebooting switches and disrupting GPU training jobs in flight.