Accelerate AI Deployment

Deploying an AI data centre fabric at 400G or 800G port density involves selecting, integrating, and validating switching hardware, optics, cabling, and EOS configuration profiles that together produce a lossless fabric capable of sustaining the traffic patterns of GPU training and inference workloads. Arista and Broadcom have jointly developed and validated a pre-tested 400G and 800G leaf-spine reference architecture that covers hardware selection, cabling topology, EOS PFC and ECN tuning profiles, and AI Job Observability configuration, reducing the validation effort that would otherwise fall on the customer's network engineering team before a production cluster can come online. The pre-validation covers 2-tier leaf-spine topologies designed to cluster tens of thousands of accelerators, with both fixed-configuration and modular-chassis options at the spine tier to match the scale of the deployment.

The 7800R4 AI Spine is the anchor platform at the spine tier for the largest deployments, providing 800G port density in a modular chassis configuration with the non-blocking, high-radix interconnect required to prevent bottlenecks at the spine layer as the cluster scales. The 7060X6 Series provides 800G port density at the leaf tier in a fixed 1RU configuration, pairing with the 7800R4 spine in a non-blocking 2-tier design. For deployments where the AI fabric is part of a mixed 400G/800G environment (GPU training on 800G, inference and general compute on 400G), the 7060X5 at 400G and 7060X4 for mixed workloads provide the leaf options. Per-flow and per-job visibility through LANZ and AI Job Observability allows operators to trace the source of tail latency and throughput variance to specific flows, jobs, or fabric tiers, the diagnostic capability that distinguishes a production-ready AI fabric from a research deployment.