Complete AI Infrastructure

AI training pipelines consume data from storage at rates that depend on both the bandwidth available between storage and compute and the latency at which individual reads complete: a storage subsystem that cannot keep the GPUs fed creates the same idle-time problem as a congested network fabric. Combining Arista's AI-optimised switching fabric with VAST Data's high-performance NVMe storage in a validated reference architecture addresses both constraints simultaneously: the Arista fabric provides the lossless RoCEv2 transport required for efficient GPU-to-GPU communication during training collectives, while the VAST Data DASE (Disaggregated Shared Everything) architecture provides the high-bandwidth, low-latency NVMe access required to stream training data to GPU memory without creating a storage bottleneck. The two systems are independently scalable: additional GPUs are added as compute nodes, and additional VAST storage nodes are added as NVMe capacity needs grow, without either tier constraining the other's scaling path.

At the spine tier, the 7800R4 AI Spine provides the 800G modular interconnect for the largest GPU clusters, with the 7700R4 Distributed Etherlink for cross-row and cross-pod fabric extension. At the leaf tier, the 7060X6 provides 800G density for GPU-rack connectivity, with one leaf-pair per 8-GPU server rack as a typical design point, and the 7060X5 serves 400G inference and mixed-workload tiers. PFC and ECN are tuned across all tiers for lossless RoCEv2 transport, with LANZ telemetry streaming micro-burst data to CloudVision for real-time fabric visibility. Multi-tenant segmentation using EVPN and VRF isolation allows the same physical fabric to carry both AI training traffic and the storage traffic to the VAST cluster without interference: separate logical fabrics run on the same physical infrastructure without separate hardware deployments for each traffic type.