Arista 7358X4 Series

The Arista 7358X4 Series is a purpose-built modular AI fabric switch platform addressing the bandwidth, congestion management, and lossless Ethernet requirements of large-scale GPU training clusters. Available in 4-slot (7358X4-4) and 8-slot (7358X4-8) chassis configurations, the series reaches 460.8 Tbps of aggregate switching capacity when fully populated with 36x QSFP-DD 400G line cards (7358X4-LC-36DD), or equivalent capacity using 800G line cards as they become available. The platform supports leaf attachment at 400G and spine interconnection at 800G in the same chassis, enabling a two-tier fabric without changing hardware at the spine.

Advanced congestion control capabilities including hardware-based ECN marking, per-destination output queue management, and deep ingress buffering address the collective communication traffic patterns (AllReduce, AllGather, ReduceScatter) that produce incast and congestion events in AI training networks. Full lossless Ethernet operation via PFC (Priority Flow Control) and DCQCN (Data Center Quantized Congestion Notification) eliminates packet loss that would degrade GPU utilisation during training iterations, while RDMA over Converged Ethernet (RoCE) support provides the low-latency GPU-to-GPU communication path that modern AI frameworks require.