Editorial read
In depth · editorial + model
Programmable network adapters (SmartNICs) and data processing units (DPUs) that terminate each node's high-speed link and offload RDMA/RoCE, congestion control, storage and tenant isolation from the host CPU. It is the fabric endpoint on every node - the layer that turns nominal link speed into usable, isolated, congestion-managed bandwidth. GPU fabrics now ride 400/800G RDMA NICs and DPUs whose generation (ConnectX-7/8, BlueField-3) paces effective bandwidth, congestion behavior and in-network collectives, and sovereign/multi-tenant AI clouds increasingly require the DPU's hardware isolation. NIC/DPU supply, firmware or congestion-control immaturity throttles effective fabric throughput and utilization even when switches and optics are available.