NIC · Network Interface Card
The server's network adapter. AI nodes carry multiple RDMA-capable high-speed NICs — often one per GPU — as the on-ramp to the scale-out fabric.
Also written as: NICs
Current numbers
~100 ns switch hop; ~1 us cited application examplefabric latency must name its measurement boundary: NVIDIA cites about 100 ns for an InfiniBand switch hop and about 1 microsecond in application-latency examples; NIC-to-NIC and RoCEv2 results are topology, NIC, software and workload specific
131,072GPUs in a two-tier fabric of 512×100G switches with each 800G NIC split into eight 100G planes (MRC consortium design; four-plane 4×200G and eight-plane 8×100G systems deployed)
~18× GB200 / ~9× GB300NVLink 5 per-GPU scale-up (~900 GB/s per direction) against one per-GPU scale-out NIC — ~18× on GB200/CX-7/400G, ~9× on GB300/CX-8/800G; keep collectives in scale-up