Data Center Switch (Spine)

Overview

A data-center spine switch is the high-radix core of a leaf-spine fabric: every leaf switch in a row or pod connects up to it, so its job is to move packets between any two ports at line rate with predictable, sub-microsecond latency. The whole forwarding plane is built around a single merchant switching ASIC that provides 12.8 Tbps of non-blocking capacity across 32 high-density cages, each running a 400G OSFP or 100G QSFP28 optic. Because it sits in a hot aisle and cannot drop traffic, the design carries hot-swap fan trays and dual load-sharing power supplies, and pushes air strictly front-to-back to match the rack's cooling layout.

How it works

Light arrives at the front panel and is converted to electrical lanes inside each Optical Transceiver, which plugs into a cage on the Port Cage Assembly. The high-speed lanes pass through the SerDes Retimer devices, which re-clock and equalize the PAM4 signaling before it reaches the silicon, then land on the Switching Mainboard.

There the Compute SoC Module switching ASIC parses each packet, looks up the destination in the TCAM, and forwards it through the shared Packet Buffer Memory toward the egress port, all in hardware. The Control-Plane CPU runs the network OS and routing protocols off to the side, programming the forwarding tables rather than touching packets. The ASIC draws hundreds of amps on its core rail, supplied by the multi-phase Voltage Regulator Module stages and cooled by the ASIC Heatsink.

Power and signaling reach the line card through the passive Midplane, fed from two supplies in a 1+1 arrangement. The Fan Tray modules pull air across the heatsinks under thermal control, and operators reach the box out-of-band through the Management Module, with the Front Panel Indicators showing health at a glance.