BCM78900 network switches have highest radix of 200GbE ports
Broadcom’s BCM78900 high radix, high bandwidth network switching devices supports up to 64 x 800GbE, 128 x 400GbE, or 256 x 200GbE ports. They feature a maximum of 64 integrated Peregrine SerDes cores, each with eight integrated 106Gbits per second PAM4 SerDes transceivers and associated PCS for native support of physical connectivity options.
Claimed to be the world’s highest radix, at 200GbE ports, a single chip supports 256 ports, enabling flat, low latency artificial intelligence / machine learning (AI/ML) clusters
The BCM78900 also delivers high bandwidth, glueless network connectivity of up to 51.2Tbits per second on a single chip.
Its architecture delivers L2 and L3 switching, routing, and tunnelling capabilities at line rate and maximum port density, with low power and latency. Software compatibility is maintained across the StrataXGS portfolio to simplify customer designs and reduce customer time to market, said Broadcom.
As server interfaces transition to higher Ethernet speeds and virtualisation continues to increase link utilisation, data centre networks demand switches with dense 50GbE to 800GbE connectivity at the edge and aggregation layers. According to Broadcom, being equipped with up to 64 x [PM8x100] SerDes and the ability to configure 50GbE, 100GbE, 200GbE, 400GbE, and 800GbE ports, allows a single switching chip to build scalable, cost-effective leaf, spine, blade, and aggregation switches across the entire data centre.
The 51.2Tbit per second shared buffer architecture is claimed to provide lowest tail latency for RoCEv2 and other new RDMA protocols. Other capabilities are advanced adaptive routing, dynamic load balancing, and support for end-to-end congestion control capabilities which have been specifically designed to handle the large, low entropy flows typical of AI/ML workloads.
There is also support for Clos and non-Clos topologies such as torus, Dragonfly, Dragonfly+, and Megafly. Another feature is hardware-based link failover for improved network resiliency and reduced JCT.
Target applications are data centre aggregation, edge and spine switches, data centre fixed and chassis/modular switches, AI/ML clusters, cloud computing and large-scale enterprise backbones.
