Broadcom Jericho4 — Deep-Buffer AI Scale-Across Fabric Router

Specifications

Overview

Broadcom's Jericho4 is the deep-buffer member of Broadcom's Ethernet switch and router portfolio, purpose-built for the scale-across tier of AI networking. Scale-across links AI clusters that sit in different buildings, campuses or regions; because RoCE traffic must not drop packets to keep distributed training and inference efficient, the fabric needs large buffers to absorb congestion over long distances.

Jericho4 addresses that with high-bandwidth memory (HBM) based deep buffering. It carries 3.2T HyperPorts, each aggregating four 800GE links, and is built to deliver lossless RoCE beyond 100 km. The device is UEC (Ultra Ethernet Consortium) compliant, matching the industry push toward open, AI-optimised Ethernet. It is in production volume as of 2026 and sits alongside Broadcom's Tomahawk 6 for scale-out and Tomahawk Ultra for scale-up, giving AI data center architects a single-vendor Ethernet stack from rack to region.

For buyers building distributed GPU clusters, Jericho4 is the tier that connects AI data centers into a single fabric. It pairs today with 800G coherent optics and is positioned for the 1600ZR class of coherent modules as they arrive from 2027.

Key Benefits

Lossless RoCE beyond 100 km keeps distributed AI traffic drop-free across sites. HBM-based deep buffer absorbs congestion at distance. 3.2T HyperPorts aggregate four 800GE links for high-density uplinks. UEC-compliant open Ethernet aligns with AI-optimised fabric standards.

Applications

Scale-across AI training and inference fabrics, multi-datacenter GPU clusters, distributed inference and agentic AI backbones, campus and regional AI superfactories, and unified Ethernet AI fabrics spanning scale-up, scale-out and scale-across tiers.

Related Products