Broadcom Launches Jericho4 Router to Power Distributed AI Across Global Data Centers

Broadcom’s Jericho4 router enables high-speed, secure AI networking across data centers with HBM and 3nm technology.

Jericho4 provides secure, scalable, and lossless interconnects for distributed AI and HPC systems, joining Tomahawk Ultra and Tomahawk 6. The Jericho4 Ethernet fabric router, which connects over a million accelerators, or XPUs, across several data centers to enable dispersed AI workloads, has started delivering, according to a statement released by Broadcom on Tuesday.

The chipmaker claimed that in order to get over scaling constraints in AI infrastructure, Jericho4 offers increased bandwidth, improved security, and lossless performance. Broadcom states that Jericho4 rounds out its networking lineup for HPC and AI workloads by working alongside its Tomahawk 6 and Tomahawk Ultra chips.

Low-power, high-bandwidth memory (HBM) packet buffering is supported by the Jericho4, offering up to 160 times more traffic buffering than on-chip memory, according to the company's product description. Broadcom claims that this allows for zero packet loss performance in networks that are extremely crowded.

Ram Velaga, senior vice president and general manager of Broadcom's core switching group, said in a statement that the Jericho4 family is designed to expand AI-scale Ethernet fabrics beyond individual data centers, supporting 3.2 Tbps HyperPort and congestion-free RoCE for previously unheard-of interconnect efficiency.

Velaga explained that Tomahawk Ultra, Scale up Ethernet (SUE), Tomahawk 6, and Jericho4 play key roles in enabling open and interoperable large-scale distributed computing across racks and data centers. Broadcom claims that a single Jericho4 system can support up to 36,000 HyperPorts with line-rate MACsec encryption, deep buffering, and RoCE transport over more than 100 kilometers, all operating at 3.2 Tbps.