Broadcom's Jericho4 boosts data center AI networking

Broadcom is now shipping its silicon division's next-generation Jericho networking chip, touting better performance for data center artificial intelligence workloads across further distances.

Jericho4 debuts several improvements and new features designed to boost networking traffic speeds in large networks operating inside and between data centers. AI appetites are placing strains on data centers, which require thousands of GPUs, tensor processing units and other accelerators to crunch increasing AI workloads.

Jericho4 delivers 51.2 Tbps of lossless, deep-buffered Ethernet, according to the company. That will allow AI workloads to be quickly scaled across racks, clusters and data centers in different geographic locations without losing performance.

While hyperscalers such as Google, Microsoft and Amazon are strengthening cloud offerings with massive data centers built on one site, Broadcom's Jericho4 enables companies using smaller, spread-out data centers a chance to speed up through interconnectivity, a key component when cobbling together AI infrastructure.

"Building these AI factories is a challenge in that they create a very heavy footprint in terms of power and cooling costs," said Jim Frey, an analyst at Enterprise Strategy Group, now part of Omdia. "There's not always enough power available within existing areas with all the build-out that's happening. So there's this need to be more flexible."

Energy efficiency has become a concern for enterprise customers trying to quickly scale AI workloads as power-hungry GPUs drive up energy consumption. Jericho4 promises 40% power reduction per bit, which will help lower operational costs.

The Jericho4 chip gives companies flexibility as they look at infrastructure options, according to Frey.

"This was a problem that needed to be solved," he said. "This won't be the only solution out there. But Broadcom has really taken on this distributed architecture challenge."

Jericho4 can interconnect more than 1 million processing units, including GPUs, CPUs, field-programmable gate arrays and other accelerators across multiple data centers. Connecting processors across multiple facilities is increasingly necessary as AI models grow and become more complex.

A single Jericho4 system scales to 36,000 HyperPorts -- each operating at 3.2 TB per second -- allowing workloads to spread over 60-mile distances. Jericho4 is manufactured on TSMC's 3-nanometer process, and the company said the units offer lower power usage, reduced cost and higher system reliability.

The new chip adds to Broadcom's Tomahawk 6 and Tomahawk Ultra switch chips in the networking lineup.

Shane Snider, a veteran journalist with more than 20 years of experience, covers IT infrastructure at Informa TechTarget.