Broadcom Chip to Help Power AI by Linking Up Smaller Data Centers

In a significant development for cloud infrastructure, Broadcom Inc. has introduced an innovative version of its Jericho networking chip, aimed at revolutionizing the integration of smaller data centers. This new chip, revealed on Monday, enhances data transfer capabilities, allowing multiple smaller facilities to function together as a cohesive unit for AI model development and execution. Ram Velaga, the senior vice president and general manager of Broadcom's core switching group, highlighted that the Jericho4 can connect over one million processors across various data centers, boasting a data handling capacity four times greater than its predecessor. This enhancement comes at a crucial time when there is a growing demand for equipment that supports the construction of AI frameworks. Broadcom's networking devices, including routers and switches, play a vital role in managing traffic between graphics processing units (GPUs), the expensive technology integral to AI model creation. While some equipment is designed for internal data movement within single data center buildings, there is an increasing need for solutions that can bridge multiple locations. The challenge of powering large GPU clusters has become more pronounced. Velaga pointed out that attempting to cluster 200,000 GPUs could easily require 300 megawatts of power, which is currently impractical for a single physical site. The Jericho chip family is poised to tackle this issue by facilitating more distributed data center architectures. Additionally, companies are seeking to position data center capacities closer to their customer base, enhancing the speed of AI responses. This strategy necessitates the use of interconnected smaller setups, particularly in densely populated urban areas. Broadcom's existing Tomahawk line effectively manages connections within a data center but typically covers distances under one kilometer. In contrast, the Jericho equipment can support connections extending over 100 kilometers. The Jericho4 is set to begin shipping to early adopters, including cloud service providers and networking hardware manufacturers. Full deployment is expected to take around nine months, marking a pivotal step towards optimizing AI data center operations.