Distance running: Broadcom’s Jericho4 chip connects AI data centers up to 60 miles apart

Broadcom has launched its next-generation Jericho4 networking chip designed to connect data centers up to 60 miles apart and accelerate artificial intelligence computation. The chip addresses growing demands from cloud computing giants like Microsoft and Amazon for faster, more sophisticated networking infrastructure as AI workloads become increasingly computationally intensive and require connecting thousands of graphics processors.

What you should know: The Jericho4 introduces several performance enhancements specifically designed for large-scale AI data center operations.

• A single system can encompass roughly 4,500 chips, enabling massive deployment across extensive networks that operate inside and between data centers.
• The chip uses the same high-bandwidth memory (HBM) found in AI processors from Nvidia and AMD to handle the enormous volume of data traffic.
• Broadcom manufactured the Jericho4 using TSMC’s advanced three nanometer process technology.

How it works: The chip’s design tackles network congestion issues that plague high-traffic AI computing environments.

• “The switch is actually holding that traffic (in memory) till the congestion frees up,” explained Ram Velaga, senior vice president and general manager of Broadcom’s Core Switching Group.
• The longer the distance data must travel from the chip to its destination, the more memory designers must include in the chip.
• The Jericho4 also encrypts data to enhance security when transferring information beyond physical data center walls.

Why this matters: Building and deploying AI has become more computationally intensive, requiring cloud companies to string together thousands of GPUs with increasingly sophisticated networking infrastructure.

• Security during data transfer is crucial for cloud companies due to potential attacks that could intercept data before reaching its destination.
• The chip’s ability to connect data centers over 60 miles apart enables more flexible and distributed AI computing architectures.