Broadcom has announced the commercial availability of its Jericho4 Ethernet fabric router, a next-generation platform engineered to support the demands of distributed AI computing. The new system, now shipping to customers, is designed to overcome the scaling limits of traditional data center networks, enabling high-bandwidth, secure, and lossless interconnectivity across geographically distributed infrastructure.
As AI workloads grow in complexity and size, infrastructure requirements are increasingly pushing beyond the capabilities of single data centers. Emerging AI models often require tens to hundreds of megawatts of compute power, distributed across multiple sites.
To address this challenge, Broadcom has developed Jericho4 to serve as a cornerstone for wide-area, Ethernet-based AI fabrics, enabling seamless connectivity across racks, data halls, and even remote data centers.
Jericho4 is capable of interconnecting over one million XPUs (e.g., GPUs, TPUs, or custom AI processors), leveraging Broadcom’s high-performance networking portfolio, which also includes Tomahawk 6 and Tomahawk Ultra.
With a focus on scalability, Jericho4 incorporates 3.2 Tbps HyperPorts, each formed by aggregating four 800GE links into a single logical port. This architecture improves bandwidth utilization by up to 70%, reduces load-balancing inefficiencies, and optimizes traffic flows across distributed environments.
Extending AI Ethernet Beyond One Data Center
Broadcom’s platform introduces several critical innovations to support AI at scale. Among them is deep buffering, which enables reliable lossless Remote Direct Memory Access over Converged Ethernet (RoCE) across distances exceeding 100 kilometers. The router’s support for line-rate MACsec encryption across all ports ensures that security is not compromised even under intensive traffic loads.
“Jericho4 is built to extend AI-scale Ethernet fabrics beyond the boundaries of a single data center,” said Ram Velaga, Senior Vice President and General Manager of Broadcom’s Core Switching Group. “Its congestion-free RoCE support and advanced HyperPort capabilities offer the performance needed for distributed AI computing at massive scale.”
Manufactured using a 3nm process and incorporating Broadcom’s 200G PAM4 SerDes technology, Jericho4 delivers high efficiency while reducing the need for external components such as retimers. This contributes to lower power consumption and increased reliability across complex network topologies.
Crucially, Jericho4 is designed with interoperability in mind. It adheres fully to standards set by the Ultra Ethernet Consortium (UEC), enabling seamless integration into open, standards-based Ethernet networks. This ensures compatibility with a broad range of network interface cards (NICs), software stacks, and switching equipment, facilitating flexible deployment across heterogeneous AI infrastructures.
With the launch of Jericho4, Broadcom is positioning itself as a foundational technology provider for AI data centers facing the challenge of interconnecting massive compute resources across dispersed environments. By delivering a fabric that scales across geographic distances without compromising performance or security, Broadcom is addressing a critical bottleneck in the next era of high-performance distributed computing.
