Broadcom Networking Chip Connects Data Centers to High-Speed AI

Broadcom’s silicon division has unveiled its next-generation networking chip, Jericho4, engineered to connect data centres separated by more than 96 kilometres and to accelerate AI workloads.

Jericho4 brings several enhancements focused on moving greater volumes of traffic across large-scale networks that span multiple facilities. As AI models grow larger and more compute-intensive, training and inference increasingly demand thousands of graphics processing units (GPUs) working together. Cloud providers and hyperscalers require higher-performance networking silicon to keep data flowing between GPUs, servers, and racks without bottlenecks.

Transferring data across physical sites also elevates security risks, since information is exposed when it crosses facility boundaries. Broadcom says Jericho4 addresses this by offering full-speed encryption on every port, maintaining wire-rate performance even under heavy traffic loads.

Ram Velaga, senior vice president and general manager of Broadcom’s Core Switching Group, said the chips are built for massive-scale deployments. An example configuration could include about 4,500 Jericho4 chips in a single system. To reduce packet loss and smooth traffic during congestion, the chip integrates high-bandwidth memory (HBM) — the same memory technology used in modern AI accelerators from major vendors — so the switch can buffer traffic until network conditions improve.

“The switch is actually holding that traffic (in memory) till the congestion frees up,” Velaga explained. “It means you need to have a lot of memory on the chip.” Longer inter-facility distances further increase the need for onboard buffering.

Manufactured on TSMC’s 3-nanometre process, Jericho4 scales to systems that support up to 36,000 HyperPorts, each capable of 3.2 terabits per second. The chip delivers deep buffering, line-rate MACsec encryption, and RDMA over Converged Ethernet (RoCE) that can operate across distances beyond 100 kilometres, enabling lossless transport for distributed workloads.

Jericho4’s HyperPort technology aggregates four 800-gigabit Ethernet lanes into a single logical port, reducing inefficiencies from load balancing and improving utilization by up to about 70%. This aggregation simplifies traffic management in very large networks. Combined with extensive buffering and sophisticated congestion control, the chip enables lossless RoCE over extended distances, making it feasible to build distributed AI clusters that are not constrained by the power or space limits of a single data centre.

Security is built into the architecture: MACsec encryption runs at line rate on every port, ensuring inter-site traffic is protected without sacrificing throughput.

The chip also incorporates Broadcom’s 200G PAM4 SerDes technology for long-reach links without additional components such as retimers. This integration reduces power consumption, lowers total system cost, and improves reliability. Jericho4 meets Ultra Ethernet Consortium (UEC) specifications to ensure interoperability with other UEC-compliant network interface cards, switches, and host software.

Jericho4 expands Broadcom’s networking portfolio alongside products like the Tomahawk 6 and Tomahawk Ultra, targeting the evolving needs of cloud operators and enterprise networks.

“As AI models grow in size and complexity, the infrastructure requirements exceed the power and physical limits of a single data centre,” Broadcom said. “Distributing XPUs (diverse processor types) across multiple facilities — each provisioned with tens to hundreds of megawatts — requires a new class of router optimized for very high-bandwidth, secure, and lossless regional transport.”

On the day of the announcement, Broadcom’s stock rose 3.2%, closing at 297.72.

(Photo by Compare Fibre)

See also: Broadcom’s latest Ethernet switch tackles AI demands on networks