Broadcom unveiled Jericho3-AI, high-radix Ethernet switch silicon designed for AI cluster intra-cluster networking.

Broadcom Inc. positions Jericho3-AI as a high-radix, AI-optimized switch silicon designed to build massive Ethernet fabrics for training and inference clusters. The chip targets configurations with hundreds to thousands of GPUs connected at 800G per port across large-scale data centers. In presentations, Broadcom networking head Ram Velaga describes Jericho3-AI as part of a broader strategy to make Ethernet a first-class citizen for AI workloads, not just storage and general cloud traffic. He points to dual-pillar designs using Jericho3-AI and Tomahawk switch families to give hyperscalers flexibility in how they build their fabrics.

Jericho3-AI's design centers on supporting very large fabrics with many thousands of endpoints while keeping tail latency for AI traffic in check. The chip is built for high-radix switching, meaning many ports per device, to reduce the number of hops between GPUs and to simplify topology choices for operators. Broadcom explains in technical material that Jericho3-AI incorporates congestion control mechanisms tuned for AI collective operations, such as all-reduce and broadcast patterns common in training. These mechanisms aim to keep queues balanced and avoid long tail delays, which slow down large-scale jobs even when average latency looks fine.

Broadcom spends visible effort arguing that Ethernet fabrics powered by Jericho3-AI can match or approach the performance of proprietary high-performance computing interconnects, while preserving ecosystem flexibility. The company highlights standard-based tooling, familiar operations models, and interoperability with a broad range of NICs and accelerators. At industry conferences, Velaga and other Broadcom engineers present benchmark results and topology diagrams that pit Jericho3-AI Ethernet designs against alternatives, often showing leaf-spine layouts with mixed Jericho and Tomahawk switches.

Jericho3-AI does not appear on retail shelves. Instead, it surfaces inside systems and line cards that hyperscalers and OEMs sell or deploy. Customers integrate the ASIC into fixed-configuration switches or modular chassis, exposing 800G or similar port speeds to connect modern GPU servers. The typical buyer is a cloud operator, telecom, or large enterprise building AI clusters—organizations that care about metrics like job completion time, fabric utilization, and failure containment as much as raw bandwidth.

For Broadcom, Jericho3-AI slots into its broader switching and routing product stack, which includes Tomahawk for high-performance data center switching and other lines for service provider networks. The AI-focused positioning reflects a wider push by Broadcom leadership to capture growing capital spending on AI infrastructure. Broadcom Inc. stock trades on Nasdaq, and the Jericho3-AI line contributes to the networking segment that investors watch for AI-related growth impulses.