TLDR: Broadcom has commenced shipping its new Jericho4 Ethernet fabric router, a next-generation AI chip designed to provide ultra-fast, scalable connectivity for large AI data centers. This chip is crucial for overcoming the limitations of single data center facilities by enabling distributed AI computing across vast distances, supporting massive AI model training, and enhancing overall AI infrastructure.
Broadcom, a leading semiconductor company, has officially begun shipping its groundbreaking Jericho4 Ethernet fabric router, a pivotal development set to redefine connectivity within the burgeoning artificial intelligence (AI) data center landscape. Unveiled at the 2025 OCP APAC Summit, this next-generation AI chip is engineered to facilitate ultra-fast and highly scalable connections between vast clusters of AI accelerators, such as GPUs and XPUs, across geographically dispersed data centers.
The Jericho4 is designed to scale beyond the confines of a single facility, capable of interconnecting over one million XPUs. This capability is critical as modern AI workloads demand unprecedented levels of bandwidth and low latency, pushing the limits of traditional centralized data center architectures. The increasing power consumption and physical footprint required by AI infrastructure are compelling cloud providers to distribute their systems across multiple locations, making Jericho4’s long-distance connectivity a commercially vital innovation.
Technically, the Jericho4 is a marvel of modern chip design. It is manufactured using an advanced 3nm process, ensuring high efficiency and performance. The chip incorporates several key features to optimize network traffic and security, including deep buffering, MACsec encryption, and support for RoCE (RDMA over Converged Ethernet) transport over distances exceeding 100 kilometers (approximately 60 miles). Furthermore, it integrates Broadcom’s 3.2T HyperPort technology, which consolidates multiple high-speed links into a single logical port, and is compliant with Ultra Ethernet Consortium standards, ensuring broad interoperability with other networking devices.
Ram Velaga, Senior Vice President and General Manager of Broadcom’s Core Switching Group, highlighted the chip’s immense scalability, noting that a single system can encompass roughly 4,500 Jericho chips. To mitigate potential network congestion, the Jericho4 chips leverage high-bandwidth memory (HBM), a technology also utilized by leading AI processor designers like Nvidia and AMD.
Also Read:
- Micron Technology Identified as a Top AI Stock Amid Surging Demand
- AI Fuels Semiconductor Investment Surge, Yet Disparities Emerge in Global Market
Analysts view the Jericho4 as a critical upgrade for AI infrastructure scaling, arriving amidst intense competition in the AI hardware sector. Its ability to maintain lossless performance over long distances transforms multiple data centers into a single, unified computing resource, addressing a fundamental scaling challenge in AI development. This technological advancement positions Broadcom at the forefront of an emerging global competitive advantage in AI capabilities, as nations and companies worldwide invest heavily in data center infrastructure to secure their technological independence and leadership in the AI race.
