“`html
CNBC AI News, August 5th – Broadcom is making waves with its latest addition to the DNX product line: the Jericho4 Ethernet fabric router. According to reports, this innovation is poised to shatter the geographic constraints of traditional data centers. The Jericho4 is engineered to interconnect distributed computing clusters across distances of 100 kilometers and beyond, providing critical infrastructure for massive AI training and inference workloads.
Jericho4’s core mission: to untangle the AI compute bottleneck. This powerhouse can securely link over a million processors (XPUs) spread across geographically dispersed data centers, effectively extending AI clusters beyond the limitations of a single facility’s physical space and power capacity. Imagine a regional synergy, unifying data center resources for seamless operation.
The cornerstone of Jericho4’s performance is its innovative HyperPort interface. Each Jericho4 system can support up to a staggering 36,000 ports, with each HyperPort delivering an aggregate bandwidth of up to 3.2 terabits per second (Tbps).
Built on Broadcom’s 3nm process and incorporating 200G PAM4 SerDes technology, HyperPort is purpose-built for long-haul, high-throughput AI data transfer by bonding four 800G channels into a single logical link.
The key advantage of HyperPort lies in its significantly enhanced link efficiency. It’s designed to overcome the inherent drawbacks of traditional multi-800G port ECMP (Equal-Cost Multi-Path) load balancing methods, such as hash collisions and uneven traffic distribution – issues that are particularly pronounced in high-volume AI deployments.
By boosting effective traffic width and improving port-level utilization, HyperPort can increase bandwidth utilization by up to 70% compared to standard 800GE solutions. This translates to faster data transfer between facilities, quicker task completion times, and the best part? No need to upgrade existing optical components or physical infrastructure.
Another significant breakthrough with Jericho4 is its extension of RDMA over Converged Ethernet (RoCE) support to distances exceeding 100 kilometers. Maintaining lossless transmission over such extended Ethernet distances has been a notorious challenge in the industry.
Broadcom has tackled this challenge by integrating deep buffering technology based on High Bandwidth Memory (HBM). This allows the router to absorb network congestion, preventing Priority Flow Control (PFC) events from spreading to adjacent data center domains. As a result, long-distance congestion impact is isolated from the local compute architecture, ensuring the stability and throughput of AI burst traffic.
Security is also paramount. Jericho4 integrates line-rate MACsec encryption on every port, ensuring the confidentiality of inter-facility traffic. Its hardware implementation guarantees that encryption, whether enabled or not, does not impact throughput performance.
Unlike proprietary interconnect solutions, Jericho4 adheres to the Ultra Ethernet Consortium (UEC) specifications, ensuring compatibility with the growing ecosystem of UEC-compliant network cards, switches, and software stacks.
HyperPort maintains standard Ethernet packet structures, simplifying integration with existing software-defined networking stacks and monitoring tools, providing operators with a unified, standards-based architecture spanning local and regional footprints. This openness mitigates vendor lock-in risks and streamlines the procurement and deployment process, making it highly appealing to customers developing long-term AI infrastructure plans.
Within Broadcom’s product portfolio, Jericho4 complements the Tomahawk series (such as Tomahawk Ultra and Tomahawk 6), which focuses on ultra-low latency, high-capacity switching within racks and facilities. Jericho4 specializes in inter-facility connectivity while maintaining the same management models and routing policies. An entire architecture based on Jericho4 can operate as a single logical router, significantly simplifying the management complexity of distributed systems.
“`
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/6418.html