Cisco AI Router Tackles Data Center Interconnect Challenge

Cisco is stepping into the increasingly crucial arena of AI data center interconnect technology, unveiling a purpose-built routing system aimed at seamlessly connecting distributed AI workloads across geographically separated facilities. This move positions Cisco as a major contender alongside Broadcom and Nvidia in a market poised for explosive growth.

On October 8th, the networking giant announced its 8223 routing system, touting it as the industry’s first fixed router capable of processing 51.2 terabits per second (Tbps) – specifically designed to link data centers actively running demanding AI applications. This bandwidth capacity is critical for minimizing latency and maximizing throughput in distributed AI environments.

The heart of the 8223 lies in the new Silicon One P200 chip, a custom-designed ASIC that represents Cisco’s solution to a fundamental challenge facing the AI industry: overcoming the limitations imposed by the physical constraints of single data centers. This challenge stems from the insatiable demand for computational resources inherent in large-scale AI model training and inference.

A Three-Way Race for Scale-Across Supremacy?

The competition for dominance in this emerging space is intensifying. Broadcom initiated the contest in mid-August with its “Jericho 4” StrataDNX switch/router chips. These chips, now sampling, also boast 51.2 Tbps of aggregate bandwidth and incorporate High Bandwidth Memory (HBM) for deep packet buffering, a key feature for managing congestion during peak AI workload activity.

Following Broadcom’s announcement, Nvidia unveiled its Spectrum-XGS scale-across network. While Nvidia secured CoreWeave as an early adopter, detailed technical specifications of the Spectrum-XGS ASICs remain somewhat limited. Cisco’s entry with the 8223 system solidifies a three-way rivalry among these networking powerhouses.

The Problem: AI is Too Big for One Building

The surging demand for AI infrastructure stems from the computational intensity required to train large language models (LLMs) and operate sophisticated AI systems. These tasks demand thousands of high-powered processors working in unison, which generates massive amounts of heat and consumes substantial power.

Data centers are increasingly facing hard limits – not just on available space, but also on the amount of power they can supply and the effectiveness of their cooling systems. The exponential growth of AI workloads is quickly exceeding the capacity of even the largest, most advanced data centers.

“AI compute is outgrowing the capacity of even the largest data center, driving the need for reliable, secure connection of data centers hundreds of miles apart,” stated Martin Lund, Executive Vice President of Cisco’s Common Hardware Group. This highlights the critical impetus behind the development of scale-across solutions.

Historically, capacity challenges have been addressed through scaling up (adding more resources to individual systems) or scaling out (connecting more systems within a single facility). However, both approaches are reaching their practical limits. Physical space is becoming scarce, power grids are struggling to provide sufficient electricity, and cooling systems are unable to efficiently dissipate the heat generated by increasingly dense computing infrastructure.

This necessitates a shift towards a third approach: “scale-across,” which involves distributing AI workloads across multiple data centers, potentially located in different cities or states. This geographic distribution introduces a new challenge: the connections between these facilities become critical bottlenecks if not properly designed and optimized.

Why Traditional Routers Fall Short

AI workloads exhibit traffic patterns that differ significantly from typical data center traffic. Training runs, in particular, generate massive, bursty traffic – periods of intense data transfer followed by relative calm. Traditional routing equipment, optimized for more uniform data flows, struggles to effectively handle these surges. If the network connecting data centers cannot accommodate these bursts, performance degrades, leading to underutilization of expensive computing resources and, critically, wasted time and money.

Traditional routers prioritize either raw speed or advanced traffic management features but often struggle to deliver both simultaneously while maintaining reasonable power consumption. AI data center interconnect applications demand all three: high speed, intelligent buffering to absorb traffic spikes, and energy efficiency to minimize operational costs.

Cisco’s Answer: The 8223 System

Cisco’s 8223 system represents a departure from general-purpose routing equipment. Housed in a compact three-rack-unit (3RU) chassis, it offers 64 ports of 800-gigabit connectivity – currently representing the highest density available in a fixed routing system. Its key performance metrics include the ability to process over 20 billion packets per second and scale up to three Exabytes per second of interconnect bandwidth.

However, the system’s distinguishing feature is its deep buffering capability, enabled by the P200 chip. These buffers act as temporary holding areas for data, akin to a reservoir that captures water during heavy rainfall. When AI training generates traffic surges, the 8223’s buffers absorb the spike, preventing network congestion that would otherwise cause expensive GPU clusters to sit idle, waiting for data.

Power efficiency is another critical advantage. The 8223 achieves what Cisco describes as “switch-like power efficiency” while retaining routing capabilities – a crucial consideration for data centers facing increasing power constraints. This efficiency is achieved, in part, through advancements in the P200’s architecture and manufacturing process.

The system also supports 800G coherent optics, enabling high-bandwidth connections spanning up to 1,000 kilometers between facilities. This long-reach connectivity is essential for geographically distributed AI infrastructure, allowing for the creation of large-scale AI clusters across multiple sites.

Industry Adoption and Real-World Applications

Several major hyperscalers have already begun deploying and evaluating the technology. Microsoft, an early adopter of Cisco’s Silicon One architecture, has found it valuable across a range of use cases.

Dave Maltz, Technical Fellow and Corporate Vice President of Azure Networking at Microsoft, stated that “the common ASIC architecture has made it easier for us to expand from our initial use cases to multiple roles in DC, WAN, and AI/ML environments.” This indicates that the Silicon One provides architectural consistency that simplifies management and deployment across different network segments.

Alibaba Cloud intends to use the P200 as a foundation for expanding its eCore architecture. Dennis Cai, Vice President and Head of Network Infrastructure at Alibaba Cloud, commented that the chip “will enable us to extend into the Core network, replacing traditional chassis-based routers with a cluster of P200-powered devices.” This suggests a move towards a more disaggregated and scalable network infrastructure.

Lumen Technologies is also exploring how the technology can fit into its network infrastructure plans. Dave Ward, Chief Technology Officer and Product Officer at Lumen, said the company is “exploring how the new Cisco 8223 technology may fit into our plans to enhance network performance and roll out superior services to our customers.” This highlights the broader interest from service providers in leveraging these technologies to improve network performance and offer advanced services.

Programmability: Future-Proofing the Investment

Adaptability is a crucial, often overlooked aspect of AI data center interconnect infrastructure. AI networking requirements are evolving rapidly, with new protocols and standards constantly emerging. Traditional hardware typically requires replacement or costly upgrades to support these new capabilities.

The P200’s programmability addresses this challenge. The silicon can be updated to support emerging protocols without requiring hardware replacement. This is significant considering that individual routing systems represent substantial capital investments and that AI networking standards are still in a state of flux. The programmability of the P200 is a key feature that allows customers to avoid costly and disruptive hardware upgrades when new standards are introduced.

Security Considerations

Connecting data centers hundreds of miles apart introduces significant security challenges. The 8223 incorporates line-rate encryption using post-quantum resilient algorithms, which addresses concerns about future threats posed by quantum computing. The integration with Cisco’s observability platforms provides detailed network monitoring to quickly identify and resolve issues. This comprehensive security posture is critical for maintaining the confidentiality and integrity of data traversing geographically distributed AI infrastructure.

Can Cisco Compete?

With Broadcom and Nvidia already vying for market share in the scale-across networking space, Cisco faces established competition. However, the company has several distinct advantages: a long-standing presence in enterprise and service provider networks, a mature Silicon One portfolio launched in 2019 (demonstrating a commitment to custom silicon development), and existing relationships with major hyperscalers already deploying its technology.

The 8223 initially ships with open-source SONiC support, with IOS XR planned for future availability. The P200 will be available across multiple platform types, including modular systems and the Nexus portfolio. This flexibility in deployment options could prove to be a decisive factor as organizations seek to avoid vendor lock-in while building out distributed AI infrastructure. Customers will appreciate the flexibility of choosing between different operating systems based on their individual needs.

Whether Cisco’s approach becomes the industry standard for AI data center interconnect remains to be seen. However, the fundamental problem all three vendors are addressing – efficiently connecting distributed AI infrastructure – will only become more pressing as AI systems expand beyond the limitations of a single facility. The increasing demand for AI will continue to drive innovation in data center interconnect.

The ultimate victor may not be determined solely by technical specifications, but rather by the vendor that can deliver the most comprehensive ecosystem of software, support, and integration capabilities surrounding their silicon. Ecosystem strength is becoming increasingly important across the technology sector, and AI data center interconnect is no exception.