NVIDIA 800G InfiniBand and Ethernet Connectivity AI Cabling Guide: A road map for operators navigating the AI landscape
Nate Hefner
Published: January 9, 2025
AI requires over 10x more fiber than traditional data centers – and more fiber means more interconnects and cabling. A typical AI data center holds tens of thousands of graphics processing units (GPUs), which are essential for their parallel processing ability, performing massive amounts of complex calculations simultaneously. That means the cabling and equipment distribution area connections will look quite different from a traditional data center.
This is new territory for many data center operators – but it represents a tremendous opportunity to serve a growing demand for AI tools and AI-as-a-service applications. As next generation hardware solutions become more broadly available, there are more opportunities for new players to tap into the AI space.
At Corning, we're committed to simplifying the path to scalable AI networks. Corning and NVIDIA jointly developed a practical guide that walks through structured cabling sets and other tools designed to enable AI networks and prepare them for the future.
NVIDIA is paving the way as the leader in the AI chip market, providing powerful GPUs and setting industry standards for AI, including A100 and H100 models. The guide provides specific solutions for optimizing data centers utilizing these solutions.
Why structured cabling?
As technology continues to evolve in the age of AI, next generation tailored connectivity solutions that can be quickly manufactured and innovated are more essential than ever. Attempting to connect 10,000 GPUs with individual patch cords is messy, time-consuming, and structured cabling fills a critical gap – especially as data centers begin moving from 400G to 800G networks.
Structured cabling is a well-planned, standardized system of organizing and connecting cabling infrastructure that’s designed to adapt to changing technology needs while maintaining reliable connectivity across various devices and systems within a facility. These strategies are necessary for deploying high-performance AI/Machine Learning (ML) computing clusters, particularly those utilizing NVIDIA's DGX systems and InfiniBand technologies. Corning’s guide outlines detailed strategies for all scenarios, including server to leaf, leaf to spine, and spine to core cabling.
Additionally, Corning has created consolidated, color-coded products that break out to all the necessary ports from rack to rack within a cluster. Because it’s easier to manage, installation is up to 4x faster with cable assemblies than with traditional approaches and reduces the opportunity for human error.
Risks of outdated approaches
With an increasing demand for data, outdated cabling approaches like traditional patch cords are risky. For one, human error is much more likely with poor cable organization, whether it’s during installation or throughout ongoing maintenance and management.
And as we prepare for the future, it's challenging to scale AI networks with traditional patch cords. Errors like improper routing, mislabeled cables and inventory mistakes can be costly, and lead to network disruptions, data loss, and unnecessary downtime.
Preparing networks for 800G and beyond (the future)
Future-proofing for AI-specific workloads is a top priority for data centers, and structured cabling will only become more critical down the line. By using a standardized system, organizations can easily manage network connections and future upgrades, allowing for more scalability and flexibility within the framework. As port density increases to accommodate higher gig optics, structured cabling offers a solution to ease installation and allow for updates to enable 800G and beyond.
Corning and NVIDIA are ensuring best-in-class solutions. In particularly, Corning's EDGE8® products, which support both single-mode and multimode optical interfaces, can provide a future-proofed, high-density connectivity infrastructure.
Corning's engineering teams can provide essential guidance for data center operators during the design phase, ensuring that cabling solutions are tailored to the specific bandwidth and architectural needs of the facility.
As the AI industry continues to evolve, so will the infrastructure inside data centers. Our guide will continue to grow, incorporating new, tailored structured cabling solutions as technology evolves and use cases shift.
Download the guide here and be sure to revisit this space for future additions to the guide.
