The rise of AI has brought about a connectivity boom that’s reshaping network architectures. High-speed, low-latency networks are essential for supporting the increasing volume and complexity of AI technology in all its forms.

A critical component driving those workloads is an interconnected system of fibers that must multiply to handle higher densities, while scaling down in size to create compact, efficient pathways throughout the data center. The way in which density and power is managed inside AI-ready facilities is changing – and fiber technology is at the fore.

Against this backdrop, in a recent DCD>Broadcast, Brian Roney, director of data center AI market development at Corning shared his insights on the surge in demand for AI deployments, and the unique connectivity challenges they present.

A second network

The demand for connectivity has led to the addition of a back end network. This ‘second network’ combines GPUs into clusters, and allows them to communicate with one another inside the data center – with traditional cable plans for supporting CPUs still intact.

“What that’s done is created a huge lift in the amount of connectivity and fibers that are needed to be fed into a server rack,” explains Roney. “We used to talk about 10 or 20 fibers to a rack switch in a traditional non-AI application. Now, we're talking about tens of thousands in a single switch rack.”

Today, a huge range of AI applications and their hefty workloads are driving a need for parallel processing. Generative AI, for instance, requires a massive number of calculations and large data sets for its training models – that's a lot of load and time for just one device. As a result, that activity needs to be spread out amongst different compute nodes across the data center.

“That division of workload requires synchronization and large bandwidths of connectivity between nodes,” adds Roney. “It's like trying to solve a complex problem, and asking someone else for help with part of the task. Nodes in the data center are like brains – the more we have and the bigger they are, the better.”

The need for purpose-built facilities

Restructuring around these increasingly complex workloads calls into question the suitability of existing infrastructure – can brownfield and legacy data centers support the connectivity levels that AI demands? Roney shares his thoughts:

“It’s definitely a mixed bag, and we have retrofitted even brand new facilities, plus existing multi-tenant data centers. But increasingly, we're seeing the need for purpose-built AI facilities with power, cooling, and extensive connectivity capacity built-in to the initial plans. And that’s enabling us to build those large clusters.”

Taking it back to basics, one of the fundamental drivers behind rethinking the data center for the AI age, is rising power demands. Today, racks with hundreds of kilowatts are the new normal, and the industry is staring down the barrel of the 1MW rack. With those unprecedented power levels, comes extreme heat outputs that also need to be managed.

Retrofitting existing facilities that were not tailor-made to handle these energy and cooling demands, can be a colossal undertaking. “Coupled with a shift towards next-generation GPUs built off of direct liquid chip cooling, it presents a whole mechanical uplift for an existing site,” adds Roney. “Supporting that shift will require specialized, purpose-built data centers.”

Density, speed, scale, and customization

Marrying its history in glass innovation that has evolved alongside the AI era, Corning’s answer to these key connectivity challenges lies in GlassWorks AI – a customized portfolio of solutions and services designed to help operators navigate the network-design challenges of AI.

“GlassWorks AI was built based on four key tenants – the first being density. With multiple density vectors at play inside data centers, we’re delivering an ultra-dense fiber cable and connector solution,” says Roney.

With millions of fibers moving around the data center, they require an efficient pathway, without taking up masses of space. Corning is rethinking cabling with its Contour Flow solution – utilizing smaller fibers, for smaller cables with reduced bend loss. Joining those cables up with compact connectors maximizes space-saving while, crucially, increasing density and connectivity.

“Second is speed of installation – that’s about how quickly we can deploy the cabling for these GPU racks, to accelerate ROI, and start supporting these workloads as soon as possible,” says Roney.

As key players race to keep pace with demand in the AI era, the phrase time is money is more relevant across the industry today than ever before. The ability to take GPUs and turn them into training models, for instance, must be delivered at speed in order to stay competitive.

Corning is delivering that need for speed with full rack integration. Building a fully-assembled off-site solution means that when it arrives at the facility, only around five connectors per rack need to be plugged in, rather than thousands, before operations are up and running.

“Third is scale. Building bigger and bigger clusters, demands higher fiber counts. That means some unique shuffling of fibers in some of the cable assemblies to create a scale-out network,” says Roney. “And finally, everybody has unique needs, so the fourth tenant is all about customization. We’re customizing at scale, based on real-life customer requirements.”

Whether it's the fiber itself, leveraging that fiber inside a cable, or deploying the full solution with fiber cable and connectivity to build clusters, GlassWorks AI is providing a tailored, one-stop-shop for hyperscalers looking to scale up for high-density workloads.

“I like to say we're the plumbers of light – we’re networking around in the data center and all this fiber is really a guiding light between all the devices,” says Roney.

Driving connectivity

As the amount of connectivity, bandwidth, and power consumed by the network rises – it’s increasingly vital to spend that power wisely. Feeding GPUs and compute, rather than pushing data around the facility, is essential for effective operations and output.

Co-packaged optics (CPO) is transforming the data center as we know it by integrating optical transceivers directly onto the same package as the switch's application-specific integrated circuit (ASIC). This integration offers improved bandwidth, energy efficiency, and latency.

Power savings equal cost savings, too. “It’s about economics, and trying to continually drive down optical cost per bit of data as well as power,” explains Roney.

At a time when operators are being pushed to locate more and more power, co-packed optics is supporting intelligent power management, reducing consumption, while increasing that all important connectivity.

Another key piece of the puzzle for driving connectivity is expanding networks. As the industry grows, it must extend beyond traditional hubs and into new, developing markets where power, space and land are available.

“As these new sites emerge, they have to be connected to the long haul network – and we need more long haul capacity to do that,” says Roney. “Late last year we announced an agreement between Corning and Lumen that dedicates 10 percent of our fiber capacity to building out that long haul network.”

That process involves overlaying or augmenting existing pathways, to connect with new hub cities and support the extension of the data center footprint in new areas.

With more operators getting connected across the globe – ensuring best practice and driving success far and wide requires a standardized roadmap. Working alongside Nvidia and other key technology partners, Corning is creating reference guides that offer practical support to customers. Roney explains:

“We’re marrying up logical reference architecture with a physical ‘how to’ cable guide, so that those ideas can be achieved in the physical space. We continue to work with those partners to create these guides, and we find them very valuable when talking to our customers – they want that guide to help them, and we're here to do it.”

The path ahead

As GPUs rack up, data center architecture and operations must evolve in response, and fast. For Corning, solving the density challenge is the driving force behind its continued efforts to develop new fiber innovations that increase connectivity and manage power, in a compact, efficient solution that suits the needs of modern data centers.

Working with key industry partners and technology vendors is an essential part of the mix, propelling the industry forward as a whole, and delivering essential future-ready solutions to market quickly and easily.

“GlassWorks AI is an evolving roadmap. We continue to work behind the scenes and workshop innovative glass solutions,” concludes Roney. “It's a never ending cycle – and we intend to be at the forefront.”

Optimizing connectivity infrastructure inside facilities and across the data center ecosystem, is an essential aspect of industry-wide efforts to support the rise of AI and the opportunities the technology has to offer. To stay ahead, operators must stay connected – both inside facilities and across the supply chain.

>> Learn more about Corning’s GlassWorks AI program here! Remember – the best time to reach out to Corning about your AI build is now! And you can check out the full DCD> broadcast ‘Through the looking GlassWorks’ here.