Neural Networking: AI Integration into Data Center Subsystems

AI integration in the data center promises to improve data center operations considerably. The incorporation of pattern recognition and decision-making abilities allows the data center to function not dissimilarly from the human brain; strengthening the connections that lead to successful outcomes and gradually improving performance through experience.

However, AI introduces unprecedented challenges in data centers; not just in power, but in cooling, latency, deployment speed, and geography. These challenges carry through all data center areas, or “subsystems,” to varying degrees.

These different subsystems often require tailored solutions in an AI data center, removing them from any semblance of a typical “uniform” design. At the same time, there are often different people responsible for each subsystem of a data center, so proper up-front planning and communication is critical. This article will take a brief look at four key subsystems and the challenges they face with AI builds: the Entry-Point, Front-End, DC Interconnect, and Back End Subsystems.

Entry Point

The Entry-Point Subsystem is a delicate ecosystem that attempts to balance available space (which is usually minimal) with the facility’s demands; and this is where AI can inadvertently create chaos. Increasing density demands from AI applications are driving increases of high fiber count cables coming into the Entry Point, and increasing power and cooling demands have further limited available gray space in the data center. This can cause complications for space utilization, cable management, and labor, resulting in speed of deployment being the most critical issue related to the entry point.

So how can data center operators mitigate this issue? Rapid and versatile installation is key; innovative splicing enclosures, splicing and interconnect trays, and consideration of pre-terminated solutions can be vital. These solutions unlock a modular approach to infrastructure, allowing for rapid and easy upgrades as needs evolve, and require much less time to install overall.  Leviton’s innovative line of STRATA™ solutions includes built-in solutions for cable management, including entry point enclosures with the capacity to contain 6,912 fibers for splicing or interconnect.

Front End

Of the four subsystems, the Front End is least impacted by the addition of AI, buoyed above the waves that hit other subsystems by its established infrastructure. The Front-End Subsystem’s foundation has only received minor updates over the last five years. 

However, this grounded, well-established infrastructure has, over time, sprouted its own evergreen issues. Local code compliance is almost always a concern when expansions occur on the front-end, especially when those expansions are in other geographic locations. Making sure that infrastructure adheres to code can significantly slow deployment, especially if you’re working with solutions that are only available in particular regions. With time and labor at a premium, you don’t want to waste either attempting to piece together solutions from various providers into a working infrastructure.

Data Center Interconnect

Somewhat ironically, AI struggles with the capability that the DC Interconnect subsystem was designed to deliver: connection over distance. An AI cluster on the back end is treated as a single system, housed in one location that then must spread itself over the rest of the facility. While this system can be broken into “chunks” to cover larger areas, physics still apply: the greater the distance covered, the more latency the network will struggle with. As data center builders aim to expand their facilities into locations with available power, some of which may be a greater distance away from their main facility, the challenge of the DC Interconnect subsystem is to manage the problem of latency.

Hollow-core and multi-core fiber are emergent technologies that may provide relief. When compared to single-mode fiber, hollow-core can cover 2.25 times the area at the same latency footprint, making it an efficient solution for organizations looking to connect data centers at greater distances. It also offers greater flexibility in regional data center placement, potential savings in real estate and power costs, and has the potential to double your core network’s geographic coverage. Multi-core fiber, on the other hand, embeds multiple cores, typically four, inside one fiber cladding, offering high bandwidth capacity and transmitting multiple signals simultaneously. The technology can be used in tandem with Wave Division Multiplexing to exponentially improve the density of the DC Interconnect subsystem.

Back End

Finally, there’s the Back End; the subsystem that houses the AI cluster, and where AI implementation really changes the subsystem’s status quo. As the density of the system increases to accommodate the AI cluster, cable management and power become critical. The substantial power demands of AI technologies require data center operators to find ways to do more with their established power infrastructure.

The challenge of managing the increased density is one that data center operators are familiar with; doing more with less. This is where Very-Small-Form-Factor (VSFF) connectors, like , come in. VSFF connectors meet the need for higher density at connect points, while taking up less space than their MPO alternatives. Leviton’s STRATA™ family of solutions has three times the port density using MMC compared to MPO connectors, with a capacity of 3,456 fibers per rack unit with a 16-fiber connector.

As for maximizing the potential of established power infrastructure, utilizing power more efficiently in the data center often boils down to cooling. Immersion cooling is an emergent technology wherein servers are submersed in non-electrically-conductive, thermally-conductive fluids, offering almost-instantaneous heat dispersion for high-density networks. While the initial setup of immersion cooling systems can be somewhat disruptive, the same immersion infrastructure can support multiple server generations. In addition to using less power and water to achieve better results than its counterparts, immersion cooling systems are less noisy, require no server vendor lock-in, and speed the adoption of newer technologies. Leviton's line of TORRENT™ immersion-ready solutions can facilitate the transition to immersion-cooling systems, as they are designed to perform reliably in dielectric cooling environments, maintaining signal integrity and resisting degradation even when fully submerged.

Conclusion

Driving efficiency across every layer of the data center is mission critical, especially with the explosive leap in fiber density and connectivity driven by AI networks. Managing such a significant change, as well as anticipating future refresh cycles, can be daunting, to say the least.

This is why instead of treating facilities as one daunting machine, we’ve taken a more granular approach – breaking down the evolution required across four distinct subsystems, and the tailored approach involved in delivering future-ready connectivity. Adjusting the lens in this manner allows us to zoom in on the people working in each subsystem, and the actual practical impact of AI integration on each of these spaces. We believe that as teams align on priorities across departments early, the data center ecosystem will gain a shared understanding of the route ahead – making it easier to build one unified solution.  

For more insight on how AI will change the data center, as well as solutions developed specifically for AI and hyperscale data centers, visit leviton.com/ainetworks.