The Impact of AI on Data Center Cabling Requirements
By Kenneth Hall
GenAI networks have changed the landscape of network implementations. Not only do they challenge power, thermal and time constraints, they also challenge the way we must approach cabling and connectivity. In a short period of time, we have seen these additive back-end networks leap from traditional duplex to one or two multipair muti-fiber push-on (MPO) connectors at each transceiver; data rates have jumped from 100G to 400G, then 800G, and this year, 1.6T, with lane rate options shortly available from 100G to 200G, as well. Performance and redundancy are critical. Optical fiber counts for many applications have increased by a factor of four and now eight. Network capacity has significantly increased, but the available working space has not scaled with it. Cabinets have grown deeper to accommodate the additional power and cooling requirements, further impacting cable management with limited pathways. Also, not every GenAI implementation occurs in an AI factory data center, reinforcing the need for an approach appropriate to the specific environment. To be clear, these developments are exciting, not bleak. It is time to clear off the whiteboard and roll up the sleeves. Pre-planning is critical. There are certain network building blocks common to most data centers which enable easy scaling, and innovative approaches are emerging in structured cabling as a result of the evolving landscape. How one prepares their designs can directly affect the efficiency and longer-term flexibility of the network. This article will provide some insight into design
and implementation considerations from practical experience that can simplify and speed up the on-site processes to help maximize the return on one’s structured cabling investment. GenAI networks use spine-leaf designs, similar to traditional data centers, but are additive purpose-built back-end networks (Figure 1). Spine-leaf provides the foundation for connecting devices throughout the network and, by design, provides redundant paths to mitigate risk from downtime. The significant jump in performance is provided by the use of GPUs for computing, which are exponentially faster than CPUs, though CPUs are still used for storage networks and in-band management of the network. The pods (i.e., cabinet rows [also referred to as scalable units (SU)]) shown in Figure 1 are composed of GPU/CPU cabinets, spine-and-leaf switch cabinets, and the cabling between them are used by hyperscale cloud and AI providers and social media platforms. These “rack-scale” cabinets and cabinet rows supporting GPUs, CPUs and spine-and-leaf switches are the building blocks that are replicated throughout the data center. GPUs for compute are connected to leaf switches for communication within the pod, and then to spine switches for routing to other pods or to core switches via super spines. Speed and redundancy are critical, so connections and media performance play a major role in designs. Each of the solid and dotted lines in Figure 1 represent cabling between devices. Depending on length, performance and application, those cables
FIGURE 1 : AI data center architecture showing connected CPU spines, GPU compute clusters, storage, and networking elements with different types of cables (copper, single-mode optical fiber, multimode optical fiber) used to connect them. Source: CommScope
can be singlemode or multimode optical fiber, duplex, or most likely multi-pair connectors for structured or point-to-point cabling, direct attach cables (DAC) and active electrical cables (AEC) for short reach point-to-point applications, or copper CAT6/6A cabling for out-of-band (OOB) management. As one may imagine, scaling out builds is costly, so reducing time to first token—or time to network availability—is imperative. Reducing on-site complexity plays a major role in determining the speed of deployment. As a general rule, anything that can be done off-site, or off the data center floor, in advance can help improve efficiency and speed up the process. This is where the building block approach is important. Some functions must be done on-site, as one would expect. Provisioning infrastructure for power, cooling, pathways and so forth must be done before network equipment is deployed. From a cabling perspective, provisioning overhead cable ladder racks or basket trays between key locations must be done on-site, early in the process. Key to this is understanding where the equipment
will be placed and what infrastructure must be ready to support those areas. In this case, the designers must know where each cabinet type will be placed and what services will need to be in place prior to equipment delivery. While this sounds like traditional data center provisioning, the rules have changed with AI requirements. With AI, and the implementation of GPUs at scale, the power and cooling requirements have jumped from legacy 10 kilowatts (kW) per cabinet in air-cooled designs to 4x, 8x and most recently, to some references specifying up to 600 kW per cabinet. Regardless of the level of the changes, the rapid progression of cooling requirements has prompted the development of a variety of cooling techniques for the GPUs. The most common seen today is direct to chip (D2C) liquid cooled. This brings plumbing into the cabinet, along with additional power to support the cooling equipment. As a direct result, GPU cabinets have grown deeper to accommodate the extra equipment and hose routing. While some cabinets have grown a bit wider, the most notable change is that the cabinets are
I
I
50
ICT TODAY
January/February/March 2026
51
Made with FlippingBook - Online catalogs