deeper—now at least four feet deep. This impact of that depth is significant to cabling for several reasons. Typically, in data centers, trunk cabling between locations enters the cabinet from above and is routed to the rear of a patch panel and then connected to provide transition to equipment cord access at the front faceplate of the panel. In a GPU cabinet, inbound and outbound hoses and power equipment now block access to the rear of the front panel, for both physical space and safety reasons. Also, human arms are generally not over four feet long, preventing the direct plugging to the rear of the panel faceplate if there was access. For GPU cabinets, cabling and cable management need to be front-facing and accessible for installation and operation. For today’s switch applications, there is more working room since they do not need the same obstructive cooling apparatus. There is still a deep cabinet to match the footprint, but there is space for installation and operations activities. Within the pod is the highest concentration of cabling and connectivity. While there are a variety of OEM-prescribed architectures, there are a few common elements with the cabinet cabling. Firstly, significantly more optical fiber connectors are used today in these applications, with the bulk of them being MPO connectors. Most common is the MPO8, MPO12 or MPO12/8 (referenced to indicate the MPO12 global industry standard interface, as well as the usage of eight of the 12 optical fiber positions) for applications. Applications within the pod may use a single MPO at the transceiver or a twin-port MPO transceiver in the same octal small form-factor pluggable (OSFP) or quad small form-factor pluggable double density (QSFP-DD) port. Whether multimode or singlemode, the eight-fiber connector most often has an angled physical contact (APC) endface, which is commonly used for AI transceiver connections and is required by systems OEMs. This endface, even for multimode, is required to minimize return loss that can be caused by light reflected from dust, dirt or imperfections on the optical fiber endface. Endface requirements should be verified with the transceiver provider before implementing to avoid challenges on-site. Secondly, test equipment is available today that can
test, verify and document performance for multi-pair connectors, such as MPO, in a matter of seconds. These can save significant time on-site compared to duplex testing. From a practical perspective, it is highly recommended that one defines testing requirements, including reference cords necessary for the application requirements and ordering materials well in advance of an installation. CONNECTION OPTIONS From the cabling perspective, structured cabling is recommended across the board to address any installation and provide flexibility, access and multi-generational network support with passive cabling. With AI builds, a pod installation will likely involve a mix of cabling types. While DAC, AEC, and active optical cables (AOC) are typically application- and speed-specific, there are use cases that should be considered within the pod (Figure 2). DAC and AEC are coaxial point-to-point cables with limited reach. DAC cables are typically less than 2 meters (m) and AEC are less than 7 m. Generally, the longer they are, the thicker and stiffer the cable due to the extra shielding required to support applications. Cabling within a cabinet, cabling to adjacent cabinet(s) or across the row are applications that
may incorporate these options from the OEM reference architectures. Many of these cables are protocol-, speed- and application-specific, and must eventually be replaced with upgrades. For short-reach applications within a cabinet, DAC or AEC cables may be used if preferred by customers. The value points for this selection are the low power consumption of these application-specific cables, as well as the fact that they are typically a less expensive option for short reach. Applications such as leaf-switch-to-spine-switch in adjacent cabinets is a common use case, as the cable’s reach is short, low-power, and inexpensive. While they do support these connections, they can still be physically challenging to manage. AOC assemblies are options but typically are not recommended by OEMs for higher data rates within a pod or beyond a pod in AI data centers. The rationale for that guidance is that when routing the cables, the connector ends with built-in transceivers and can be susceptible to damage. The connector weight—combined with small outside diameter (OD) optical fiber cabling—can pose a risk of damage or optical fiber breakage when pulling through pathways and cabinet openings. These are single- purpose cables that are generally replaced with system upgrades. As noted, there are options from a cabling perspective. Timing is critical and anything that can be done in advance, off-site or off the data center floor, can save time and improve deployment efficiency. To save time on-site, cables preset for the pod can be bundled, labeled, even color-coded according to cabinet, application, or other repeated requirements. This improves on-site efficiency with visual cues for field technicians. Another process that can improve deployment efficiency is that, when trunk cables are in place on-site, a cabinet can be built, cabled and internally configured and commissioned off-site with a top-of-rack optical fiber patch panel. It can then be delivered to the site and rolled into position, where staff can dress down and connect the trunk cables to the patch panel. The availability of transceivers can play a major role in media and architectural options. Industry shortages
can cause a delay or force the re-architecture of an installation. Projects designed around a high-demand transceiver type may be able to make a design change, for example, from multimode to singlemode, 4-pair to duplex. It is important to work with suppliers early to secure materials needed. Optical fiber cable count choices should be based on covering the number of connectors needed, as well as spares when possible. Pulling additional cable on day 2 should be avoided. Using common optical fiber counts like 144f and 288f can help provide more reliable market availability when needed. As referenced earlier, optical fiber counts have grown significantly to support AI developments. This increased bulk—especially for trunk cables between network segments—can make pathways and spaces challenging to manage. One answer to this challenge is using preterminated cable assemblies from factory production lines to help provide options for site efficiency with documented performance. Recent connectivity innovations for these assemblies include very small form factor (VSFF) multipair connectors, which deliver the same optical fiber counts as MPO connectors but in roughly a third to half of the space required for traditional MPO connectors. With similar performance to MPO connector options, the reduced size makes VSFF attractive for trunk applications, as they can be grouped or ganged to provide more optical fibers to panels. Transition to MPO connectors and equipment cords for transceiver usage is then managed from the panel faceplate. At this stage, transceivers use common, readily available MPO interfaces, so transition would be necessary to connect to equipment. CONCLUSION As standards and capabilities continue to evolve, so will the need for agile, flexible and scalable optical fiber and copper connectivity solutions in AI data centers. Dealing with the realities of new deployments requires thinking that not only addresses standards and speeds, but also floor space, man-hours and tolerance for downtime. AI data centers looking at scaling up or scaling out
FIGURE 2 : Different types of optical fiber breakout cables, specifically DAC/AEC and AOC assemblies. Source: CommScope
I
I
52
ICT TODAY
January/February/March 2026
53
Made with FlippingBook - Online catalogs