Assuming Microsoft goes ahead with a proposed project of a flattened switch architecture in its data centers, the use of Active Optical Cables (AOCs) for the 0-to-30-meter lengths would undoubtedly remain under consideration, even though the operator is leaning towards the use of VCSELs. However, we believe there would definitely be obstacles to the utilization of AOCs in this application, which would probably make them inherently unattractive. It is interesting to note that from an installation perspective, as opposed to an inventory standpoint, Microsoft does not make a distinction between MultiMode Fiber (MMF)/VCSELs and AOCs. The deployment of either would be as if it were doing so with passive copper cable. In this article, we will also be looking at the future deployment of AOCs in general.
One hurdle with AOCs in the 30-meter-and-under app, which would likely make Microsoft unhappy, is that there would apparently need to be some kind of clever cable management solution under the control of the operator. Another drawback is that the AOCs would be hardwired from the source to the destination. So, there would be a certain rigidity with a passive patch panel in between, which is part of Microsoft’s “wish list” (such an inclination to include as many desires as possible appears unique to this hyperscaler in, once again, acting to some degree more like a traditional ISP).
With multiple hundreds of AOCs, it probably becomes cumbersome to pre-build a part or a set of racks, which would be the goal of Microsoft. With its approach of rolling in the self-contained racks, it may be tested, at least to some extent, at the optical interface, which exits somewhere at the top or the middle of the rack.
Thus, this is one case in which along with AOCs, Microsoft does not want pigtails, because they have to be handled as they dangle. Of course, the potential downside is that the extra interface there adds loss and reflections, so one does not know whether it is really going to work in the long run.
In looking at the down-the-road apps, when the industry goes to 100-gig electrical, there is a good number of people contending that there will be loads of AOCs because they just do not see passive copper working, despite it being ultra-cheap. While active copper could be employed, we have recently discussed (as noted above) the resistance by hyperscalers in having too many part types. Microsoft itself has stated when it reaches 400G networks, and the server connections go up in speed, it does not see copper playing much of a role in that server-to-first switch connection, unlike today where passive copper is pervasive.
In the meantime, the IEEE standards group for 100G electrical has been adopting objectives and approaches identical to the previous project, the 50-gig electrical – so, still PAM4, but just double the baud rate. The electrical channel loss at the higher data rate is even higher, and the idea to use even better PCB materials, has resulted in companies, such as Cisco Systems, being alarmed by the expense getting even higher.
There seems to be about an even split at the group between those individuals with a high confidence level in doing business as usual, because it was able to pull it off before, and about the other half of the folks believing other options should be considered. Alternatives offered include looking at more active elements, including optics. The idea is to perhaps cut the channel loss down resulting in somewhat shorter reaches on PCB traces, and use flyover copper cables, AOCs, etc.
Some engineers are uncertain on the use of AOCs, such as in the future, running at 400G (100G/lambda) over a short-reach, 100 meters of MMF, because the bundles on the ends are really big, including the QSFP-DDs and OSFPs. Once they start routing them through cable trays, it gets more painful, and so their preference is for a cable. It will come down to whether the difference in price is big enough for the AOCs compared with the labor costs of pulling the modules through the racks.
As always, fibeReality does not recommend any securities, and this writer does not invest in any companies being analyzed by us.
To follow us on our totally separate, quick update company blog, which is exclusively on fibeReality’s LinkedIn page, please click here.
[written by Mark Lutkowitz]
SHARE THIS