While fibeReality is willing to double down on our wakeup call to the hyperscale data center operators on the need for them to help reconstruct an optical ecosystem for development efforts on 4x100GbE for the long term, and for 50G in the short term, it is important to take a step back on what they could actually live with — for an indefinite period of time. In the past, we have consistently not only pointed out the technical difficulties in moving to higher data rates, but cases in which we believed executives at these big private networks were disingenuous in suggesting they lacked the means of adaptability in a capacity crunch. Also, when fibeReality saw the initial unviable schemes produced by standards bodies for 400GbE, we were an early advocate for moving down to working on a more conservative 200GbE standard. Moreover, although there is an undeniable enigma associated with stranded ports on Broadcom’s Tomahawk 3 offering, which might be solved in a more perfect world at a quicker pace (the enablement of 4×100-gig), there are advantages with the chip in forcing a more significant effort in optics R&D as well as the larger switch radix allowing for a flatter network topology.
On this last point, the hyperscale folks (Microsoft has been the most outspoken about the idea) definitely want to increase the radix, resulting in a higher number of ports. As we have discussed before, they are now going up through layers of switches, from top of rack to leaf and spine, and in going vertically (north to south), more latency is being incurred.
In fact, it is quite possible that the advantages of a flatter switch structure, including saving on power utilization, combined with hooking up to less expensive optics at lower rates could be a winning arrangement for at least one of these large customers. Theoretically, it could make the stranded capacity caused by the high radix less of a burden.
Actually, we think that Facebook is moving in this exact direction. Despite its share of hype in supposedly needing to move as soon as possible to 400G, it is understood in the industry that it could get away with far less transport capacity.
At an OFC 2017 presentation by Facebook, in reading between the lines, we surmised that “100G will be probably be sufficient indefinitely.” We collected some intelligence recently that indicated it will just use higher-density 100G along with its efforts to make its switch-to-server infrastructure more efficient.
fibeReality has not found part of the explanation that we had heard Facebook is giving out on avoiding the use of 400G optics to be compelling. We believe it is saying Broadcom kind of sandbagged when it would have the Tomahawk 3 come out, and that it was going to be available earlier than the buyer anticipated.
It should be noted that Facebook has put out an RFQ for 200-gig. We are not aware of any other hyperscale entities that are officially committed to the data rate. It is possible they are either in denial on how long it will take to deliver 4x100G, or they will choose to use Google’s band aid method.
Turning to Google, over a year ago, it was not planning to roll out 400G first. It then decided to do 200G in a QSFP-DD slot, packaging two 200Gs into one module, as mentioned in our last article. While we have heard Google accused of “technically cheating” to get to 400G, it gets the firm to the desired speed, and once again, demonstrates just how much work needs to be done in achieving the four-lane goal.
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]