In June, this writer was honored to be a keynote speaker at EPIC’s meeting on “Singlemode vs Multimode Communication” in Kessel-Lo, Belgium. Our preliminary judgement is that we are cautiously optimistic about the prospects for Multimode Fiber (MMF) deployment growth within hyperscale data center operators. It is significantly based on the assumption that in striving to lessen latency and lower power dissipation, Microsoft’s plan to flatten its network by ultimately replacing the top-of-rack switch with a passive patch panel, and use MMF to connect the end/middle-of-row switch to a rack of wheel with servers, is adopted by all of the major hyperscale players. Obviously, such a transition would be good news to the two major, short-reach VCSEL transceiver vendors, Finisar and Foxconn Technology Group. However, given the delays that have become recently evident at the former in ramping up its 3D-sensing chip production, the stakes are somewhat higher for the firm that this opportunity in the 30-meter range comes to fruition, hopefully within two to three years.
There is no question that the immediate assumption has been that for intra-data center applications at these big network operators that Singlemode Fiber (SMF) is utilized. Yet, although the application is unclear, Google actually uses a lot of SR4 over OM3 fiber now in places in which it proves in better on cost. It also told the IEEE P802.3cm 400 Gb/s over Multimode Fiber Task Force that it desires standardization of an SR8 type of module. Alibaba, which also deploys MMF, told the standards group that it really needed SR4.2.
Concerning Microsoft’s plan, our understanding is that the use of MMF has not been fully endorsed at the high end of its networking department, yet. While there is no particular prejudice against it, and there is a favorable view of the cost/power/size of the VCSEL/MMF combination, there is resistance to the idea of being stuck with two types of fiber. In a more perfect world, the preference would be for a 32-channel DWDM duplex solution. The decisive factor for Microsoft may be that about half of its problems in the data center involves cleaning SMF connectors.
Regarding Finisar, its preoccupation with 3D-sensing, and the time it took to come up with a strategic plan with the change in leadership, has apparently put its 50Gbaud development on the back-burner. It has taken a fairly aggressive stance on 50-gig MMF using NRZ for HPC. Finisar views the sweet spot as the much cheaper/smaller (one-ninth the size) than a COBO module using MMF (can fit several hundred on a line card). Despite some suspicions in the marketplace to the contrary, the supplier also claims it has solved the reliability issue at high temperatures by improving the design. Certainly, the NRZ makes sense because the PAM4 is more sensitive to noise and reflections. Even the need for some kind of equalization on the receive side with NRZ is not a bad tradeoff.
We are starting to really wonder if Finisar’s heart was originally in the 3D-sensing space in the beginning. Over a year-and-a-half ago, we expressed doubts about the attractive nature of the space, and Broadcom’s decision not to enter the sector was further confirmation. Finisar’s former CEO did not make a secret of the challenges in producing six-inch VCSEL wafers. It appears quite likely that the vendor just could not pass up on the big Apple PR maneuver.
While there seems to be little doubt that Finisar is all-in now in the 3D space, we are picking up that there might be some problems with morale. In fairness, anytime there is a transition to a new CEO, especially one who was in charge for 30 years, there is going to be some disenchantment. Clearly, it must be disappointing that the experience gained by Apple’s engineers in working with Lumentum/Win Semiconductors, was apparently not enough to make a difference in staying on schedule with the production of the new VCSEL bare die.
In addition, it is notable that the Director, New Product Development at Finisar, who had been with the VCSEL group since the 1990s, left the company last month. Of course, it may just be the case that he was attracted to the challenge of heading a new startup.
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 company blog, which is exclusively on fibeReality’s LinkedIn page, please click here.
[written by Mark Lutkowitz]