Broadcom on CPO/NPO: Countervailing Theories

Avatar photoPublished on
shutterstock 1931052791 copy

Last October, fibeReality wrote an article defending a proposition that Broadcom’s big expenditure push for co-packaged optics has been based on a bluff. Since that time, we have received other postulates from our contacts, one of the most intriguing being that the supplier is so involved with near package optics/CPO because of the hardships in getting to 200-gig SerDes. Naturally, there is also the kind of Machiavellian explanation that the top leadership at Broadcom has just been blinded by ambition in striving to offer a very expensive solution as it takes control of both the switch and optics TAMs (our understanding of how the ex-GM of the Optics System Division mainly sold the concept internally). Of course, another excuse is that a couple of large customers simply compelled the supplier to move in this direction. As we persist in addressing these matters, it becomes even clearer that while there may be some limited merit to the first new motivation expressed, the last two continue to end up not being very compelling. For it is our enduring contention that the ultimate arrival of OSFP-XD, despite pushback by the current GM of OSD, Near Margalit, will lead to an automatic diminution in the R&D efforts in the NPO/CPO direction (assuming the cutbacks do not occur sooner). Therefore, we remain adamant in asserting that Broadcom’s CEO, Hock Tan, who undoubtedly has a mercilessly realistic view of the marketplace, is making such otherwise irrationally high expenditures as a very costly, defensive smokescreen to help protect the company’s current strong position in the marketplace as well as its future viability. Certainly, in reviewing its investor conferences, we could find no mention of NPO/CPO since the original announcement of its silicon photonics product line in early 2021. Even more critically, the supplier’s intention to buy VMware places its NCO/CPO development efforts in the proper perspective.

We believe that the 200G SerDes rationale is a minor/indirect one at best. Advocates for this perspective will refer to very few examples of operational SerDes and that there have been numerous delays.

At least a small number of engineers in this camp will even legitimately find NPO interesting because of the relaxation of the requirements necessary in the ASIC. Nevertheless, we think that a large contingent of technicians is not anymore confident about NPO because it fails to provide any additional power or cost advantages, nor the flexibility of a pluggable.

A kind of related school of thought is the contention that, at least on the module side, Broadcom did not meet its expectations with 100G PAM4 chips. Inphi/Marvell was the big winner and Broadcom’s 7nm approach did not pan out as planned. So, supposedly now, it may be attempting to push CPO with 200G PAM4 to make up for that shortcoming.

Still, we are cognizant of a potential concession made by Broadcom regarding any hardships resulting from 200G SerDes itself. Apparently, customers have been admonished that the supplier may only offer bare die. Thus, the responsibility for dealing with the burden of the packaging losses and signal integrity would be shifted to the buyers anyway.

As fibeReality implied earlier in this piece, Broadcom will not have too much of a problem in saving face because after reviewing 14 meetings with financial analysts since January 12, 2021, we found that CPO/NPO was not brought up by the upper corporate management. In fact, on that original date, the subject was just the third part of an overall networking presentation to a particular investment house. The matter of SiPh itself was addressed only once after that time at such conferences and only as part of a general list of offerings receiving substantial funds of money.

Apparently, these solutions were not deemed important enough to give them their own separate technology evaluations to investors by Broadcom as has been done with other products in the past. Lower-level executives at the supplier have only been outspoken about CPO/NPO at optical tech shows such as at OFC and ECOC or at OIF and COBO meetings.

While at OFC 2022, there was arguably frustration demonstrated on the part of Margalit that a particular panel was not totally moving in the direction most favorable to his particular message at that time, there is definitely room for nuance on his team. A product line manager at an OFC’22 workshop made a point “of not going as far as referring to a paradigm shift, in that advancements in both pluggables and CPO are going to happen.” Exceptionally revealing was at a traditionally slow news time, Broadcom demonstrated its willingness from a direct PR perspective to hop on board of the OSFP-XD train.

It should also be noted that any pressure from the two hyperscalers supporting CPO/NPO being a truly coercive factor at Broadcom holds less water than ever. For example, Meta has evidently felt forced to put its own skin in the game. If one is being charitable, perhaps the webscaler installs a switch blending NPO and pluggable sockets together — and declares victory.

Finally, Broadcom’s desire to buy out VMWare places its R&D expenditures on CPO/NPO in the right context — exceedingly small potatoes even under the most optimistic scenarios. The really extreme, dramatic paradigm shift for the firm as a whole would happen with the successful completion of the buyout of the software-centric VMWare entity. In what appears to be some foreshadowing the month prior to the announcement, Tan stated, “start converting software, which is what we call domain specific hardware, and make them into hardware.”

While we have a great deal of confidence that pluggable modules will continue to be viable even with 100T switches, such an immense device might be changed via virtualization to a bunch of modular units. Clearly, nobody can necessarily anticipate the design of a 200T solution (perhaps constructed with many chiplets) and the stakes for Broadcom could be incredibly higher in taking over the virtual realm so aggressively, as its cost and IP advantages in the switching space may turn out to be quite substantial. Nevertheless, it will hardly be shocking, if at the end of the day, pluggables are still utilized at 200T.

To comment on this post, please click here.