Co-Packaged Optics (CPO) is a serious risk to the well-being of the photonics industry, which can ill-afford the expenditure of such valuable resources. The concept is being heavily pushed by a very small minority of engineers, principally within Microsoft and Facebook – in fibeReality’s opinion, for its own self-glory, as this group has a lot more to prove than the large majority of product and network designers, who are against it or non-committal. The technology is simply not in the ballpark of being ready, most specifically for intra-data center usage, and pluggable optics will continue to thrive for quite a long time. During a four-hour, interactive presentation on 5G X-haul optics, when the subject of pluggables came up, this writer implored the audience to keep their spending on CPO to the absolute minimum, while they continue to provide the necessary lip service to the cause, especially if it would otherwise mean a loss of business at the two hyperscalers. It has been gratifying to see examples of engineering folks openly being critical of aspects of CPO recently, despite the high likelihood of getting bullied later. The main aspect to consider is that sheer desire and force of will cannot speed up the historically snail’s pace of substituting gradually shorter electrical links with optical connections in just getting to the current one-meter point.
In fact, with data rate increases, optics had to be applied in novel ways with each transition. There was never just a wholesale replacement of wires or coax cables at the lower speeds. Yet, illogically, with CPO, there is an expectation of an ASIC I/O, despite an over-constrained optical solution, when such constricts were avoided in the past.
One day, there will be the ability of putting an optical I/O in the same package as a switch ASIC. However, it will be so far out into the distant future as to be extraneous as a priority to current R&D efforts for optical suppliers.
This past December, fibeReality presented, “Co-Packaged Optics Pipe Dream, in which we discussed the following eight points:
1. collision of high-reliable ASICs with low-reliable lasers;
2. elimination of pay-as-you-go economics;
3. relatively slow deployment of 25.6T switches;
4. switch ASIC suppliers are incentivized to design at highest power;
5. CPO would not be enough by itself to reduce power sufficiently at 800G;
6. path divergence between Microsoft and Facebook;
7. two webscalers promoting CPO viewed as followers, while Google and Amazon are considered leaders; and
8. our belief that Microsoft’s track record with homegrown concepts will remain dismal, as happened with COLORZ and COBO projects.
Over the last few years, fibeReality has addressed the variety of technical issues related to CPO. A big one continues to be determining which type of supplier takes on the role of integrator.
Additionally, changing the optics power seems even more disingenuous than messing with the wattage on the switch ASICs. Simply put, CPO is missing any value creation. While there have been several announcements by vendors of work being done on CPO, the focus is on future availability, and we do not take them seriously.