The Need for CPO in the Latency-sensitive Data Center

Evolution of CPO in 2021 and Early 2022

Co-packaged optics (CPO) has emerged as technology of considerable interest as a platform for deploying optics in large data centers. A considerable segment of data communications sees CPO as a practical way forward to deploy 800G and above interfaces.  It is recognized by just about everybody in the industry that a monolithic optoelectronic chip with this capability lies some way into the future and is largely impossible given current commercial materials and manufacturing processes. There is also a (declining) group of technologists who believe that CPO is ahead of demand and—in the data center at least—pluggables will be all that is needed for many years to come.

In early 2021, CIR produced a best-selling analysis report pointing out (1) the advantages, disadvantages and likely applications for CPO and (2) the kind of products in which CPO will be used and the CPO standards that will evolve. CIR’s market study “Co-packaged Optics Market 2021-2025” (https://cir-inc.com/reports/co-packaged-optics-markets1799/) was widely read.

This new report brings the story up to date and points out the new and growing opportunities in this space.  That said, in this new report we are putting a much stronger emphasis on new traffic types which make co-packaging an absolute necessity and also on CPO developments in the supply chain, noting here that the switching components and the optoelectronics component have traditionally been supplied by what is essentially two industries.  We think that as optics and electronics merge in CPO it will become necessary for these two industries—and their supply chains—to merge, too.

CPO and the Changing Nature of Data Center Traffic

By the second half of 2021 the CPO fuss that had been so prevalent at the beginning of the year had declined. Nonetheless, in the year or so since CIR’s previous report, there have been important developments as far as (2) is concerned and these form the subject matter of Chapter Two.  It has also become clear that most of the developments in CPO in the near-to-medium future, will be in the data center, which is our primary concern in this report.

CIR believes that a primary question that has to be answered before assessing the prospects for CPO in the data center is the nature of traffic in the data center in the coming years.  The view that this report has adopted is that data centers will need to adapt to a growing amount of high data-rate traffic that is also highly sensitive to latency.  The most important examples of this kind of traffic are AI, machine learning (ML), augmented reality (AR) and virtual reality.  What all these kinds of traffic have in common is that (like video) they entail a lot of new bandwidth, but they are much more latency sensitive than video. In this report we conjecture that there will be enough of this new kind of traffic around that CPO will be needed sufficiently that a CPO product market will emerge.

About 90 percent of traffic is now video.  While video will increase in absolute terms, this 90 percent share can be expected to decline for obvious mathematical reasons.  We are now in a new era where latency-sensitive traffic will become pre-eminent. AI, ML, AR and VR traffic will have to be dealt with.  Simple kinds of AI traffic (e.g., for Siri) are now common, although the pace at which they will grow is something of a speculation.  Also, increasingly important will be data traffic from augmented reality (AR) and virtual reality (VR), perhaps reborn as the “multiverse.”

The role of this kind of traffic in the data center is the subject of Chapter Three of this report.  Like video, the evolution of latency sensitive networks may start slow. Then everything will probably happen at once—typical exponential growth in fact.  A variety of techniques may be applied to solve the latency problem, but higher bandwidths can help.  High data-rate interconnects like CPO will certainly be part of dealing with latency issues but they are part of it.

Five-year Forecast of CPO Products

All the excitement about CPO at the beginning of 2021 may well have overshot, but the need for speed coupled with the need to minimize power consumption makes the ultimate victory of optoelectronic integration in the next-generation data center all but certain.  Over time, a lot of work needs to be done and this will take a considerable thinking through of design, materials and technology considerations.  For example, an open question is how shared cooling schemes can be arranged for a situation that will increasingly occur where hot, high-performance chips will sit next to optics and the cooling technology must be shared and optimized between electronics and optics.

But for 2022, however, we don’t expect too much drama around CPO—significant launches of CPO switches, most notably, and steady progress on the standards/vendor agreement front. The organizations to look to on the “standards” (the term “standards” is not always used) are the OIF, the Consortium for On-Board Optics (COBO), less so the Co-Packaged Optics Collaboration.  Another interesting likely direction is the emergence of External Laser Small Form Factor Pluggable (ELSP) products that may go some way to address the apparently retrograde implications of early CPO which abandoned pluggability in a major sense.

Chapter Four of this report consists of a forecast of CPO transceivers and other products for the data center. As the data has come in, we have revised our numbers from the 2020 report.  In particular, we have changed prices based on additional information received.  As usual, the report includes granular forecasts of CPO in the data center, with a discussion/breakout by data rates, size of data center, switching generation, CPO design approach, standards, MSAs, pricing, etc.  These forecasts are based on how CIR expects price per GB for CPO evolves.

The market prospects for CPO products are shaped by many factors, but the ones set out below are the most important.

Factors Shaping the Market for CPO Products in the Data Center
Type of trafficFor almost a decade now, traffic in the data center has become ever more video oriented.  Video is now reaching saturation point, while AI/ML traffic is taking off—AI/ML and AR/VR may be the new video but is much more latency sensitive than video ever was.  Although the penetration of this kind of traffic is already well advanced, various roadmaps and timeframes (discussed later in this report) are possible here as are various scenarios for their impact on CPO standards and product development
Number of data centersData centers have been consolidating for some time.  This means fewer purchasers of interfaces but more need for 400G and 800G interfaces, some of which will no doubt use CPO.  For the forecasts in this report, we assume that this consolidation will continue and a noticeable pace
Move of high data rates downwardInitially, very-high speed transceivers will be used for building-to-building and router-to-router applications, but they will increasingly be used lower in the enterprise data center hierarchy, eventually all the way down to the rack level
Use of CPO for high data rates800G will almost certainly use some pluggables but the proportion will decline over time.  Once data centers begin to adopt data rates above 1 Tbps, it should be assumed that CPO (or something like it) will begin to dominate data center communications.  The “or something like it” is supposed to indicate that CPO is at a very early stage of technical and standards and may well evolve into new forms as packaging and materials technology develops

Source:  CIR

From switching generations to the rack:  CPO is currently driven by new switching generations.  Each generation has greater capacity—room for more/higher-capacity transceivers, but increasingly demanding in terms of power.  Anecdotally, switch bandwidth doubles every two years.  According to Cisco, to achieve the 80x switch bandwidth increase from 640 Gbps to 51.2 Tbps the power utilized by the switch increases by 9.5x. So, power efficiency gets a lot better, but the absolute power demand continues inexorably to increase.  For example, according to one source, a single building on a data campus can consume 100MW.

CPO comes to the rescue by bringing the optics and electronics closer together.  The closer two devices get the lower the power needed to transmit a signal between them. In CPO the optical engine is shifted into the switching silicon package saving significant amounts of power.

This switching story is dominant in accounts of CPO today; it reflects the current switching environment where the vendors are primarily concerned with doing something about the sudden switching port count bottleneck created by the arrival of commercial 800G and the need for switches that can accommodate 800G ports.  There will be no significant revenues from 800G transceivers unless switches that can accommodate a reasonable number of 800G transceivers come into being. And a CPO-based 51.2 Tbps switch design is much more power-efficient than a traditional 51.2 Tbps with pluggable optics.

It is likely a 51.2 Tbps switch device will arrive in 2022.  Then switching evolution will continue, with power considerations making it harder to build switches without CPO with each generation. The next generation of switches—102.4 Tbps switches in 2024 or 2025 will all be CPO switches.  But after 51.2 switches, the focus of network design will shift from an immediate 800G port “crisis” to the racks.  As rack speeds inevitably increase, the DACs and other copper will give way to optical cable. While rack switches mostly will not need CPO right away, they will inevitably, expanding the market for CPO, at least once they get beyond a certain speed.

Some consider that the sweet spot for CPO will begin with 100 Tbps.  Once 200 Tbps switches appear, co-packaged optics will become a necessity.