• Categories:Industry News
• Time of issue:2022-06-07

(Summary description)      The transition from plug-in Optics to co-packaged Optics (CPO) is exciting for the optical communications industry, but it is important for data centers to have realistic expectations for CPO adoption. In addition to the numerous manufacturing challenges and meeting the goal of reducing power consumption, end users must also accept CPO as a viable way to continuously reduce costs.       Artificial intelligence (AI) clusters and high performance computers (HPC) have more uncertainty about the timing of CPO adoption, but this market is more willing to take risks and use innovative solutions, even if they are Proprietary. These systems require even more bandwidth than computing clusters in data centers. Experts believe that GPU interconnects can now use up to 10 times the interconnection bandwidth currently used. Disaggregated clusters need to increase about 10 times in the future.       AI clusters and HPC architectures are evolving. We may see CPOs deployed on Gpus, TPus, and Ethernet, InfiniBand, or NVLink switches. There are a variety of FPGA-based accelerators that may also benefit from CPO. LC's current projections combine all of these use cases into one "AI clustering and HPC" application category.       Computing clusters in large data centers will be the second largest application of CPO. Some large customers are not going to use proprietary CPO designs, preferring to wait for a new competitive ecosystem based on standard CPO solutions to emerge. This will limit the scale of early deployments, but there will be customers willing to take risks.       The figure below shows LC's forecast for CPO ports and pluggable Ethernet optical transceivers and AOC shipments. Pluggable devices will continue to dominate the market for the next five years and beyond. However, CPO ports will account for nearly 30% of the total 800G and 1.6T ports deployed by 2027.       CPO supporters may regard this view as too conservative, but LC suspects it may be too optimistic. Forecasters tend to underestimate how long it takes to change an industry's direction.       In this analysis, LC calculates CPO based on 800G and 1.6T transceiver equivalents (or ports), but a CPO engine can combine multiple 800G or 1.6T ports into an opto-chiplet. For example, a single 3.2Tbps engine is equivalent to four 800G CPO ports. Each AOC is calculated as two ports.       LC's current projections do not include shipments of Near Package Optics (NPO) developed by OIF. Meta does plan to use Npos on 51.2T switches, but these are likely to be a limited number of proof-of-concept trials.        Will all CPO solutions be based on silicon Photonics (SiP) technology? Probably not. IBM is developing systems based on VCSEL. A start-up company, Avicena, is developing GaN Micro leds to enable extremely low-power, short-range (<10m) connections. More startups are still in stealth mode, and LC says it's looking forward to seeing more technology enter the fray. But at least for now, silicon light is the leader and the ultimate integration platform.