LightTrends Newsletter

AI Clusters and HPCs are the right entry points for co-packaged optics

May 2022

LightCounting updates forecasts for silicon photonics and co-packaged optics

The transition from pluggable to co-packaged optics (CPO) is an exciting one for our industry, but it is important to set realistic expectations for the adoption of CPO in data centers. Apart from numerous manufacturing challenges and meeting the targets set for lower power consumption, end-users have to accept CPO as a viable approach for continuing cost reduction.

There is more uncertainty about the timing of CPO adoption in AI Clusters and HPCs, but this market is more open to taking risks and using innovative solutions, even if they are proprietary. These systems are even more starved for bandwidth than compute clusters in data centers. Experts argue that GPU interconnects could use 10 times more interconnect bandwidth now. Emergence of disaggregated clusters will need yet another 10x increase in the future.

The architecture of AI Clusters and HPCs is evolving. We may see CPO deployed on GPUs, TPUs as well as Ethernet, InfiniBand or NVLink switches. There is a variety of FPGA-based accelerators, which may also benefit from CPO. Our current forecast combines all these use cases into a single “AI Cluster and HPC” application segment.

Compute clusters in mega datacenters will be the second largest application of CPO. Some of the large customers are not planning to use proprietary CPO designs and prefer 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 a risk.

The Figure below presents LightCounting’s forecast for shipments of CPO ports and pluggable Ethernet optical transceivers and AOCs. The pluggables will continue to dominate the market for the next 5 years and probably longer than that. However, CPO ports will account for almost 30% of the total 800G and 1.6T ports deployed in 2027.

CPO enthusiasts may argue that this view is too conservative, but we suspect that it may be too optimistic. Forecasters tend to underestimate the time required to change course in an industry.

In this analysis we count CPOs in terms of 800G and 1.6T transceiver equivalents (or ports), but CPO engines can combine multiple 800G or 1.6T ports into one opto-chiplet. For example, a single 3.2Tbps engine is equivalent to 4 ports of 800G CPO. Each AOC is counted as two ports.

Our current forecast does not include shipments of near package optics (NPO), developed by the OIF. Meta does plan on using NPO on 51.2T switches, but these are likely to be very limited proof of principle trials.

Will all CPO solutions be based on silicon photonics SiP technology? Most likely not. IBM is developing a VCSEL-based system. A start-up Avicena is working on GaN micro LEDs for very low power, short reach (<10m) connectivity. There are more start-ups still in stealth mode, and we are looking forward to seeing more technologies entering the race. SiP is the front runner and the ultimate integration platform, at least for now.

The updated report also presents forecast for adoption of SiP technology. It also includes estimates for market shares of optical transceivers based on InP, GaAs and LiNbO3, including thin film LiNbO3.

Ready to connect with LightCounting?

Enabling effective decision-making based on a unique combination of quantitative and qualitative analysis.
Reach us at info@lightcounting.com

Contact Us