This means that first products using ASICs co-packaged with optics will be announced in 2021. Do we know this for sure? No. Market research is just like quantum mechanics – nothing is 100% certain, but the probability of this event is very high, as detailed in the report.
Skeptics would argue that development of the co-packaged optics technology is just starting. COBO and OIF just announced their working groups in the end of 2020. Co-packaged optics (CPO)collaboration launched by Facebook and Microsoft in 2019 is just offering guidelines for suppliers. The standards for CPO are years away.
All of this is true, but first implementations of CPOs will be based on proprietary solutions. Several large companies are working on this technology now, including AMD, Broadcom, Cisco, Huawei, Intel, Nvidia, Samsung and TSMC. All this work is highly confidential and LightCounting has no insider knowledge of the status of these projects, but this is a race and first announcements are probably just months, if not weeks away.
Public discussions are focused on co-packaging optics with switching ASICs, but it is a much higher priority for CPUs, GPUs and TPUs used in HPCs and AI Clusters. These systems are starved for bandwidth. A factor of 10x increase is needed now and another 10x will be needed for the next generation architectures based on disaggregated compute and storage.
As illustrated in the figure below, HPC and AI Clusters will be the largest market segment for CPO optics by 2025. HPCs were the first to use massive quantities of EOMs back in 2010, but the majority of EOM shipments in 2020 are going into military and aerospace systems, shown in the figure as “Other.”
EOM and CPO technologies are also of interest in the data center as the capacity of Ethernet switch chips continues to double every two years. The first 25.6Tb/s switch chips with 256, 112Gbp/s serdes are now appearing and in a year, the first 51.2Tb/s silicon is expected. 400/800G AECs, AOCs and optical transceivers will enable the majority of such switch designs, but some may use CPO. More designs will adopt CPO in 102Tb/s switches to improve power efficiency and port density. It is very likely that first CPO designs will appear in HPC, AI, military and aerospace applications, which use DACs, AOCs and EOMs now.
The distinction between EOMs and CPO is blurry. CPO is really the next generation more compact EOMs, designed to be placed much closer to an ASIC, enabling co-packaging. While EOMs have largely failed to compete with DACs, AOCs and optical transceivers, CPO is expected to be more successful. Reaching a 10x increase in bandwidth is almost impossible without CPO technology.
These are still early days in the development of real-world applications for AI. Many new products and technologies will fail, but a few will succeed and change our lives as we now know them. Regardless of who the winners arethe winners, these systems will need a lot of bandwidth. Our forecast for CPO may look too optimistic now, but it is very likely just the tip on an iceberg. It is hard not to get overexcited about the market in 2030 and this is one reason we limit our forecast to 2025.
The report includes historical data (2016-2020) and forecast (2021-2025) for shipments, revenues and average selling prices for the products mentioned above. We analyze technologies, market trends, protocol transitions, data rates, and MSAs for InfiniBand, Ethernet and other protocols. Application segments are reviewed in detail and 20 categories of products are individually tracked, forecasted and mapped into four application segments: High Performance Computing (HPC) and AI Clusters, Cloud (hyperscale datacenter), Core Routing & Optical Transport, and lastly Military/Aerospace/Other applications.
More Information on the report is available at: https://www.lightcounting.com/products/HighSpeedCables/