Datacom will be the main driver for Silicon Photonics

Despite the dashed hopes of using Si photonics for consumer applications in the short term, this technology is still very promising, especially for data centers.

By Eric Mounier, Ph.D., is Chief Analyst at Yole Intelligence, part of Yole Group

Over the past 50 years, mobile technology innovations have been rolled out each decade. Mobile bandwidth requirements have evolved from voice calls and texting to ultra-high-definition (UHD) video and a variety of augmented reality / virtual reality (AR / VR) applications. Despite the profound implications of the COVID-19 outbreak for the telecom infrastructure supply chain, consumers and business users worldwide continue to create new demand for networking and cloud services. Social networking, business meetings, UHD video streaming, e-commerce, and gaming applications will continue to drive growth.

The driving forces are mainly social media, digital entertainment, and cloud computing, which have voracious appetites for more bandwidth and faster interconnect speeds. This, in turn, affects the data center network architecture (disaggregation, AI and ML, strong focus on power consumption). All this will require new high-speed, low-latency technologies for data transfer (heterogeneous packaging, interposers, PICs,).

Figure 1

Optical interconnects have become an essential technology in data/telecommunication infrastructure. In the Co-Packaged Optics for Datacenter 2023 report, Yole Intelligence, part of Yole Group, asserts that the development of semiconductor technologies such as lasers, modulators, and DSPs enabled an increase in bandwidth and accelerated the development of the overall infrastructure. Over the past years, many technological evolutions occurred and will continue to develop at different network levels (see figure 1). For long haul and metro core, InP PICs might play a significant role in future optical modules. DSP will improve with higher levels of modulation. As we move from metro access (below 100 km) to intra- and inter-data center networks, silicon photonics will be increasingly used; optics will become standard for ultra-short reach connections, coherent technology will be used, and new switching architecture (co-packaged optics) may become commercially viable. But we currently see some macroeconomic headwinds that will negatively impact budget-intensive projects, particularly for emerging technologies such as CPO. Recent news indicates that most of the main CPO proponents have suspended support for CPO programs. There are several reasons why CPO is losing attractivity, and the first is the well-established industrial ecosystem around pluggables. Also, new optical technologies for pluggable form factors, including TFLN, BTO, carbon, and polymer modulators, can help achieve the required low power and be introduced to the market without changing any existing network system designs.

Historically, integrated photonics has been developed on an SOI platform. The goal was to benefit from wafer-scale manufacturing in the CMOS industry and use it for photonic chips. But SOI is expensive, and silicon is not the ideal material for all the different photonic functionalities. Since the very start, laser has been one of the greatest challenges for silicon photonics. The considerable development effort made by Intel for InP chiplet integration on SOI has been key to Intel’s actual business success in silicon photonics. More companies are now trying to duplicate this model to offer Si photonic wafers with integrated laser chips. Today, as data rates increase, high-speed modulation using Mach-Zehnder on Si is becoming a bottleneck. There are numerous developments in new materials to overcome the current limitation (TF LNO thin films, InP, BTO, polymer, plasmons). This creates opportunities for companies focusing on materials for silicon photonics: Lumiphase and Polariton, Hyperlight and Liobate in China for thin film LNO, and Riber for BTO. As integrated optics moves towards increased functionalities, the definition of Si photonics would expand to incorporate other materials (figure 2).

But the situation for AI / ML systems is different. The potential for billions of optical interconnects (chip-chip, board-board) in the future drives big foundries to prepare for mass production. Since most of the photonics manufacturing IP is held by non-foundry firms, big foundries such as Tower Semiconductor (Intel), GlobalFoundries, and others are preparing silicon photonics process flows to accept any PIC architecture from design houses. All of them are joining forces in industry consortiums such as PCIe, CXL, and UCIe.

So datacom and servers will continue to be the main drivers for integrated optics. Silicon photonics will be central in data center evolution in the short-term for 100G (already well implemented in data centers) and then 400G and 800G pluggables. It will also be an enabling technology for disaggregating data centers and a possible future CPO approach. The Covid-19 pandemic and the Ukraine-Russia crisis have strongly impacted the semiconductor supply chain. However, the impacts have been less in silicon photonics than in other semiconductor markets. Indeed, silicon photonics requires low volumes of wafers and strategic technology, so allocated capacity does not affect the global fabs’ capacity, which is generally much larger.

The market value of silicon photonics to the industry was more than 20% of the optical transceivers market for datacom in 2021, and this proportion is still growing. It will be more than 30% in 2027. Silicon photonics is increasingly used for 500 m DR and is also increasingly used as coherent technology enters datacom applications. There is also an increasing demand for 400ZR (figure 3). And besides datacom, other applications are promising in computing, interconnects in data centers, and sensing. Optical interconnects using silicon photonics will enable disaggregated data centers with more power available for high-performance computing (HPC) and data communications. The industrial ecosystem (Ayar Labs, Intel, Ranovus, Lightmatter, AMD, GlobalFoundries, and others) around ML machine vendors Nvidia and HPE has made decent progress, with a plan to ship products in volume between 2024 - 2026. Photonic computing, which allows for analog AI computations much faster than today’s digital AI, has been developed and will also hit the market soon. Other applications include sensors (immunoassays, gyroscopes, and LiDARs), 5G, optical processing, and CPO. Medical has begun to hit the market, with Genalyte and many other startups using Si-integrated optics as a manufacturing platform. In the automotive domain, more and more manufacturers are integrating LiDAR into their products.

At the supply chain level, Intel is strengthening its market leadership with a 58% market share in units in 2021, followed by Cisco and other smaller companies (Marvell/Inphi, Sicoya, Acacia, and others) that, step by step, are gaining market. After decades of investments from Intel, the silicon photonics market is now the centerpiece of the evolution of the data center. Intel is now securing its access and manufacturing capabilities to leverage all its core competencies for the next decades. From autonomous mobility to data centers and the future of cloud computing, Intel is now strengthening its position in the silicon photonics landscape, with great results recently. From a decentralized on-the-edge central processing leader, Intel is transforming itself into the centerpiece of the future of computing based on photonics and data centers.

Figure 3.

While some players are taking the opportunity offered by silicon photonics to enter the optical transceiver market, others are seeking new applications, and the silicon photonics industrial landscape has remained very active.
New potential applications of silicon photonics have resulted in the creation of various companies in the past three years. So, while some players are taking the opportunity offered by silicon photonics to enter the optical transceiver market, others are looking for new applications: medical, sensors, interconnects, computing. Co-packaged optics seem to require a considerable investment which is not yet acceptable to any player, delaying its introduction. Maximum shipments of optical transceivers will probably arrive after 2026. This will push players that cannot get into the co-packaged optics application to move to new applications in 5 - 10 years. We should therefore see a shift in applications in the coming years. And China continues to be very active in the development of silicon photonics, with many players involved.

• Silicon Photonics 2022 report, Yole Intelligence
• Optical transceivers 2022 for Datacom & Telecom 2022 report, Yole Intelligence
• Co-Packaged Optics for Datacenter 2023 report, Yole Intelligence

About the author
Eric Mounier, Ph.D., is Chief Analyst at Yole Intelligence, part of Yole Group. With more than 30 years’ experience within the semiconductor industry, Eric provides daily in-depth insights into emerging semiconductor technologies such as quantum technologies, the Metaverse, terahertz, photonics, and sensing.Based on relevant methodological expertise and a significant technological background, Eric works closely with all of Yole Group’s teams to highlight disruptive technologies and analyze business opportunities through technology & market reports and custom consulting projects. Eric has spoken at numerous international conferences, presenting Yole Group’s vision of emerging semiconductor technologies, markets, and applications.Previously, Eric held R&D and Marketing positions at CEA-Leti (France). Eric Mounier has a Ph.D. in Semiconductor Engineering and a degree in Optoelectronics from the National Polytechnic Institute of Grenoble (France).