Search

Silicon photonics: from transceivers to speed-of-light AI

Since 1985, silicon photonics has progressed from the initial development of high confinement waveguides to strategically incorporate CMOS techniques, establishing its dominance in the transceiver space. In the coming years, it has the potential to expand to a wide range of innovative applications. – An article written by Martin Vallo, Senior Analyst, Photonics, at Yole Intelligence, part of Yole Group, for PIC MAGAZINE.

Since early 2023, there has been a lot of hype around – and massive investments into – silicon photonics, especially optical computing, optical I/O, and diverse sensing applications. It seems logical that primary technologies in various applications will be replaced relatively quickly by optical-based designs and architectures. The giants forecast that optics will be necessary and become ubiquitous relatively soon, while startups are developing new applications through R&D. So, can we expect this prediction to be realised any time soon?

While there are many arguments about the necessity of photonics coupled with electronics, the largest silicon photonics market – datacom pluggables – generates only around 12 percent of datacom transceiver revenue (projected to reach 30 percent by 2028). The semiconductor market is suffering from an extended period of decline, leading to more pragmatic buying behaviour on the part of customers. DC operators prefer longestablished and low-cost technology solutions. Yole Intelligence’s market research shows that silicon photonics is not yet a primary technology, even for intra-data-centre interconnects with up to 500m reach.

In this context, silicon photonics remains a technology in active development, with a wide array of potential applications, hinting at promising opportunities on the horizon. In the coming decade, frontrunners will emerge, leading to industry consolidation. Nevertheless, the broad spectrum of applications will ensure abundant opportunities for the technology to expand and proliferate.

Yole Group, in its new Silicon Photonics 2023 report, estimated the silicon photonics PIC market was worth US$68 million in 2022 and is forecast to generate more than US$600 million in 2028 at a 44 percent Compound Annual Growth Rate for 2022-2028 (CAGR2022-2028). This growth will mainly be driven by 800G high-data-rate pluggable modules for increased fibre-optic network capacity. Additionally, projections of rapidly growing training dataset sizes show that data will need to use light to scale ML models using optical I/O in ML servers.

Note from the author – Forecasts have been substantially revised downward compared to the Silicon Photonics report, 2022 edition. Shipments of silicon photonics-based products have seen a notable decrease, primarily attributed to reduced adoption in datacom applications. This is compounded by a careful revision of the manufacturing yield and the average selling price of silicon photonic devices. Yole Group now excludes the tunable laser in the PIC circuit for telecom. Consequently, these adjusted estimations impact the silicon photonics forecast between 2023 and 2028.

Substantial data centre requirements, particularly in the domains of artificial intelligence (AI) and machine learning (ML), are expected to fuel the ongoing over the next decade. With the conventional processor-centric computing architecture and copper interconnects, the state-of-the-art chips based on 3nm technology are approaching their physical limitations, while the necessity for faster data transmission has surged. Silicon photonics, with its ability to facilitate high-speed communication, has therefore become a prime focus.

Architectures that include optical I/Os can take out streamline access between compute nodes and memory pools, harnessing the fan-out capabilities of optics to minimise the number of switching hops required to access resources. Broadcom’s strategic plan outlines a trajectory for switching chips, projecting an increase from 51.2 Tb/s (5 nm process node) this year to 102.4 Tb/s (3 nm process node) in 2025, and an impressive 204.8 Tb/s (2 nm process node) by 2027. This exponential growth could serve as a significant catalyst for the advancement of silicon photonics in networking applications, paving the way for significantly enhanced data capacity in the future. Silicon photonics provides a versatile platform for applications with high-volume scalability demands.

The primary and most immediate domain for its application is data centres, where Intel holds a dominant position. A second major high-volume application is telecommunications, as exemplified by Acacia, benefiting from the consistent and superior performance of silicon processing. A third broad application area encompassing optical LiDAR systems has significant potential but faces cost and 2D beam-scanning challenges. 3D integration, housing both chips on the same silicon substrate, is vital for seamless control. Optical gyroscopes need sizeable chips for sensitive rotation sensors, benefiting from silicon substrates and SiN waveguides. Quantum computing is pivotal in the evolving AI and machine learning landscape. Optical computing, ideal for efficiency-focused tasks, garners industry attention and promises substantial impact.

Advanced photonic components and their integration for medical use can transform healthcare, enabling faster, more precise diagnostics, treatment, and patient monitoring. Overcoming regulatory and standardisation challenges may be necessary for clinical adoption. The outlook for silicon photonics-based medical applications is promising and holds significant potential for various healthcare and medical fields. Extending silicon photonics into the visible spectrum shows potential for future developments, offering a vast range of innovative applications.

The silicon photonics industrial landscape is forming around diverse players, including: major vertically integrated players (Intel, Cisco, Marvell, Broadcom, Nvidia, IBM, etc.); actively engaged in the silicon photonics industry; startups and design houses (AyarLabs, OpenLight, Lightmatter, Lightelligence); research institutions (UCSB, Columbia University, Stanford Engineering, MIT, etc.); foundries (GlobalFoundries, Tower Semiconductor, imec, TSMC, etc.); and equipment suppliers (Applied Materials, ASML, Aixtron, etc.). All these players contribute to significant growth and diversification.

Intel is a leader in this field, investing heavily in research and development…

… Read the full article here (page 24).

up