Photonic computing is experiencing renewed interest, driven by the growing demands of AI computations. The Mach-Zehnder Interferometer (MZI) serves as a cornerstone of Photonic Integrated Circuits (PICs), enabling light phase modulation for parallel processing. Recent innovations, including lithium tantalate (LiTaO3) implementation and Schrödinger microcombs, are advancing the field's efficiency. As silicon photonics technology matures, it emerges as a promising solution for AI applications.
While photons can carry information like electrons, their applications extend to both communication and computing.
Historically, photons were primarily used for long-distance communication, with optical fibers replacing traditional telephone lines for internet applications. The technology has gained renewed attention due to the rise of AI servers. Looking ahead, most communications will occur at chip-to-chip and server-to-server levels, with massive data transmission remaining the most energy-intensive aspect of information processing and transfer.
However, server chip design currently prioritizes performance based on traditional PPA (Performance, Power, Area) metrics, often overlooking power consumption and heat dissipation during the design phase. These challenges must then be addressed during manufacturing and advanced packaging, driving industry investment in silicon photonics, which the semiconductor industry sees as a promising solution.
The potential of photonic computing has long been recognized, dating back to the 1960s and 1970s with the advent of lasers and analog signal processing, leading to significant developments of photonic components in the 1980s. However, by the 1990s, when the technology reached application and mass production stages, advanced manufacturing processes were operating at around 0.5–0.8 micrometers, while photonic components typically exceeded one micrometer. This scale disparity made it difficult for photonic products to compete with electronic components on silicon wafers, causing the field's progress to diverge significantly.
The emergence of AI has prompted a reevaluation of photonic computing. AI computations, such as those performed by Convolutional Neural Networks (CNNs) or transformer models used in large language models, rely heavily on parallel matrix multiplication. Given these workloads' characteristics - large data volumes and relatively straightforward algorithms - photonic computing presents a promising solution. In 2016, Yichen Shen and his research partners proposed using photonic computing to process deep learning tasks.
The Mach-Zehnder Interferometer (MZI) has emerged as a fundamental component of silicon photonics. As the basic unit of silicon photonics, the MZI primarily modulates light phase. When light enters an MZI, it first passes through a splitter, dividing it into two beams that follow separate optical paths. One path remains unaffected, while a controllable voltage is applied to the material in the other path, altering its refractive index and consequently changing the light phase. The two beams then recombine, interfering with each other. Through phase control of one path, the resulting interference creates different light intensities at the exits of the two paths, enabling the MZI to function similarly to a transistor in computation.
The MZI's capabilities make it the foundation of Photonic Integrated Circuits (PICs), enabling the construction of circuits capable of performing matrix multiplications and establishing a foothold for photonic computing in AI applications.
Computational efficiency in photonic systems can be significantly enhanced through Schrödinger microcombs. These microcombs separate continuous wave laser sources into multiple evenly spaced frequency sources, enabling parallel computation. A single microcomb can generate dozens to hundreds of frequencies for simultaneous calculations, effectively compensating for the size limitations of conventional optical components.
When photonic computing proposals first emerged in 2016, silicon photonics technology was still immature. The previous Heterogeneous Integration Roadmap (HIR) did not anticipate significant silicon photonics adoption until 2020, and even then, the timeline for mass production has lagged considerably.
Recent proposals suggest utilizing lithium tantalate (LiTaO3) for silicon photonic components, enhancing the feasibility of using MZIs for photonic computing. Lithium tantalate, already employed in 5G communication, is compatible with semiconductor manufacturing processes. It offers low production costs and several advantageous physical properties for MZI fabrication. Its low birefringence simplifies circuit design and increases optical component density, while its low optical loss supports effective signal maintenance.
Moreover, high-performance MZIs made from lithium tantalate can achieve an electro-optical bandwidth of 40 GHz and a half-wavelength voltage-length product of 1.9 V•cm, indicating ease of phase modulation.
While photonic computing theoretically offers high speed and low power consumption, representing a potential solution to various contemporary computing challenges, it previously gained little traction due to the immaturity of silicon photonics technology and the larger sizes of optical components compared to microelectronic components.
However, driven by the rise of AI servers, silicon photonics technology is advancing with additional industry support, capitalizing on new opportunities. Current applications focus on ASIC-type computations closely related to AI, seeking a foundational foothold in the market.
About the author
Albert Lin received his Ph.D. in Physics in 1988, taught at National Central University, and then moved to the technology industry. Lin served as director and vice president of ProMOS Technologies, and president of ConDel International Technologies. He chaired the Lithography Forum and served as the chairman of the Supervisory Board of the Taiwan Semiconductor Industry Association. Lin is now a visiting research fellow in the Department of Physics at National Taiwan University. His main research fields are new materials, new mechanisms, and basic research on quantum information. He is the standing supervisor of the Taiwan Association of Quantum Computing and Information Technology.
