A Survey on Silicon Photonics for Deep Learning (2101.01751v2)

Abstract: Deep learning has led to unprecedented successes in solving some very difficult problems in domains such as computer vision, natural language processing, and general pattern recognition. These achievements are the culmination of decades-long research into better training techniques and deeper neural network models, as well as improvements in hardware platforms that are used to train and execute the deep neural network models. Many application-specific integrated circuit (ASIC) hardware accelerators for deep learning have garnered interest in recent years due to their improved performance and energy-efficiency over conventional CPU and GPU architectures. However, these accelerators are constrained by fundamental bottlenecks due to 1) the slowdown in CMOS scaling, which has limited computational and performance-per-watt capabilities of emerging electronic processors, and 2) the use of metallic interconnects for data movement, which do not scale well and are a major cause of bandwidth, latency, and energy inefficiencies in almost every contemporary processor. Silicon photonics has emerged as a promising CMOS-compatible alternative to realize a new generation of deep learning accelerators that can use light for both communication and computation. This article surveys the landscape of silicon photonics to accelerate deep learning, with a coverage of developments across design abstractions in a bottom-up manner, to convey both the capabilities and limitations of the silicon photonics paradigm in the context of deep learning acceleration.

Silicon Photonics: A Survey for Deep Learning Acceleration

The academic paper titled "A Survey on Silicon Photonics for Deep Learning" authored by researchers from Colorado State University, provides an extensive overview of the potential of silicon photonics in accelerating deep learning architectures. This research is timely given the increasing energy and bandwidth demands of deep learning models that push the limits of traditional electronic accelerators based on CPU and GPU architectures.

The paper commences by highlighting deep learning’s significant impact across various domains, particularly in computer vision, natural language processing, and pattern recognition. It attributes these feats to advances in both neural network training techniques and improvements in hardware platforms. Yet, traditional electronic accelerators face inherent bottlenecks due to limitations in CMOS scaling and metallic interconnects. Consequently, alternative computing paradigms are required, with silicon photonics emerging as a viable and promising solution.

Silicon photonics leverages light for computation and communication, potentially circumventing the bottlenecks created by conventional electronic architectures. By using optical signals, this paradigm offers substantial improvements in bandwidth and energy efficiency. Specifically, the paper argues that photonics-based multiply-accumulate operations can achieve nearly 1000 times better energy efficiency compared to the best electronic accelerators available, with bandwidths approaching hundreds of GHz.

In essence, silicon photonics could lead to revolutionary advancements in deep learning accelerators, as detailed in the survey. The paper provides a bottom-up perspective, starting from fabrication alternatives and devices, moving through neuron microarchitectures, and culminating in integrated system-level neural networks. It employs a design layer approach to classify photonic neural networks, highlighting both the inherent capabilities and limitations within the silicon photonics paradigm.

Several key devices are instrumental in realizing photonic neural networks: lasers, modulators, waveguides, couplers, photodetectors, and phase-change materials. Notably, lasers serve as light sources and can also implement neural activation functions, whereas modulators like microring resonators (MRs) and Mach-Zehnder Interferometers (MZIs) tune optical signals for effective communication and computation. Photodetectors are pivotal in converting optical signals to electrical ones and summing optical inputs, which are crucial for emulating neural activities.

Moreover, the paper introduces various types of neuron architectures, ranging from all-optical neurons to opto-electronic neurons, and coherent to noncoherent neurons. Each approach has distinct advantages and disadvantages concerning energy efficiency, scalability, and operational bandwidth. The paper acknowledges challenges such as thermal sensitivity and integration complexities, driving future research directions for low-loss components and robustness improvements.

Significant consideration is given to the architecture level, which encompasses implementations across deep neural networks, spiking neural networks, and reservoir computing. The research discusses the versatility of silicon photonics in deploying various deep learning models, illustrating their potential through coherent, noncoherent, and diffractive approaches.

Ultimately, silicon photonics offers transformative implications for deep learning systems, both theoretically and practically. This paper provides a pathway for researchers focusing on enhancing existing architectures with optical paradigms and invites further exploration into silicon photonics' robustness, scalability, and real-world performance applicability. Future investigations might center on the integration of photonic components in existing structures, developing cross-layer designs, and optimizing neural network training processes using silicon photonic networks. This outlook on photonics-based acceleration positions silicon photonics as a cornerstone for subsequent developments in AI hardware implementations.