Photonics for Neuromorphic Computing: Fundamentals, Devices, and Opportunities

TL;DR

This paper surveys photonic neuromorphic computing as a non-von Neumann path to address the data- and parameter-scale growth in AI. It synthesizes fundamentals of photonics, active/passive photonic components, optical interconnects, and device fabrication platforms, emphasizing metrics such as energy efficiency $OP/J$ and compute density $TOP/mm^2/s$ to evaluate state-of-the-art photonic accelerators. Key milestones are highlighted, including all-optical and hybrid optical–digital networks (e.g., PDNN, VCSEL-ONN, D2NN, AONN) and the emerging photonic tensor core that leverages soliton microcombs and phase-change materials for massive MVM. The article also discusses emerging light sources and SOI-based integration, identifies remaining challenges (cost, scalability, reliability, integration), and outlines future directions that could enable PICs to co-exist with or eventually complement electronic ICs in large-scale AI and scientific computing tasks.

Abstract

In the dynamic landscape of Artificial Intelligence (AI), two notable phenomena are becoming predominant: the exponential growth of large AI model sizes and the explosion of massive amount of data. Meanwhile, scientific research such as quantum computing and protein synthesis increasingly demand higher computing capacities. Neuromorphic computing, inspired by the mechanism and functionality of human brains, uses physical artificial neurons to do computations and is drawing widespread attention. Conventional electronic computing has experienced certain difficulties, particularly concerning the latency, crosstalk, and energy consumption of digital processors. As the Moore's law approaches its terminus, there is a urgent need for alternative computing architectures that can satisfy this growing computing demand and break through the von Neumann model. Recently, the expansion of optoelectronic devices on photonic integration platforms has led to significant growth in photonic computing, where photonic integrated circuits (PICs) have enabled ultrafast artificial neural networks (ANN) with sub-nanosecond latencies, low heat dissipation, and high parallelism. Such non-von Neumann photonic computing systems hold the promise to cater to the escalating requirements of AI and scientific computing. In this review, we study recent advancements in integrated photonic neuromorphic systems, and from the perspective of materials and device engineering, we lay out the scientific and technological breakthroughs necessary to advance the state-of-the-art. In particular, we examine various technologies and devices employed in neuromorphic photonic AI accelerators, spanning from traditional optics to PICs. We evaluate the performances of different designs by energy efficiency in operations per joule (OP/J) and compute density in operations per squared millimeter per ...

Figures (24)

• Figure 1: Trend of increase in representative AI model's parameters over time. Parameters (i.e. weights) are variables in an Al system whose values are adjusted during training to establish how input data is correlated with the output labels. The most recent chatGPT has a record-breaking 1.8 trillion parameters and is still growing. No. of parameters are estimated based on published statistics in respective papers and naturally come with some uncertainty.
• Figure 2: a. Biological neuron in animals; b. Multi-layer perception neural networks (MLP) or fully-connected (FC) layers; c. Forward propagation of artificial neurons in MLP, including the input, weighs, summation, activation function, and the output. Data obtained from online image libraries.
• Figure 3: Timeline of the evolution of photonic neuromorphic systems, especially for AI applications. Existing works can be categorized into three branches: 1. simulate the principles of forward propagation in an artificial neuron; 2. achieve image classification or pattern recognition; 3. realize convolutional operations by performing MVM. Sources are from milestone works spanning from the 90s to 2023: psaltis1990holographyShen_Harris_Skirlo_Prabhu_Baehr-Jones_Hochberg_Sun_Zhao_Larochelle_Englund_etLin_Rivenson_Yardimci_Veli_Luo_Jarrahi_Ozcan_2018feldmann2019allZuo_Li_Zhao_Jiang_Chen_Chen_Jo_Liu_Du_2019feldmann2021parallelxu202111ashtiani2022chipliu2022programmablechen2023allchen2023deep. For detailed specs of each work, refer to Table 1.
• Figure 4: Overall hierarchical organization of this paper and logical links between all sections.
• Figure 5: Comparison of state-of-the-art AI accelerators between conventional digital electronics and the emerging ONNs empowered by photonic neuromorphics. Comparison is embodied by the FoM such as energy efficiency ($TOP/J$) and compute density ($TOP/mm^2/s$). Corresponding FoM values of each reference can be found in Table 1. On both the x and y axis, larger values are better. As indicated by the legend, it should be noted that while the electronic systems shown here are commercial products, the photonic (ONN) systems are laboratory demonstrations. Future endeavor calls for continued efforts to improve the energy efficiency and compute density of photonic systems. Adapted with permission.chen2023deep Copyright 2023, Springer Nature.