智能光计算作为近年来新兴的计算模态，在后摩尔时代展现出远超传统硅基电学计算的通量与能效，有望解决人工智能领域的算力与功耗难题。然而，现有智能光计算面临通用传播模型不明、计算架构缺失、计算规模受限等桎梏，其计算任务往往局限于简单的字符分类、基本的图像处理等。对于亟需高算力与高能效的复杂大模型智能计算，光的高性能潜力难以施展。针对大规模光电智能计算难题，本文刻画干涉-衍射多通道联合传播模型，建立光子-电子神经元对偶互易理论，首创分布式并行广度光计算架构，刻画通用矩阵张量化分解重构计算机制。研制了大规模通用智能光计算系列芯片Taichi与DTU等，能效与算力超越现有商用AI芯片2-3个数量级。突破光子智能计算规模与通用性桎梏，首次赋能光计算实现自然场景千类对象识别、跨模态内容推理生成等通用人工智能任务。本文贡献及主要创新点总结如下：（1）本文独辟蹊径，提出大规模智能光子计算通用理论与架构，通过建立光学干涉-衍射多通道联合传播调制模型与光子-电子对偶互易训练理论，摒弃传统电子深度计算的范式，首创多通道分布式广度光计算架构与光学神经元时空互易重构机制，为光子实现通用大规模智能推理训练计算提供理论支撑。（2）针对现有智能光计算规模受限难题，本文基于光学衍射-干涉通用传播模型与分布式广度光计算架构，刻画衍射光计算大规模并行优势与干涉光计算灵活重构特性，提出衍射编码-干涉特征计算-衍射解码的融合计算方法，研制了干涉-衍射异构集成智能光芯片Taichi，实现160 TOPS/W高效大规模光计算。（3）针对现有智能光计算重构通用性难题，本文基于光学神经元时空互易重构机制，将具有空间分布难以重构的衍射计算权重映射为与输入信号并行的权重时间序列，研制高通量可重构衍射计算芯片DTU，以高速光信号调制器时序复用，实现大规模通用任意矩阵计算，支撑百万TOPS级算力的光子通用人工智能。大规模智能光计算有望在如今大模型通用人工智能蓬勃发展的时代，以光子之道，为百亿像素大场景光速智能分析、百亿参数大模型训练推理、毫瓦级低功耗自主智能无人系统等探索新灵感、新架构、新路径。
The rapid development in the field of artificial general intelligence (AGI) imposes stringent energy efficiency and area efficiency requirements on next-generation computing. As a brand-new technology to break the plateauing of Moore’s Law, integrated photonic neural networks have shown great potential in achieving superior processing speed and high energy efficiency. However, it suffers severely from limited computing capability and scalability that only simple tasks and shallow models (e.g., 4-class vowels classification, 10-category digits classification, etc., with up to thousand-parameter-level networks) were experimentally realized in photonics, hardly supporting modern AGIs. In response, this dissertation proposes novel photonic computing theories and architectures to exploit the high parallelism and high connectivity of wave optics, implementing computation with high computing density and capability. Specifically, the proposed approaches explore a multi-channel diffractive-interference-hybrid propagation and modulation theory, and an optical-artificial dual-neuron learning theory for photonic computing. Based on the theories, a distributed encoding-embedding-decoding photonic computing architecture, and a tensorized spatial-temporal-equivalenced diffractive reconfiguring architecture are introduced to effectively increase the scalability and generalizability of optical neural networks (ONNs). For the first time, the approach enables neural networks with million-level parameters to be realized on-chip optically, achieving general-purpose advanced intelligence tasks e.g., multi-view three-dimensional perception, thousand-category-level classification, and artificial intelligence-generated content (AIGC), with up to 2 orders of magnitude improvement in area efficiency and energy efficiency compared to current artificial intelligence (AI) chips. Main innovations of this dissertation are summarized as follows:(1) This thesis proposes a multi-channel diffractive-interference-hybrid propagation and modulation theory with an optical-artificial dual-neuron learning theory for large-scale photonic computing, effectively enabling large neural networks and advanced intelligence tasks to be both trained and inferenced optically. Combining the high-throughput advantage of optical diffraction and high-flexibility advantage of optical interference, such a novel photonic computing paradigm validates the scalability and capability potentials of photonic computing. (2) This thesis reports the on-chip integration for large-scale photonic intelligent computing, and proposed Taichi, a large-scale photonic AI chiplet, owning millions-of-neuron on-chip computing capability with superior robustness and energy efficiency. By integrating scalable wavefield diffraction and reconfigurable interference, the entire inputs are encoded passively and modulated in a highly-parallel way, achieving 160 TOPS/W on-chip energy efficiency and 879 T MACS/mm? area efficiency in versatile real-world applications (up to 2 orders of magnitude improvement in both energy and area efficiency than existing AI chips).(3) This thesis addresses the challenge of general-purpose photonic intelligent computing and application adaptivity by proposing DTU, a fully reconfigurable large-scale diffractive tensorized unit for high-throughput general-purpose computing. Instead of regulating in-core neurons, DTU takes novel near-core modulations to rearrange spatial-distributed neuron weights into temporal domain, iteratively combining input data with recurrent neuron weights via high-speed modulators. Such a near-core modulation allows dynamic reconfiguration of diffractive computing to perform universal large-scale matrix multiplications, and endows DTU with arbitrary network structures across various AI tasks.Leading to a giant leap in scalability, efficiency, and generalizability, the proposed methods pave a viable route to real-world large-scale photonic computing and represents an important step for next generation high-capacity high-efficiency AI chips, supporting applications in beyond-billion-neurons foundation models, unmanned-system edge computing, etc.
