Minimalist Photonic Processor for High‐Volume and Versatile Linear Computation
Zhenhua Li, Zhaoang Deng, Jie Liu, Chuyao Bian, Jiaqing Li, Ziliang Ruan, Ranfeng Gan, Zihao Chen, Kaixuan Chen, Changjian Guo, Liu Liu, Siyuan Yu · Laser & Photonics Review · 2025
Abstract By fully exploiting the rich parameter dimensions of the light wave including time, wavelength, transverse space, or mode, photonic integrated circuits potentially offer low‐latency, high‐throughput, and energy‐efficient solutions for acceleration of multimodal linear data processing in artificial intelligence‐related computational tasks. However, many existing schemes tailor specific parameter dimensions and construct specific architectures to suit specific computational operations and, therefore not making full use of optical resources and lacking versatility in adapting to different operations. Their scale is often linked to that of the operands, therefore lack flexibility when dealing with variable data sizes. A novel multi‐dimensional minimalist photonic processor (MD‐MPP) architecture is demonstrated, capable of simultaneously and scalably utilizing time, wavelength, and space multiplexing to achieve high throughput, versatile operations, and flexible data adaption, performing all‐optical multiply‐and‐accumulate (MAC) operations for vector dot‐products, matrix‐vector‐multiplication, single‐/multi‐kernel convolution in time‐recursive, wavelength‐parallel and spatial‐parallel fashions. As a verification, a processor chip fabricated in thin‐film lithium niobate (TFLN) experimentally implements single‐/multi‐kernel and multi‐wavelength convolution in optoelectronic convolutional neural networks with up to 36.7 billion MAC operations per second (or 73.4 GOPS) per device per wavelength, underscoring its potential to be a promising candidate for flexible optical computing at high data volumes with lower energy consumption.