With the emergence of various large models and deep neural networks, how to manufacture the next generation of AI chips that meet the development of artificial intelligence, have both high computing power and high energy efficiency has become an international frontier hotspot. The research group of Associate Professor Fang Lu from the Department of Electronic Engineering at Tsinghua University and Academician Dai Qionghai from the Department of Automation have abandoned the traditional paradigm of electronic deep computing and taken a new approach. They have pioneered the development of the world’s first large-scale interferometric diffraction heterogeneous integrated chip, Taichi, based on a distributed breadth intelligent optical computing architecture, achieving universal intelligent computing of 160TOPS/W. The research findings were published in the latest issue of Science on April 12th Beijing time.CHIP

Optical computing, as the name suggests, is the transformation of computing carriers from electricity to light, utilizing the propagation of light in chips for computation. With its ultra-high parallelism and speed, it is considered one of the most powerful competitive solutions for future disruptive computing architectures. However, its computational tasks are limited to simple character classification, basic image processing, and so on. The pain point is that the computing advantage of light is trapped in an unsuitable electrical architecture, with limited computing scale and inability to support complex large-scale intelligent computing that urgently needs high computing power and energy efficiency.

According to Xu Zhihao, the first author of the paper and a doctoral student in the Department of Electronics, in the “Tai Chi” architecture, a top-down encoding splitting decoding reconstruction mechanism simplifies complex intelligent tasks into multi-channel and highly parallel subtasks. The distributed “large receptive field” shallow optical network is constructed to divide and conquer the subtasks, breaking through the inherent computational error of multi-layer deep cascading in physical simulators.

The team was inspired by the classic book of the Book of Changes, “Yi has Tai Chi, it is the birth of two instruments.” They established an interference diffraction joint propagation model, integrating the advantages of large-scale parallel diffraction light calculation with the flexible reconstruction characteristics of interference light calculation. They partially/comprehensively reconstructed and reused diffraction encoding and decoding with interference feature calculation, breaking through the flux bottleneck through time-series multiplexing, and supporting the distributed breadth light computing architecture from bottom to top, exploring new paths for large-scale intelligent light computing on chip.

It is understood that the “Tai Chi” optical chip has an area efficiency of 879TMACS/mm ^ 2 and an energy efficiency of 160TOPS/W, empowering light computing for the first time to achieve artificial intelligence complex tasks such as natural scene object recognition and cross modal content generation. The “Tai Chi” optical chip is expected to provide computing power support for large-scale model training and inference, general artificial intelligence, and autonomous intelligent unmanned systems.