Recently, the research group led by Associate Professor Fang Lu from the Department of Electronic Engineering and Academician Dai Qionghai from the Department of Automation at Tsinghua University abandoned the traditional paradigm of electronic deep computation, pioneering a distributed breadth intelligent optical computing architecture, and developed the world's first large-scale interference diffraction heterogeneous integrated chip named "Tai Chi", achieving a universal intelligent computation of 160 TOPS/W. This research achievement was published in the latest issue of "Science" on April 12th Beijing time.
As one of the three pillars of artificial intelligence, computing power is crucial for training AI models and performing inference tasks. If we liken large models to exquisite dishes, computing power is akin to a set of handy cooking utensils.
Optical computing, as the name suggests, converts the computing carrier from electricity to light, utilizing the propagation of light within the chip for computation. With its ultra-high parallelism and speed, optical computing is considered one of the most powerful competitive solutions for future disruptive computing architectures.
Optical chips possess advantages in high-speed and high-parallel computing, holding promise for supporting advanced AI applications such as large models.
Intelligent optical computing, as an emerging computing paradigm, demonstrates the potential to surpass silicon-based electronic computing in the post-Moore era. However, its computing tasks are limited to simple character classification, basic image processing, and the like. The pain point lies in the fact that the computing advantage of light is trapped in unsuitable electronic architectures, limiting the scale of computation and the support for complex large-model intelligent computing requiring high computing power and efficiency.
Different from electronic neural networks that rely on network depth to achieve complex computation and functions, the architecture of the "Tai Chi" optical chip originates from the unique 'fully connected' and 'high parallel' attributes of optical computing, transforming deep computation into distributed breadth computation. This exploration of a new path enables scalable, highly parallel, and robust general intelligent optical computing systems.
According to Xu Zhihao, the first author of the paper and a doctoral student in the Department of Electronic Engineering at Tsinghua University, in the "Tai Chi" architecture, the top-down encoding-splitting-decoding-reconstruction mechanism simplifies complex intelligent tasks by splitting them into multiple channels of highly parallel subtasks. The constructed distributed 'large receptive field' shallow optical network addresses subtasks separately, overcoming the inherent computational errors of multi-layer deep cascade physical simulator devices.
Title: Transforming Depth into Breadth: Taichi Distributed Breadth Optical Computing Architecture
Inspired by the classic Chinese text, the "Book of Changes," with its famous quote "There is Taiji which gives rise to two polarities," the team has developed an interference-diffraction joint propagation model. This model integrates the massive parallel advantages of diffraction optical computing with the flexible reconstruction characteristics of interference optical computing. It achieves partial/whole reconstruction and reuse of diffraction encoding/decoding and interference feature computation. By breaking through throughput bottlenecks with temporal reuse, it supports the bottom-up construction of a distributed breadth optical computing architecture, paving the way for exploring new avenues in on-chip large-scale general-purpose intelligent optical computing.
In simpler terms, the combination of interference and diffraction is akin to playing with LEGO bricks. LEGO blocks can be joined together through the interlocking of one module's groove with another module's protrusion. In the eyes of the research team, once interference and diffraction are transformed into basic modules and reconstructed for reuse, one can use rich imagination to build an infinite variety of configurations.
Team members posing for a project group photo.
According to the paper, the "Taiji" optical chip boasts an area efficiency of 879 T MACS/mm2 and an energy efficiency of 160 TOPS/W, enabling complex AI tasks such as recognizing thousands of objects in natural scenes and generating cross-modal content through optical computing for the first time.
The "Taiji" optical chip holds promise in providing computational power for large-scale model training and inference, general artificial intelligence, and autonomous intelligent unmanned systems.
Link to the relevant paper: https://doi.org/10.1126/science.adl1203
