Innovative Breakthrough in AI Chip Development by Tsinghua University

On the 11th, it was reported by Tsinghua University that, in response to the challenges of large-scale optoelectronic intelligent computing, Associate Professor Fang Lu’s research team from the Department of Electronic Engineering and Academician Dai Qionghai’s research team from the Department of Automation have broken away from the traditional paradigm of electronic deep computation. They have pioneered a distributed breadth-wise optical computing architecture, and developed a large-scale interferometric-diffraction heterogeneous integrated chip named Tai Chi, achieving a universal intelligent computing capacity of 160 TOPS/W.

In today’s era of booming general artificial intelligence with large models, this research achievement, utilizing the way of photons, presents new inspirations, architectures, and pathways for exploring high-performance computing resources. This relevant research output has been published in the latest issue of the international journal “Science”.

Transformation from “Deep” to “Wide”: Distributed Breadth-wise Optical Computing Architecture.

Intelligent optical computing, as a nascent computing mode, has shown performance and potential far beyond silicon-based electronic computing in the post-Moore era. However, its computational tasks are limited to simple character classification and basic image processing. The pain point lies in the fact that the computational advantages of light are trapped in unsuitable electronic architectures, limiting computational scale and failing to support the urgent need for complex large model intelligent computing with high computational power and efficiency.

In an interview with the Science and Technology Daily, Fang Lu mentioned, “In contrast to the deep calculation and functions achieved by electronic neural networks relying on network depth, the architecture of the Tai Chi chip stems from the unique ‘fully connected’ and ‘high parallelism’ attributes of optical computing, transforming deep computation into breadth-wise computation. We have established a top-down coding-splitting-decoding-reconstructing mechanism, breaking down complex tasks into multi-channel high-parallel sub-tasks and constructing a distributed ‘large receptive field’ shallow optical network to divide and conquer sub-tasks, breaking through the inherent errors of multi-layer cascaded analog computation.”

Fang Lu further explained, “Inspired by the ‘Tai Chi gives birth to yin and yang’ concept from the Book of Changes, during the chip development process, we established an on-chip interferometric-diffraction joint propagation model, integrating the advantages of large-scale parallelism in diffractive optical computing and the flexible reconstruction characteristics of interferometric optical computing. We broke through the bottleneck of throughput through temporal reuse, exploring a new pathway for on-chip large-scale universal intelligent optical computing.”

Yin and Yang as One: Interferometric-Diffraction Fusion Computing Chip.

It is reported that the computing efficiency of the Tai Chi optical chip exceeds existing intelligent chips by 2 to 3 orders of magnitude, providing computational support for high-speed intelligent analysis of scenes with billions of pixels, inference of large models with billions of parameters, and low-power autonomous intelligent unmanned systems at the milliwatt level.