The team of Zou Weiwen from the School of Electronic Information and Electrical Engineering of Shanghai Jiaotong University has made important progress in photonic tensor processing chips

Recently, the team of Professor Zou Weiwen of the Department of Electronic Engineering, School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, proposed innovative ideas at the intersection of photonics and computational science, and developed a new photonic tensor processing chip to realize high-speed tensor convolution operations, and the related results are based on "High-order tensor flow processing using integrated photonic circuits" (Advanced tensor flow processing based on integrated photonic chips) was published in the journal Nature Communications
.

Research background

The tensor form composed of multidimensional data stacking is an efficient form of data processing, which is conducive to exploring the intrinsic structural characteristics of data, and is widely used
in radar, communications, artificial intelligence, life sciences and other fields.
With the development of information technology in the future, the speed of data generation has exploded, and the multi-dimensional stacking of massive data and its efficient and rapid processing have become important scientific challenges
at present.
In order to meet this challenge, traditional electrical processors usually use the Generalized Matrix Multiplication (GeMM) strategy to convert high-order tensor operations into matrix multiplication operations, convert multi-level nested loop operations into parallel operations, and increase the overall speed
of tensor processing by increasing the number of computing cores.
However, the GeMM strategy relies on a large number of data replication, requires additional memory occupation and repeated communication overhead between memory and processor, which is one of the core bottlenecks to improve the speed of
multidimensional data tensor operation.

Innovative results

This research innovation puts forward the interdisciplinary research idea of constructing tensor operation process based on photon integration method, which can not only give full play to the broadband high-speed characteristics of optics to increase the computing clock frequency to tens of GHz, but also use the multi-degree of freedom of optics to directly characterize different dimensions
of tensor data.
This idea eliminates the need for tensor-to-matrix conversion, and realizes flow computing
from input tensors to output tensors.

Principle architecture of photonic tensor processing chips

Based on this innovative idea, the team designed and developed a photonic tensor processing chip, which comprehensively used the three degrees of freedom of wavelength, space and delay of optics, and successfully verified the high-speed tensor convolution operation with a clock frequency of 20 GHz on a multi-channel image, with a chip computing density of 588 GOPS/mm2, and the integration scale of photonic devices is expected to reach more than
1 TOPS/mm2 。 The research team used the chip to build a convolutional neural network for video action recognition, and the convolutional layer in the network was completed on the photon tensor processing chip, and finally achieved a recognition accuracy of 97.
9% on the KTH video dataset, which was close to the ideal recognition accuracy of 98.
9%.

The results of this study show that photonic integrated chips can realize tensor streaming at ultra-high clock frequencies, solve the problem of additional memory occupation and memory access, and provide a new technical approach
for the construction of advanced information systems such as high-performance computing and broadband signal processing.

Photon tensor processing chips

Multi-channel image convolution calculation results

Other relevant achievements

Professor Zou Weiwen's team carried out in-depth research in the direction of photonics and computational science in the early stage, and the formed photoelectric fusion processing principle and photoelectric hybrid integration method provided an important foundation
for the successful development of photonic tensor processing chips.
Recently, a series of research has been carried out on the core principles of photonic computing, innovative architecture and advanced application verification, and Opt.
Express 2022, 30(23): 42057, Opt.
Lett.
2022, 47(24): 6409 and other results
。 Related research has also promoted the smooth development of the Microcomb-based integrated photonic processing unit (the work is led by the team of Professor Wang Xingjun of Peking University and participated by Professor Zou Weiwen's team), and the relevant research results have also been accepted and published in the journal Nature Communications volume 14, Article number : 66 (2023)
。

Dissertation information

From left to right: Xu Shaofu, Wang Jing, Yi Sicheng, Zou Weiwen

Shanghai Jiao Tong University independently completed the study, with Xu Shaofu, assistant professor of the Intelligent Microwave Light Fusion Innovation Center of the Department of Electronic Engineering, Wang Jing, a doctoral student, as the co-first author, and Professor Zou Weiwen as the only corresponding author
.

Funding Information: This research work was supported
by the National Natural Science Foundation of China (T2225023), the Youth Fund Project (62205203), and the National Key R&D Program of China (2019YFB2203700).

Link to paper: style="text-indent: 2em; text-align: justify;">Nature Communications is an internationally peer-reviewed, multidisciplinary open access journal dedicated to publishing high-quality research in various fields such as biology, medicine, health, physics, chemistry, and earth sciences, with an impact factor of 17.
694
.