CUHK Professor Chaoran Huang Shares about Silicon Photonic Neural Networks for High-performance Computing and Machine Learning


Photonic communication has long been regarded as an effective means to alleviate the data movement bottleneck of conventional microelectronic processors. Prof. Chaoran Huang, Associate Professor, Department of Electronic Engineering, The Chinese University of Hong Kong (CUHK) paid a visit to UMEC today and gave a talk on the latest silicon photonics technologies.

As a fundamental bottleneck of existing AI hardware, poor interconnect performance leads to high power consumption and latency. To tackle the problems, Prof. Huang has developed a silicon photonic neural network which is perfect for interconnect, that can reduce power loss and increase bandwidth significantly. Her team emulated neurons and neural network model with photonic devices and circuits, enabling accurate training using conventional training algorithms and faster, more efficient execution of the neural model. The network allows simultaneous data movement across chips and matrix multiplication with low cost. It was successfully applied to different scenario, such as channel equalisation and real-time system control.

Although there are still many considerations for mass adoption of photonics in analog computing systems, such as high costs related to periphery, reprogramming weights during training, and A/D and D/A conversion, the higher-level communication protocols between multiple neuromorphic cores, etc., the successful research result proved it is a right track for future development. UMEC is interested in exploring its potential applications, and also looks forward to partnering with CUHK to co-develop innovative photonic and electronic coprocessors, which target the niche AI applications with considerable market potential.


Prof. Charoran Huang explained her research on silicon photonic neural networks and applications

UMEC is interested in adopting the neural network to their AI chips to improve the performance of deep learning training and inference