[A-10] Integrated Photonic Devices

[A-10-06] Experimental Demonstration of a Novel Self-adaptive Photonic Tensor Core on the SOI Platform with Noise Suppression towards High Precision and Computational Efficiency

〇Rui Shao¹, Gong Zhang¹, Xuanqi Chen¹, Haibo Wang¹, Yue Chen¹, Yuxuan Wang¹, Xiao Gong¹ (1. National Univ. of Singapore (Singapore))

Presentation style: On-site (in-person)

https://doi.org/10.7567/SSDM.2022.A-10-06

We report the first self-adaptive photonic tensor core employing noise suppressing structures using cascaded programmable Mach-Zehnder modulators (MZMs) on the standard 220 nm SOI photonic platform. We take advantage of the waveguide crossing structures to help achieve a significant reduction of errors when the number of MZMs raises. We further propose and utilize the self-adaptive mechanism to realize a ±0.5 dB dynamic control. By using our self-adaptive photonic tensor core to realize the deep convolution neural network (DCNN) with 15 convolutional layers, we demonstrate the lowest accuracy degradation (~0.2%) with high computational efficiency (~333 FLOP/W) among reported silicon photonic hardware.