On-chip Phase Change Optical Matrix Multiplication Core
Xuan Li, Nathan Youngblood, Wen Zhou, Johannes Feldmann, Jacob L. Swett, Samarth Aggarwal, Abu Sebastian, C. David Wright, Wolfram H. P. Pernice, Harish Bhaskaran
Abstract
We demonstrated a non-volatile photonic matrix computation core which contains a 3×3 photonic phase change in-memory computing matrix to carry out matrix vector multiplication on a silicon-based substrate. We established an optical computing core model and hardware implementation based on our photonic phase change material (PCM) devices and fabricated a 3×3 matrix using the silicon photonic foundry. We demonstrated the functionality of this matrix as the linear convolution layer in a convolutional neural network (CNN) and demonstrated simple pattern recognition. Finally, we simulated scaling up matrix limits using experimental data from smaller matrices. This on-chip, non-volatile, photonic computation matrix transfers optical computing from single device to matrix, paving the way for practical wide usage of optical computing systems.