A Multi-Mode Visual Recognition Hardware Accelerator for AR/MR Glasses

• Published in: IEEE International Conference on Circuits and Systems (Online) pp. 1 - 5
• Main Authors: Yunhui Zhu, Yaohua Zuo, Tong Zhou, Guoping Fan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.05.2018
• Subjects: AR/MR glasses; Arrays; Bandwidth; Clocks; Engines; Hardware; hardware accelerator; multi-mode; Neural networks; PE utilization rate; visual recognition; Visualization
• ISSN: 2379-447X
• DOI: 10.1109/ISCAS.2018.8350918

A multi-mode visual recognition hardware accelerator for AR/MR glasses is designed in this paper. The accelerator supports state-of-the-art deep neural networks, including DNN, CNN, RNN and LSTM. To achieve higher utilization rate of computational components, the accelerator supports two mapping modes of neural networks to physical computational structures in a single PE (Processing Engine) array: a) 2D systolic flow of both filter and image data, b) each neural network output maps to a PE. The accelerator adaptively chooses the more efficient mapping mode layer by layer, achieving a higher PE utilization rate than single mapping mode accelerators. When benchmarking with Inception-v4 network, the accelerator's PE utilization rate is 79.5%, which is 20 points higher than state-of-the-art embedded accelerators. Higher PE utilization rate contributes to lower latency, higher throughput, less power consumption and smaller chip area. When benchmarking with AlexNet, the accelerator processes 108 images per second, which fully meets the real-time requirement for AR/MR applications.