Invited Talk II
Prof. Juinn-Dar Huang
National Yang Ming Chiao Tung University, Taiwan
Ingredients of Efficient Hardware Accelerators for Neural Networks
Abstract
Today, neural networks (NNs) are broadly used for numerous artificial intelligence (AI) applications including computer vision, image/video processing, speech recognition, and natural language processing (NLP). Though NN-based algorithms can provide better solutions on several AI application domains, those advantages come at the cost of extremely high computational complexity. Currently, GPU-based computing engines are most commonly used platforms for NN computation. Nevertheless, they are pricey, power-hungry and therefore inappropriate for certain application areas, such as edge computing. Therefore, specialized hardware accelerators optimized for a specific class of NNs are getting more attention these days. This short talk aims to introduce some common ingredients of efficient NN hardware accelerators, including dataflow-based optimization, weight pruning and compression, quantization, and numeric data formats. Algorithm-level design considerations facilitating efficient hardware implementations are also discussed. Accelerators for convolutional neural networks (CNNs) and multilayer perceptron (MLP) are used as examples for demonstrations.