Ultra-Low Power Neural Network Processors using Analog-Based Computation

In this paper, two types of low power and energy efficient neural network (NN) inference processors are introduced. First, an energy efficient accelerator for convolutional neural networks (CNNs) using lookup tables (LUTs) based on the vector quantization (VQ) technique is shown. VQ requires a nearest neighbor search (NNS), resulting in a large energy overhead. To overcome this issue, a timedomain associative memory (TAM) using analog-based computation is proposed. Thanks to energy efficient NNS by TAM, the proposed LUT-based CNN accelerator achieves 7.5 times higher efficiency compared with the conventional accelerators. Then, an ultra-low power NN-based audio feature extractor for keyword spotting (KWS) is demonstrated. In the conventional KWS processors, the large portion of the power of the KWS system is consumed in an analog-to-digital converter (ADC) for feature extraction. To tackle with this issue, a NN-based feature extractor is proposed in this work. It extracts audio features in analog-domain, which can remove the powerhungry ADC. Consequently, the power dissipation of the feature extraction can be reduced by $-88 \%$.