SNPU: Always-on 63.2μW Face Recognition Spike Domain Convolutional Neural Network Processor with Spike Train Decomposition and Shift-and-Accumulation Unit

Recently, always-on face recognition (FR) and action recognition chips are widely developed in battery-limited mobile devices for event detection [1]. CNNs with high accuracy are unable to realize <100μW ultralow-power inference because many fixed numbers of operations should always process the 2D inputs [2]. Neural networks with low precision weight such as BNN and XOR-Net were proposed to realize low power operation but their accuracy is too low to compare with CNN [3], [4]. Replacement of multiplication with simple operations such as ShiftNet and AdderNet were proposed to achieve low power but they required very high input/weight precisions [5], [6].