第十九届全国容错计算学术会议 (CFTC 2021)

When the neural network technology is applied to the battery-powered terminal equipment, the energy efficiency of its hardware calculation has become the key problem to be considered. In view of this, this paper designs and realizes a reconfigurable approximate multiplication architecture for CNN-Based speech recognition. First, a convolutional neural network reconfigurable computing cell structure is presented. Second, it is extended to the design and implementation of a low power precision controllable convolutional neural network, which includes the Wallace tree tensor multiplier unit and the design of an approximate compressor. As case study, the proposed approximate designs are applied to a CNN-based keywords speech recognition system. Under TSMC 22nm ULL UHVT process condition, compared with the speech keyword recognition system without approximate computation, the power consumption of the processing engine with approximate multiplication computation unit is reduced by 51.55%, while the recognition accuracy is reduced by only 1%.

Keywords Spotting (KWS); Approximate Multipliers; Low Power Circuits; Convolutional Neural Network