An Energy-Efficient Multibit ReRAM Computing-in-Memory Processor for Ultra-low-energy Keyword Spotting

Abstract

An energy-efficient resistive random-access memory (ReRAM)-based computing-in-memory (CIM) processor is proposed for the ultra-low-energy keyword spotting (KWS) system for mobile devices. Previous ReRAM CIM processors have faced several limitations, such as excessive computing energy caused by undesired current branches, significant computation non-linearity, and substantial throughput degradation due to analog-to-digital conversion (ADC). It also suffered from difficulties in efficiently supporting diverse operations, such as depthwise separable convolution, which further constrained its applicability in modern AI workloads. This paper introduces three key improvements to address these issues: 1) charge-mode ReRAM bitcells (CRB) eliminate DC branches during CIM operations, reducing computing energy by 61.39% and improving computation linearity by 1.65×; 2) additive powers-of-two (APoT) quantized ADC reduces analog-to-digital (A-to-D) conversion energy by 65.42% compared to previous ADCs, while maintaining conversion accuracy. It also shortens A-to- D conversion latency, achieving a 1.33× improvement in system throughput; 3) the layer fusion depthwise separable convolution cluster (DCC) and pipelined CIM architecture enhance overall system throughput by 1.35×. The proposed processor is designed in 45 nm CMOS technology and a compatible ReRAM device. It achieves 46.25 TOPS/W of macro energy efficiency and 22.59 TOPS/W of system energy efficiency. As a result, the processor realized 0.74 μJ inference energy and 92.7% of KWS accuracy at the Google Speech Commands dataset.