Convolutional Window-Inspired Similarity-Aware Computation-in-Memory for Energy Saving

Abstract

Various data-driven computation-in-memory (CIM) architectures have been proposed to reduce inference energy. However, most data-driven CIM architectures require specific conditions to achieve energy savings (e.g., zero skip requires a ReLU activation function). This letter proposes a convolutional window-inspired similarity-aware CIM that saves energy by predicting the current output based on the previous one, which is applicable in most cases where the neural network is based on convolution. In addition, this letter introduces a novel transposable architecture to enhance linearity and an analog-to-digital converter (ADC) for improved area efficiency. The prototype was fabricated with 65 nm process and achieved the highest SWaP FoM as 19.04 TOPS/W $\times $ Mb/mm2 among the state-of-the-art transposable CIMs.