Similarity-Aware Transposable SRAM Computation-In-Memory with Pulse-Based SAR ADC

Computation-in-memory (CIM) is a promising approach that aims to improve the energy efficiency in calculating the multiply-and-accumulate (MAC) of deep neural networks (DNNs) by performing computations directly in memory. However, the data-driven feature has not been explored enough, which could drastically improve energy efficiency. LowMAC first [7] and SS-SAR hybrid CIM [8] have been proposed considering bell-shaped MAC, but those works have a high detection fail rate as they did not consider statistical information such as correlation. This work found a high correlation between the preceding and the current MAC result due to the overlap of the convolutional window, as shown in Fig. 1 (top-left). This paper proposes similarity-aware CIM (SA-CIM) with three features: 1) similarity-aware ADC (SA-ADC) to exploit convolutional similarity, leading to a 31.5% ADC power reduction, 2) capacitor-based transposable bitcell, improving linearity and symmetricity, 3) pulse-based ADC to overcome dual ADC area overhead of transposable CIM, leading to 44.4% transistors reduction.