A Resonant Time-Domain Compute-in-Memory (rTD-CiM) ADC-Less Architecture for MAC Operations

Abstract

In recent years, Compute-in-memory (CiM) architectures have emerged as a promising solution for deep neural network (NN) accelerators. Multiply-accumulate (MAC) is considered a de facto unit operation in NNs. By leveraging the minimal data movement required and inherent parallel processing capabilities of CiM, NNs that require numerous MAC operations can be executed more efficiently. Traditional CiM architectures execute MAC operations in the analog domain, employing an Analog-to-Digital converter (ADC) to digitize the analog MAC values. However, these ADCs introduce significant increase in area and power consumption, as well as introduce non-linearities. This work proposes a resonant time-domain CiM (rTD-CiM), an ADC-less architecture that reduces the power consumption of traditional CiM architectures with ADCs. The feasibility of the proposed architecture is evaluated on an 8KB SRAM memory array using TSMC 28 nm technology. The proposed rTD-CiM architecture demonstrates a throughput of 2.36 TOPS with an energy efficiency of 28.05 TOPS/W.