A 4-bit Mixed-Signal MAC Array with Swing Enhancement and Local Kernel Memory

Abstract

Modern deep neural networks require energy- and area-efficient multi-bit multiply-accumulate (MAC) functions. In-memory computing (IMC) with analog accumulation has shown the potential to outperform purely digital solutions but lacks efficient multi-bit computation. In this work, we explore the design of a mixed-signal, charge-sharing compute array that performs 4-bit MAC operations within one clock cycle. Its key features include 7×4 bit kernel memory in each MAC cell, as well as a per-kernel binary combiner that simplifies the unit cell design and enables efficient in-column ADC integration. In addition, it leverages a differential switching scheme that improves the signal swing by 4× relative to single-ended schemes, thereby reducing ADC energy. Post-layout simulations in 28 nm CMOS indicate an energy efficiency of 6.4 fJ/MAC and a compute density of 3.17 TOPS/mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> for an input activation vector size of 160.