Mixed-signal Dot-product Processor with Switched-Capacitors for Machine Learning

Abstract

A mixed-signal dot-product computation has growing use cases in embedded sensory systems and emerging computing platforms (such as in-memory or neuromorphic computings) for the ultra low-power implementation of machine learning (ML) algorithms. This paper proposes a compact and energy-efficient mixed-signal dot-product circuit with switched capacitors which has an analog input as one operand and a digital input as the other operand. The proposed dot product processor requires only two unit-sized capacitors per multiplication, thereby highly energy- and area- efficient. The proposed processor also supports flexible input bit-precision without any hardware overhead by simply iterating more cycles to provide higher bit precision. The simulated results for the proposed circuit designed in a 14nm CMOS show 10.5 and 7.9 improvements in energy efficiency and computation delay, respectively, compared to a conventional switched-capacitor based implementation while maintaining 9 bit output resolution.