Energy-Efficient Time-Domain Vector-by-Matrix Multiplier for Neurocomputing and Beyond

TL;DR

This paper introduces a highly energy-efficient time-domain vector-by-matrix multiplier suitable for large-scale neurocomputing, utilizing time-encoded signals and current sources to minimize static power consumption.

Contribution

It presents a novel mixed-signal, time-domain approach for vector-by-matrix multiplication that reduces energy use and allows scalable chaining for large neural networks.

Findings

Achieved ~6-bit precision limited by drain-induced barrier lowering.

Estimated energy consumption of ~7 fJ per operation for large matrices.

Demonstrated potential for sub-1 fJ/Op energy efficiency with optimized design.

Abstract

We propose an extremely energy-efficient mixed-signal approach for performing vector-by-matrix multiplication in a time domain. In such implementation, multi-bit values of the input and output vector elements are represented with time-encoded digital signals, while multi-bit matrix weights are realized with current sources, e.g. transistors biased in subthreshold regime. With our approach, multipliers can be chained together to implement large-scale circuits completely in a time domain. Multiplier operation does not rely on energy-taxing static currents, which are typical for peripheral and input/output conversion circuits of the conventional mixed-signal implementations. As a case study, we have designed a multilayer perceptron, based on two layers of 10x10 four-quadrant vector-by-matrix multipliers, in 55-nm process with embedded NOR flash memory technology, which allows for compact implementation of adjustable current sources. Our analysis, based on memory cell measurements, shows that at high computing speed the drain-induced barrier lowering is a major factor limiting multiplier precision to ~6 bit. Post-layout estimates for a conservative 6-bit digital input/output NxN multiplier designed in 55 nm process, including I/O circuitry for converting between digital and time domain representations, show ~7 fJ/Op for N>200, which can be further lowered well below 1 fJ/Op for more optimal and aggressive design.