Scaled resistively-coupled VO2 oscillators for neuromorphic computing

Elisabetta Corti,Bernd Gotsmann,Kirsten Moselund,Adrian M Ionescu,John Robertson,Siegfried Karg

doi:10.1016/j.sse.2019.107729

Abstract

New computation schemes inspired by biological processes are arising as an alternative to standard von-Neumann architectures, to provide hardware accelerators for information processing based on a neural networks approach. Systems of frequency-locked, coupled oscillators are investigated using the phase difference of the signal as the state variable rather than the voltage or current amplitude. As previously shown, these oscillating neural networks can efficiently solve complex and unstructured tasks such as image recognition. We have built nanometer scale relaxation oscillators based on the insulator–metal transition of VO2. Coupling these oscillators with an array of tunable resistors offers the perspective of realizing compact oscillator networks. In this work we show experimental coupling of two oscillators. The phase of the two oscillators could be reversibly altered between in-phase and out-of-phase oscillation upon changing the value of the coupling resistor, i.e. by tuning the coupling strength. The impact of the variability of the devices on the coupling performances are investigated across two generations of devices.

Highlights

Complex and unstructured problems like speech or image recognition are currently solved most effectively on by deep learning algorithms running on specialized electronic circuits such as graphical processing units (GPUs) [1]
In this paper we investigated the properties of VO2 coupled oscillator fabricated on a Si substrate
We experimentally showed frequency and phase-locking of VO2 on Si resistively-coupled oscillators

Summary

Introduction

Complex and unstructured problems like speech or image recognition are currently solved most effectively on by deep learning algorithms running on specialized electronic circuits such as graphical processing units (GPUs) [1]. In [17] it has been demonstrated that it is possible to build a network that is suitable for performing tasks as image recognition using resistances as coupling elements for the oscillators, envisioning implementation of the coupling with resistive RAM [18] This would bring to the network the advantage of having reconfigurable weights on chip, allowing online learning of the network. The best performing oscillators operate at a maximum frequency of 9 MHz [19] and operate at a scaled voltage of 1 V with a power consumption around 10 μW [20] These results refer to devices build on TiO2 substrate, that being lattice-matched to VO2, allows deposition of crystalline material. The effect of the realization of these devices on Si is explored: it is shown how the variability of the devices resulting from granular films influences the value of the coupling element that is necessary to use to bring the oscillators at the frequency locking condition, and how this impact on the performances of the network

Fabrication

Experimental results

Conclusion