Comparison of Two Different Classes of Memristive Devices for Neural Network Inference Tasks
Felix Cüppers a, Stephan Aussen a, Rainer Waser a, Susanne Hoffmann-Eifert a
a Forschungszentrum Jülich GmbH, PGI 7 &10, Jülich, Germany
Proceedings of Neuronics Conference (Neuronics)
València, Spain, 2024 February 21st - 23rd
Organizers: Sabina Spiga and Juan Bisquert
Oral, Felix Cüppers, presentation 022
DOI: https://doi.org/10.29363/nanoge.neuronics.2024.022
Publication date: 18th December 2023

Memristive devices based on the valence change mechanism (VCM) are promising candidates for emerging memory and neuromorphic applications. Due to their two-terminal structure and the possibility to be arranged in large arrays, one of their largest potential is seen in inference-based tasks such as the operation of neural networks once the training is finished.

VCM-type memristive systems can be further classified based on the area scaling of the device conductance states. In particular, area-independent, i.e. filamentary-switching, devices and area-type switching devices exist. Recent advances in the understanding of the different conductance mechanisms [1] and of the area-dependent switching variety in particular [2] have identified this second class of VCM-type memristive devices a new prospect for aforementioned inference tasks. Due to their area-scaling conductance values and the improved signal-to-noise ratio over filamentary-switching devices [3], they appear to be the superior choice in this context. In contrast, area-dependent devices frequently suffer from slower switching speeds at increased switching voltage and limited retention times compared to the filamentary class.

In this work, the two classes are compared with one another in the context of an inference-based task applied on two typical candidates, namely Pt/HfO2/TiOx/Ti cells for the filamentary and Pt/Al2O3/TiOx/Pt devices for the area-dependent class. Their conduction mechanism, switching kinetics, noise and retention properties are determined experimentally and fed into the simulation of a single-layer neural network for classification of the MNIST dataset. A comprehensive comparison highlights advantages and drawbacks of the device choice. Finally, strategies for exploiting the respective advantages while mitigating the respective drawbacks are evaluated, and a synergetic approach is proposed.

References

[1]      

[2]      

[3]      

 

This work was in part funded by the German Research Foundation (DFG) under Grant No. SFB 917, in part by the Federal Ministry of Education and Research (BMBF, Germany) in the projects NEUROTEC (Project Nos. 16ME0398K and 16ME0399) and NeuroSys (Project No. 03ZU1106AB) and is based on the Jülich Aachen Research Alliance (JARA-FIT).

© FUNDACIO DE LA COMUNITAT VALENCIANA SCITO
We use our own and third party cookies for analysing and measuring usage of our website to improve our services. If you continue browsing, we consider accepting its use. You can check our Cookies Policy in which you will also find how to configure your web browser for the use of cookies. More info