Abstract
We thoroughly investigate the performance of the Dynamic Memdiode Model (DMM) when used for simulating the synaptic weights in large RRAM-based cross-point arrays (CPA) intended for neuromorphic computing. The DMM is in line with Prof. Chua’s memristive devices theory, in which the hysteresis phenomenon in electroformed metal-insulator-metal structures is represented by means of two coupled equations: one equation for the current-voltage characteristic of the device based on an extension of the quantum point-contact (QPC) model for dielectric breakdown and a second equation for the memory state, responsible for keeping track of the previous history of the device. By considering ex-situ training of the CPA aimed at classifying the handwritten characters of the MNIST database, we evaluate the performance of a Write-Verify iterative scheme for setting the crosspoint conductances to their target values. The total programming time, the programming error, and the inference accuracy obtained with such writing scheme are investigated in depth. The role played by parasitic components such as the line resistance as well as some CPA’s particular features like the dynamical range of the memdiodes are discussed. The interrelationship between the frequency and amplitude values of the write pulses is explored in detail. In addition, the effect of the resistance shift for the case of a CPA programmed with no errors is studied for a variety of input signals, providing a design guideline for selecting the appropriate pulse’s amplitude and frequency.
Highlights
The Matrix-Vector Multiplication (MVM) method is a key piece for computation in Artificial Neural Networks (ANNs), which have demonstrated outstanding results in the field of pattern recognition and classification, among others [1]
We demonstrate that a novel approach, compatible with the previous one but fully time-dependent, can be used for the SPICE simulation of large-scale memristor-based cross-point arrays (CPA) intended for pattern recognition tasks without significantly increasing the associated computational cost
We have demonstrated the viability of the Dynamic Memdiode compact Model (DMM) for realistic SPICE simulations of large Resistive Random Access Memory (RRAM)-based Cross-Point Arrays (CPA) intended for neuromorphic applications
Summary
The Matrix-Vector Multiplication (MVM) method is a key piece for computation in Artificial Neural Networks (ANNs), which have demonstrated outstanding results in the field of pattern recognition and classification, among others [1]. Hu et al.[5] reported a simulation-based case study of character recognition using two CPAs with 256 × 26 synapsis (i.e., 256 rows by 26 columns, totaling ∼13 k devices) to represent both the positive and negative weights using a Verilog-A nonlinear memristor model [16]. Aiming to reduce both the area and power consumption which arises from having two CPAs, an alternative architecture was considered by Truong et al.[12] (64 × 26, ∼1.6 k devices) using the same memristive device model. This model was successfully tested in voice recognition using a set of CPAs summing up to ∼2.5 k memristors [13]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
