Performance Analysis of Convolutional Neural Network Using Multi-level Memristor Crossbar for Edge Computing
In this paper, the performance analysis of convolutional neural network (CNN) with multi-level memristor crossbar is presented. Multi-level memristor crossbar is used to implement the Vector-Matrix Multiplication (VMM), which is the most computationally intensive step in the CNN algorithm. A procedure to convert a classical CNN model with floating-point accuracy weights to finite-bit weights implemented with multilevel memristor is presented. The impacts of memristor levels, crossbar line resistance and crossbar array size to the VMM calculation accuracy and the CNN classification accuracy are analyzed in details. As an example, one converted CNN is tested with a down sampled 14x14 hand-writing digits dataset. Classification accuracy of 93% and 95% are achieved with 4-bit memristor crossbar and 5-bit memristor, respectively.
Paper
Full text
Performance Analysis of Convolutional Neural Network Using Multi-level Memristor Crossbar for Edge Computing
Semantic Scholar · Computer Science · 2020
Abstract
In this paper, the performance analysis of convolutional neural network (CNN) with multi-level memristor crossbar is presented. Multi-level memristor crossbar is used to implement the Vector-Matrix Multiplication (VMM), which is the most computationally intensive step in the CNN algorithm. A procedure to convert a classical CNN model with floating-point accuracy weights to finite-bit weights implemented with multilevel memristor is presented. The impacts of memristor levels, crossbar line resistance and crossbar array size to the VMM calculation accuracy and the CNN classification accuracy are analyzed in details. As an example, one converted CNN is tested with a down sampled 14x14 hand-writing digits dataset. Classification accuracy of 93% and 95% are achieved with 4-bit memristor crossbar and 5-bit memristor, respectively.