Variability-Aware Memristive Crossbars—A Tutorial

Abstract

Memristor crossbar architecture is one of the most popular circuit configurations due to its wide range of practical applications. The crossbar architecture can emulate the weighted summation operation, called multiply and accumulate operation (MAC). The errors to MAC computing get introduced due to a range of crossbar variability. We broadly group the variability in three categories: (1) device-to-device variations, (2) programming nonlinearity, and (3) those from peripheral circuits. This tutorial provides insights into the variability and compensation approaches that can be adopted to reduce its impact when designing for practical applications with crossbars. .

PyMem: A Graphical User Interface Tool for Neuromemristive Hardware–Software Co-Design

“…Memristors [1] represent a broad class of resistive switching devices that when used in a crossbar array can efficiently emulate the multiply and accumulate (MAC) [2] operation required for dot-product computations. However, in practice, memristor devices are prone to a wide range of device-todevice and cycle-to-cycle variability [3]. In addition, when implemented in a crossbar architecture [4], the performance of the system is subject to various circuit parasitic and noise.…”

“…which maps with each column the current outputs of the crossbar. Thus, each crossbar column can be considered a neuron computation, and the crossbar as a single neuron layer [3], [6], [7]. Using this concept, this study introduces a GUI to design such crossbar based neural networks [8].…”

“…temp.neural_weight([[1,2,3],[4,5,6]]) 2 new_weight-temp.neural_memristor_weights() Invoking functions to get corresponding memristance Algorithm 4 Algorithm for implementing mathematical model for memristors in python Start Parameter 1: R of f , R on , R init , Amplitude, frequency, time duration, sample rate, p, j, model #variabilities temp ← memristor_models(P arameters1) #call the main PyMem class Parameter 2: Quantization_value, percentage of variability R on , percentage of variability R of f temp.memristor_with_variability(P arameters2) #setting up the memristor with variability Parameter 3: current_weights #variabilities temp.neural_weight(P arameters3) #changing neural weights to memristor resistance value new_weight ← temp.neural_memristor_weights() #changing neural weights to memristor resistance value Output: new_weight End Python-based implementation for Applying a Mathematical Model to Implement Memristors is mentioned in Algorithm 4. The algorithm starts working by mapping the weight values from the neural network into quantization of G of f (1/R on ) and G of f (1/R of f ) of the ideal memristor.…”

Magnetic Flux Sensor Based on Spiking Neurons with Josephson Junctions

“…Nanoscale thin film (metal-insulator-metal) resistive switching devices were introduced as memristors only in 2008 [16]. Being compatible with complementary metal-oxide-semiconductor (CMOS) technology [17,18] is the main advantage of this type of device, while a major drawback is the large device-to-device and cycle-to-cycle variability [19]. In turn, the early mention of a JJ as a memristor dates back to 1974 [20].…”