4K-memristor analog-grade passive crossbar circuit

Abstract

The superior density of passive analog-grade memristive crossbar circuits enables storing large neural network models directly on specialized neuromorphic chips to avoid costly off-chip communication. To ensure efficient use of such circuits in neuromorphic systems, memristor variations must be substantially lower than those of active memory devices. Here we report a 64 × 64 passive crossbar circuit with ~99% functional nonvolatile metal-oxide memristors. The fabrication technology is based on a foundry-compatible process with etch-down patterning and a low-temperature budget. The achieved <26% coefficient of variance in memristor switching voltages is sufficient for programming a 4K-pixel gray-scale pattern with a <4% relative tuning error on average. Analog properties are also successfully verified via experimental demonstration of a 64 × 10 vector-by-matrix multiplication with an average 1% relative conductance import accuracy to model the MNIST image classification by ex-situ trained single-layer perceptron, and modeling of a large-scale multilayer perceptron classifier based on more advanced conductance tuning algorithm.

Applications and Techniques for Fast Machine Learning in Science

“…In this section, the most prominent emerging technology proposals, including those based on emerging dense analog memory device circuits, are grouped according to the targeted low-level neuromorphic functionality - see, e.g., reviews in Burr et al ( 2017 ), Bavandpour et al ( 2018 ), Yang et al ( 2013 ), and Yu ( 2018 ) and original work utilizing volatile (Ohno et al, 2011 ; Pickett et al, 2013 ; Chu et al, 2014 ; Sheridan et al, 2017 ; Wang et al, 2017 , 2018c ; Adda et al, 2018 ; Lashkare et al, 2018 ; Zhang et al, 2018b ; Cai et al, 2019a ; Yeon et al, 2020 ) and nonvolatile (Mahmoodi et al, 2009 , 2019 ; Alibart et al, 2012 ; Govoreanu et al, 2013 ; Prezioso et al, 2015 , 2016 , 2018 ; Li et al, 2016a ; Adam et al, 2017 ; Pedretti et al, 2017 ; Bayat et al, 2018 ; Hu et al, 2018b ; Wang et al, 2018c ; Kim et al, 2019 ; Cai et al, 2020a ; Lin et al, 2020b ; Liu et al, 2020b ; Yao et al, 2020a ) memristors, phase change memories (PCM) (Kuzum et al, 2011 ; Burr et al, 2015 ; Tuma et al, 2016 ; Ambrogio et al, 2018 ; Ríos et al, 2019 ; Joshi et al, 2020 ; Karunaratne et al, 2020 ), and nonvolatile NOR (Bayat et al, 2015 ; Guo et al, 2017a , b ; Mahmoodi et al, 2019 ), and NAND (Bavandpour et al, 2019 , 2020 ; Lee et al, 2019 ), and organic volatile (Fuller et al, 2019 ) floating gate memories, as well as multiferroic and spintronic (Sengupta et al, 2016 ; Ni et al, 2018 ; Ostwal et al, 2018 ; Romera et al,…”

SDEX: Monte Carlo Simulation of Stochastic Differential Equations on Memristor Crossbars

“…Because fabrication techniques are still in development for such devices, the resolution of analog memristive devices is quite low compared to conventional computers. The low "resolution" is due to limited programming tuning accuracy, with 99% of memristors within 4% of the target conductance in the most advanced crossbars [12].…”

Nonideality-Aware Training for Accurate and Robust Low-Power Memristive Neural Networks

“…Each weight is typically made proportional to the difference of G + and G − (with k G acting as the constant of proportionality) which enables to encode any real number within a finite interval. However, infinite conductance combinations will produce the same difference [32], thus the network designer may have to make an arbitrary choice of how to perform this mapping. For example, to encode weights w ∈ w, the two conductances may be picked symmetrically around the average value [32], as shown in Equation 4.…”

“…However, infinite conductance combinations will produce the same difference [32], thus the network designer may have to make an arbitrary choice of how to perform this mapping. For example, to encode weights w ∈ w, the two conductances may be picked symmetrically around the average value [32], as shown in Equation 4.…”