Fig. 4: Multibit switching characteristics. | Nature Communications

Fig. 4: Multibit switching characteristics.

From: Wafer-scale fabrication of memristive passive crossbar circuits for brain-scale neuromorphic computing

Fig. 4

a Plot of the current evolution for the memristor with the tuning algorithm. The top panel shows the programming pulses corresponding to 160 consecutive VSET and VRESET pulses at a fixed W = 80 μs. For the target G, the amplitude started from ±0.4 V and then was incrementally increased by 60 mV. When the current G approached the target, the incremental amplitude became smaller to 24 mV to perform fine tuning. To verify the current G after programming, VREAD was used with 0.2 V for 1 ms. Tuning error was set to be <5%. b Reproducibility of consecutive analog and multibit switching characteristics during 106 programming pulses. The pulse endurance test was performed based on the tuning algorithm with target ION = 10−5 A and IOFF = 2 × 10−6 A. The tuning error was set to be <5 %. The inset shows a magnified plot of the analog switching curve. c Retention capability of 16 G values with an interval of 1.75 μA (i.e., 4 bits) for 104 s. All G values were repetitively read at VREAD = 0.2 for 1 ms with Δt = 1 s. d, e Plots of 4-bit operations with respect to 16 memristors (d) and 30 cycles (e). The 16 distinguishable states were also clearly observed at VREAD in I–V curves (red dotted line in d), corresponding to 16 G values (the inset in d). The error bars in e are the standard deviation. f, g Plots of 5- (f) and 6-bit operations (g) with smaller tuning errors of 3 % and 1 %. The same VREAD was used. h Examples of fashion-item images (Shoes, T-shirt, and Trousers). Each image consists of 28 × 28 gray-scale pixels. i A 28 × 28 sub-array to be encoded by the three images (yellow dotted box) based on the tuning algorithm with W = 80 μs. The scale bar for (i) is 50 μm. j Resultant conductance maps visualizing the encoding results for the three fashion-item images. The purple box indicates the tuning failure that is typically caused by low DR (<3). k Operation results for estimated (blue bar) and experimental VMM (red dot). The small difference between the two results suggests the successful VMM operation in combination with the precise encoding using the tuning algorithm.

Back to article page