CoChat

Application of the Quasi-Static Memdiode Model in Cross-Point Arrays for Large Dataset Pattern Recognition

Open access
Aguirre, Fernando Leonel, Pazos, Sebastian Matias, Palumbo, Felix, Sune, Jordi, Miranda, Enrique · 2020 · IEEE Access

The DOI matches its Crossref record; 98 of the 100 references it cites were found in a registry or an index.

Abstract

We investigate the use and performance of the quasi-static memdiode model (QMM) when incorporated into large cross-point arrays intended for pattern classification tasks. Following Chua's memristive devices theory, the QMM comprises two equations, one equation for the electron transport based on the double-diode circuit with single series resistance and a second equation for the internal memory state of the device based on the so-called logistic hysteron or memory map. Ex-situ trained memdiodes with different MNIST-like databases are used to establish the synaptic weights linking the top and bottom wire networks. The role played by the memdiode electrical parameters, wire resistance and capacitance values, image pixelation, connection schemes, signal-to-noise ratio and device-to-device variability in the classification effectiveness are investigated. The confusion matrix is used to benchmark the system performance metrics. We show that the simplicity, accuracy and robustness of the memdiode model makes it a suitable candidate for the realistic simulation of large-scale neural networks with non-idealities.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1109/access.2020.3035638

Cites

98 of 98 references verified0 citation errors · 0 not checkable

Statistics

Figures

FIGURE 1 (a) Sketch of the CPA structure. Red and blue arrows exemplify the electron flow through the memdiodes connecting the top (Word lines -WL-) and bottom lines (Bit lines -BL-). Different resistance states are schematically represented (High Resistance State -HRS- to Low Resistance State -LRS-). The dashed blue line depicts the so-called sneak path problem. The parasitic wire resistance is indicated for WLi and BL i . (b) Equivalent circuit representation of the CPA sketched in (a).
FIGURE 1 (a) Sketch of the CPA structure. Red and blue arrows exemplify the electron flow through the memdiodes connecting the top (Word lines -WL-) and bottom lines (Bit lines -BL-). Different resistance states are schematically represented (High Resistance State -HRS- to Low Resistance State -LRS-). The dashed blue line depicts the so-called sneak path problem. The parasitic wire resistance is indicated for WLi and BL i . (b) Equivalent circuit representation of the CPA sketched in (a).CC-BY · From the PDF
FIGURE 2 (a) Hysteron model with logistic ridge functions 0 C (Eq. 3) and 0 (Eq. 4). is the space of feasible states S . The red thick faded line superimposed to the hysteron model indicates the trajectory of the state variable inside from an initial ( S 1 ) to a final ( S 2 ) state. The inset in the left shows the equivalent circuit model for the current equation (1) including the series resistance. The diodes are driven by the memory state of the device and one diode is activated at a time. (b) Typical I -V characteristic for a memdiode obtained via simulation of the proposed model. Superimposed blue arrows indicate the current evolution. (c) Detail of the HRS (exponential) to LRS (linear) transition, showing a few intermediate states. Linear [13], [15], [16] and sinh() -like [34] models are superimposed for comparison.
FIGURE 2 (a) Hysteron model with logistic ridge functions 0 C (Eq. 3) and 0 (Eq. 4). is the space of feasible states S . The red thick faded line superimposed to the hysteron model indicates the trajectory of the state variable inside from an initial ( S 1 ) to a final ( S 2 ) state. The inset in the left shows the equivalent circuit model for the current equation (1) including the series resistance. The diodes are driven by the memory state of the device and one diode is activated at a time. (b) Typical I -V characteristic for a memdiode obtained via simulation of the proposed model. Superimposed blue arrows indicate the current evolution. (c) Detail of the HRS (exponential) to LRS (linear) transition, showing a few intermediate states. Linear [13], [15], [16] and sinh() -like [34] models are superimposed for comparison.CC-BY · From the PDF
FIGURE 3 Experimental I -V loops of different materials reported in the literature fitted with the QMM model: (a) HfO 2 [43], (b) Al 2 O 3 [44], (c) MnO 3 [45], (d) CuO 2 [46], (e) La 1-x Cax MnO 3 [47] and (f) TaO X [48]. The QMM fitting parameters are shown for each case. As reference, the HRS and LRS curves are indicated in (a) and the SET and RESET points in (b). Note that in (a) a current compliance of 200 A was imposed to prevent permanent dielectric breakdown, which can be also represented by the QMM.
FIGURE 3 Experimental I -V loops of different materials reported in the literature fitted with the QMM model: (a) HfO 2 [43], (b) Al 2 O 3 [44], (c) MnO 3 [45], (d) CuO 2 [46], (e) La 1-x Cax MnO 3 [47] and (f) TaO X [48]. The QMM fitting parameters are shown for each case. As reference, the HRS and LRS curves are indicated in (a) and the SET and RESET points in (b). Note that in (a) a current compliance of 200 A was imposed to prevent permanent dielectric breakdown, which can be also represented by the QMM.CC-BY · From the PDF
FIGURE 4 (a) Schematic representation of the Write-Verify procedure for programming the CPA memdiodes. The upper waveform stands for the alternating write and verify voltage pulses, while the lower plot represents the associated conductance changes. A simplified circuit schematic for the writing connection is indicated in the upper right inset. (b) Experimental and simulation results for the reset characteristics of SiO X from UCL (Data from [55]) using the QMM model. Notice the control of the intermediate memory states. The inset shows the input signal.
FIGURE 4 (a) Schematic representation of the Write-Verify procedure for programming the CPA memdiodes. The upper waveform stands for the alternating write and verify voltage pulses, while the lower plot represents the associated conductance changes. A simplified circuit schematic for the writing connection is indicated in the upper right inset. (b) Experimental and simulation results for the reset characteristics of SiO X from UCL (Data from [55]) using the QMM model. Notice the control of the intermediate memory states. The inset shows the input signal.CC-BY · From the PDF
FIGURE 5 Flowchart diagram for the simulation procedure. Starting with the image size specification, R W , V read , and connection scheme, the routine creates the database, trains the ANN (single layer perceptron), translates it into a CPA, adds the peripheral control circuit, performs the simulations and processes the results. MATLAB tasks are grouped by the red box and SPICE operations by the green box.
FIGURE 5 Flowchart diagram for the simulation procedure. Starting with the image size specification, R W , V read , and connection scheme, the routine creates the database, trains the ANN (single layer perceptron), translates it into a CPA, adds the peripheral control circuit, performs the simulations and processes the results. MATLAB tasks are grouped by the red box and SPICE operations by the green box.CC-BY · From the PDF
FIGURE 6 Samples of the three different images databases considered in this article. In all cases images are represented in 28 28 px. Pixel brightness (or intensity) is codified in 256 levels ranging from 0 (fully OFF, black) to 1 (fully ON, white). (a) MNIST database of handwritten numeric digits, (b) MNIST-F database [57] of fashion articles and (c) MNIST-K database [58] of handwritten Kanji Japanese ideograms.
FIGURE 6 Samples of the three different images databases considered in this article. In all cases images are represented in 28 28 px. Pixel brightness (or intensity) is codified in 256 levels ranging from 0 (fully OFF, black) to 1 (fully ON, white). (a) MNIST database of handwritten numeric digits, (b) MNIST-F database [57] of fashion articles and (c) MNIST-K database [58] of handwritten Kanji Japanese ideograms.CC-BY · From the PDF
FIGURE 7 Simplified equivalent circuit of the single layer perceptron with n 2 inputs and m outputs ( n 2 m ). 2 CPAs are considered, one of them is used to represent the positive weights and the other the negative weights. The Dual Connection Scheme (DSC) implies biasing the wordline terminals. V i 1 -V i n 2 indicates the elements of the input image vectors ( n 2 1).
FIGURE 7 Simplified equivalent circuit of the single layer perceptron with n 2 inputs and m outputs ( n 2 m ). 2 CPAs are considered, one of them is used to represent the positive weights and the other the negative weights. The Dual Connection Scheme (DSC) implies biasing the wordline terminals. V i 1 -V i n 2 indicates the elements of the input image vectors ( n 2 1).CC-BY · From the PDF
FIGURE 8 Impact of the memdiode resistance window on the detection performance. Different resistance windows are considered ( R ON and R OFF ) by scaling the LRS and HRS curves (a) HRS and LRS are scaled, (b) only HRS is scaled and (c) only LRS is scaled. Recognition accuracy as function of the models OFF resistance (d) and ON resistance (e). For every model each point represents a different value of R W , swept from 100 m to 10 (darker markers indicate higher values or R W ). CPAs are connected from both sides (DSC) and images are not downsized (img. resolution: 28 28 px.). Confusion matrix with R W =4.53 for model plays (f) A 1, (g) A 2, (h) A 3, (i) B 1, (j) B 2, (k) B 3, (l) C 1, (m) C 2 and (n) C 3.
FIGURE 8 Impact of the memdiode resistance window on the detection performance. Different resistance windows are considered ( R ON and R OFF ) by scaling the LRS and HRS curves (a) HRS and LRS are scaled, (b) only HRS is scaled and (c) only LRS is scaled. Recognition accuracy as function of the models OFF resistance (d) and ON resistance (e). For every model each point represents a different value of R W , swept from 100 m to 10 (darker markers indicate higher values or R W ). CPAs are connected from both sides (DSC) and images are not downsized (img. resolution: 28 28 px.). Confusion matrix with R W =4.53 for model plays (f) A 1, (g) A 2, (h) A 3, (i) B 1, (j) B 2, (k) B 3, (l) C 1, (m) C 2 and (n) C 3.CC-BY · From the PDF
FIGURE 9 (a) CPA Inference accuracy for different model plays ( R ON = R OFF ) and wire resistances ( R W ) plotted against its associated read margin. Avg. Read margins below 10% increasingly affects the CPA inference accuracy. (b) Map of read margin as function of the spatial location for a 784 10 CPA, with R W =1 . Significant voltage variation can be seen. (c) I -V characteristics of the memdiode showing the exponential (HRS) to linear (LRS) transition by varying . The red shaded region indicates the possible voltages applied to the device as the read margin reduces. I HRS and I LRS currents are pinpointed at nominal V read with the grey and white circle markers, respectively. Overestimation of I HRS may occur when considering a linear model for the HRS regime and lower effective V read voltages as indicated by the cyan, blue and black ball markers. The Linear model does not account for the R ON = R OFF ratio reduction with V read , and thereby cannot account for its impact on the accuracy (d), as well as on the CPA Power consumption (e).
FIGURE 9 (a) CPA Inference accuracy for different model plays ( R ON = R OFF ) and wire resistances ( R W ) plotted against its associated read margin. Avg. Read margins below 10% increasingly affects the CPA inference accuracy. (b) Map of read margin as function of the spatial location for a 784 10 CPA, with R W =1 . Significant voltage variation can be seen. (c) I -V characteristics of the memdiode showing the exponential (HRS) to linear (LRS) transition by varying . The red shaded region indicates the possible voltages applied to the device as the read margin reduces. I HRS and I LRS currents are pinpointed at nominal V read with the grey and white circle markers, respectively. Overestimation of I HRS may occur when considering a linear model for the HRS regime and lower effective V read voltages as indicated by the cyan, blue and black ball markers. The Linear model does not account for the R ON = R OFF ratio reduction with V read , and thereby cannot account for its impact on the accuracy (d), as well as on the CPA Power consumption (e).CC-BY · From the PDF
FIGURE 10 (a) CPA Power consumption vs. CPA size (in number of devices). Note that Power increases as R ON reduces from model plays ( C 1C 4). As reference, reported CPA power for other works [6], [8], [14], [79], [80] is also plotted. (b) Ratio between the memdiode power and the total CPA power consumption. (c) Obtained Signal-to-Noise ratio (SNR) for different model plays ( C 1C 4) as function of the line resistance. The marker color indicates the inference accuracy for each simulation case. Note that reduced R ON in C 1 comes at the cost of a reduced SNR and thereby limits accuracy.
FIGURE 10 (a) CPA Power consumption vs. CPA size (in number of devices). Note that Power increases as R ON reduces from model plays ( C 1C 4). As reference, reported CPA power for other works [6], [8], [14], [79], [80] is also plotted. (b) Ratio between the memdiode power and the total CPA power consumption. (c) Obtained Signal-to-Noise ratio (SNR) for different model plays ( C 1C 4) as function of the line resistance. The marker color indicates the inference accuracy for each simulation case. Note that reduced R ON in C 1 comes at the cost of a reduced SNR and thereby limits accuracy.CC-BY · From the PDF
FIGURE 11 (a) Impact of the wire resistance ( R W ) on the detection performance considering four different resistance windows corresponding to models plays C 1C 4. (b) Inference accuracy is plotted against the R W = R ON ratio showing a unified trend among all model plays. Confusion matrix for model C 2, with R W equal to (c) 0.1 , (d) 4.53 (16 nm [15]) and (e) 81.3 (10 nm [76]). In all cases CPAs are connected from both sides (DSC) and images are not downsized (image resolution: 28 28 px.).
FIGURE 11 (a) Impact of the wire resistance ( R W ) on the detection performance considering four different resistance windows corresponding to models plays C 1C 4. (b) Inference accuracy is plotted against the R W = R ON ratio showing a unified trend among all model plays. Confusion matrix for model C 2, with R W equal to (c) 0.1 , (d) 4.53 (16 nm [15]) and (e) 81.3 (10 nm [76]). In all cases CPAs are connected from both sides (DSC) and images are not downsized (image resolution: 28 28 px.).CC-BY · From the PDF
FIGURE 12 (a) Expected inference latency for a CPA of 1Mb (1,000 1 : 000) as function of the process feature size, considering the model proposed by Meindl et al. in [76], [87], SPICE simulations considering the QMM, and the simplified circuital model depicted by Eq. 13. (b) Circuit schematic of a RRAM cell in a CPA structure considering the associated wire parasitic resistance and capacitance. (c) Schematic representation of the simplified model of Eq. 13. (d) Estimated Latency based on Eq. 13 as function of R OFF .
FIGURE 12 (a) Expected inference latency for a CPA of 1Mb (1,000 1 : 000) as function of the process feature size, considering the model proposed by Meindl et al. in [76], [87], SPICE simulations considering the QMM, and the simplified circuital model depicted by Eq. 13. (b) Circuit schematic of a RRAM cell in a CPA structure considering the associated wire parasitic resistance and capacitance. (c) Schematic representation of the simplified model of Eq. 13. (d) Estimated Latency based on Eq. 13 as function of R OFF .CC-BY · From the PDF
FIGURE 13 (a) Readability loss as the resolution decreases from 28 28 px (case I) to 8 8 (case IV). (b) Impact of the CPA size on the detection performance for model C 2 and with R W swept from 1 to 100 . Two different partitions schemes are considered: P1 indicates non-partitioned arrays whereas P4 stands for CPA partitioned in 4 sub-arrays. Markers indicate max. size with max. accuracy. Partitioned arrays allow higher accuracy in larger CPAs. (c) Recognition accuracy against the ratio R W = R ON for two different array sizes: 28 28 images (two 784 10 arrays, 15.6k devices, empty markers) and 8 8 images (two 64 10 arrays, 1.2k devices, filled markers). Resulting confusion matrix considering model play C 2 and R W =10 for (d) 28 28, (e) 20 20 and (f) 8 8 px. image sizes, considering non-partitioned CPAs.
FIGURE 13 (a) Readability loss as the resolution decreases from 28 28 px (case I) to 8 8 (case IV). (b) Impact of the CPA size on the detection performance for model C 2 and with R W swept from 1 to 100 . Two different partitions schemes are considered: P1 indicates non-partitioned arrays whereas P4 stands for CPA partitioned in 4 sub-arrays. Markers indicate max. size with max. accuracy. Partitioned arrays allow higher accuracy in larger CPAs. (c) Recognition accuracy against the ratio R W = R ON for two different array sizes: 28 28 images (two 784 10 arrays, 15.6k devices, empty markers) and 8 8 images (two 64 10 arrays, 1.2k devices, filled markers). Resulting confusion matrix considering model play C 2 and R W =10 for (d) 28 28, (e) 20 20 and (f) 8 8 px. image sizes, considering non-partitioned CPAs.CC-BY · From the PDF
FIGURE 14 Simplified equivalent circuit schematic for a partitioned CPA based single layer perceptron. Each CPA is subdivided into N identically sized partitions to minimise the parasitic voltage drops. Partial output current vectors are indicated in the output of each partition.
FIGURE 14 Simplified equivalent circuit schematic for a partitioned CPA based single layer perceptron. Each CPA is subdivided into N identically sized partitions to minimise the parasitic voltage drops. Partial output current vectors are indicated in the output of each partition.CC-BY · From the PDF
FIGURE 15 Impact of the device variability on the test accuracy. (a) Different sources of variability considered: R ON , R OFF and . (b) Impact of the variability on the accuracy. Model plays A 1 and C 1 are equivalent and thereby only one is considered (See. Fig 8a-8c). Inference accuracy is studied as function of the combined variability of R ON and R OFF for model plays A 2 (c) and C 2 (d) showing a similar behaviour to the one exhibit by (b).
FIGURE 15 Impact of the device variability on the test accuracy. (a) Different sources of variability considered: R ON , R OFF and . (b) Impact of the variability on the accuracy. Model plays A 1 and C 1 are equivalent and thereby only one is considered (See. Fig 8a-8c). Inference accuracy is studied as function of the combined variability of R ON and R OFF for model plays A 2 (c) and C 2 (d) showing a similar behaviour to the one exhibit by (b).CC-BY · From the PDF
FIGURE 16 Performance metrics of the memdiode-based CPA simulated with the QMM, compared against the ideal Software case computed in MATLAB. 5 different databases are considered: MNIST [78], MNIST-F [57], MNIST-K [58], CIFAR-10 [94] and SVHN [95]. Each CPA is divided into 4 partitions to further reduced the impact of R W . Note that as R W tends to 0, the CPA mimics the MATLAB results.
FIGURE 16 Performance metrics of the memdiode-based CPA simulated with the QMM, compared against the ideal Software case computed in MATLAB. 5 different databases are considered: MNIST [78], MNIST-F [57], MNIST-K [58], CIFAR-10 [94] and SVHN [95]. Each CPA is divided into 4 partitions to further reduced the impact of R W . Note that as R W tends to 0, the CPA mimics the MATLAB results.CC-BY · From the PDF
FIGURE 17 The computational cost (run-time and RAM memory usage) of the memdiode-based CPA implemented with the QMM model is compared against other memristor models reported by Yakopcic et al. [34], Laiho et al. [96] and the University of Michigan [97]. (a) Simulation run-time as function of the CPA size (in number of devices) increases quadratically. (b) Simulation RAM memory usage increases linearly with the CPA size. (c) Total CPU time normalised against the case of a resistor of fixed value to represent the programmed memristor. (d) Total RAM memory usage normalised against the resistor case (linear model). In both cases simulations were performed both in HSPICE and Fast-SPICE (FineSim) environments and for different image sizes.
FIGURE 17 The computational cost (run-time and RAM memory usage) of the memdiode-based CPA implemented with the QMM model is compared against other memristor models reported by Yakopcic et al. [34], Laiho et al. [96] and the University of Michigan [97]. (a) Simulation run-time as function of the CPA size (in number of devices) increases quadratically. (b) Simulation RAM memory usage increases linearly with the CPA size. (c) Total CPU time normalised against the case of a resistor of fixed value to represent the programmed memristor. (d) Total RAM memory usage normalised against the resistor case (linear model). In both cases simulations were performed both in HSPICE and Fast-SPICE (FineSim) environments and for different image sizes.CC-BY · From the PDF
Image 18
CC-BY · From the PDF
Image 19
CC-BY · From the PDF
Image 20
CC-BY · From the PDF
Image 21
CC-BY · From the PDF
Image 22
CC-BY · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 98 references

Research works printed but not linked: 3

  1. 2010, Resistance Random Access Memory (RRAM)
  2. lecun, 1998, MNIST Handwritten Digit Database
  3. hagan, 2014, Neural Network Design

Cited by

See all 36

Metadata

Authors
Aguirre, F. L., Pazos, S. M., Palumbo, F., Sune, J., & Miranda, E.
Affiliations
  • Universitat Autònoma de Barcelona (ES)
  • National Technological University (AR)
Venue
IEEE Access
Published
2020
Volume
8
Pages
202174-202193
ISSN
2169-3536 (ISSN-L)
OpenAlex
W3096521724
Cited by
36
References
100
Citations by year
2026: 6 · 2025: 6 · 2024: 6 · 2023: 6 · 2022: 7 · 2021: 4 · 2020: 1
FWCI
1.54
Type
journal article
Language
en
Open access
Locations
  • IEEE Access (publishedVersion, cc-by)
  • Dipòsit Digital de Documents de la UAB (Universitat Autònoma de Barcelona) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • CONICET Digital (CONICET) (submittedVersion, cc-by)
  • Greater South Information System
Funders
  • Universidad Tecnolgica Nacional (PID-UTN CCUTIBA4764TC,, PID-UTN CCUTNBA5182, PID-UTN CCUTNBA6615, PID-UTN EIUTIBA4395TC3, PID-UTN MATUNBA4936)
  • Electronic Components and Systems for European Leadership (WAKeMeUP 783176)
  • Consejo Nacional de Investigaciones Cientficas y Tcnicas (PIP-11220130100077CO)
  • Ministerio de Ciencia Tecnologa e Innovacin Productiva (PICT 2013/1210, PICT 2016/0579, PME 2015-0196)
  • Ministerio de Ciencia e Innovacin (TEC2017-84321-C4-4-R)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Ferroelectric and Negative Capacitance Devices
Keywords
cross-point array, in-memory computing, pattern classification, MNIST, synaptic weights, device-to-device variation, neuromorphic hardware

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Application of the Quasi-Static Memdiode Model in Cross-Point Arrays for Large Dataset Pattern Recognition. https://app.cochat.ai/papers/10.1109/access.2020.3035638
The paper
Aguirre, F. L., Pazos, S. M., Palumbo, F., Sune, J., & Miranda, E. (2020). Application of the Quasi-Static Memdiode Model in Cross-Point Arrays for Large Dataset Pattern Recognition. IEEE Access. https://doi.org/10.1109/access.2020.3035638

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.