Student or educator? Sign up with your school email and get 6 months of Pro free.
Cited by 265Advanced Memory and Neural ComputingFerroelectric and Negative Capacitance DevicesNeural Networks and Applications

The DOI matches its Crossref record; 58 of the 58 references it cites were found in a registry or an index; 9 of the 23 sources its claims cite were checked for support.

4K-memristor analog-grade passive crossbar circuit

Open access
Kim, H., Mahmoodi, M. R., Nili, H., Strukov, D. B.·2021·Nature Communications

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.

Overview

CoChat Summary

## TL;DR

This paper reports a foundry-compatible 64 × 64 passive metal-oxide memristor crossbar with ~99% functional devices and low enough variation to support analog tuning and in-memory vector-by-matrix multiplication. The authors demonstrate 4K gray-scale pattern programming with <4% average relative error, a 64 × 10 MNIST classifier with ~1% average relative conductance import accuracy, and simulations showing algorithmic gains against device variation.

## Key findings

  • The fabricated 64 × 64 passive crossbar had ~99% functional nonvolatile memristors, and only 45 devices were unswitchable, i.e. ~1.125% of the array.
  • Device switching thresholds averaged 1.19 V for set and –1.39 V for reset, with standard deviations of 0.31 V and 0.37 V; the device-to-device switching-voltage coefficient of variance was <26%.
  • The tuning algorithm programmed an Einstein gray-scale image on the full array: excluding unswitchable devices, ~98% of devices were tuned within 5% relative error, with an average relative error of ~3.76%; the abstract also states <4% average relative tuning error for a 4K-pixel gray-scale pattern.
  • In the 64 × 10 MNIST single-layer perceptron demo, the average relative conductance import accuracy was ~1%; the measured classifier accuracy was 1.87% lower than the ideal software model for the most accurate weight import.
  • For the modeled multilayer perceptron, the advanced tuning approach improved tuning accuracy by 9% versus baseline at 26% device variation, and increased the tolerable variation for no classification loss from ~14% to ~20%; the baseline classifier accuracy at 26% variation improved to ~97.3%.

## Methods

  • The paper combines fabrication, electrical characterization, conductance tuning experiments, and modeling of ex-situ-trained neural-network inference circuits.
  • The device sample is a 64 × 64 passive crossbar on a 4-inch Si wafer, built from Ti/Al/TiN electrodes and an Al2O3/TiO2−x switching stack, with 500 devices used for retention testing and 500 crossbar-integrated memristors used to fit the conductance-change model.
  • Electrical measurements used a Keysight B1500A analyzer and switching matrix; the MNIST hardware demo used a 64 × 10 single-layer perceptron trained on 60,000 training and 10,000 8 × 8 down-sampled test images, and the modeling study used an ex-situ-trained 784-64-10 multilayer perceptron with differential conductance mapping and simulated tuning algorithms.

## Limitations

  • The authors state that the demonstrated MNIST classifier was “quite low” in accuracy compared to state-of-the-art, because it used a single-layer network and down-sampled black/white images.
  • The paper explicitly notes that the experiment demonstrating classifier fidelity was aimed at conductance tuning capability rather than high classification accuracy.
  • Evident from the text: 45 unswitchable devices remained in the array, and these devices contributed to tuning error and one misclassified MNIST example.
  • Evident from the text: retention is reported via accelerated tests at 100 °C with extrapolation to room temperature, so the room-temperature lifetime claims are inferred rather than directly measured.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-021-25455-0

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Device fabrication details. a–h Growth and patterning process steps (see “Methods” section for details). For clarity, panel e shows device cross-section turned out-of-plane by 90 degrees with respect to drawings shown in panels a–d, f–h. Scanning electron microscopy (SEM) images of (i) patterned bottom electrodes, (j) partially planarized bottom electrodes through chemical-mechanical polishing and etch-back, and (k) a fragment of completed crossbar array. l SEM image of the full 64 × 64 memristor crossbar array. Bottom left and bottom right insets show, correspondingly, material layers at the device cross-section with corresponding thicknesses in nanometers, and the packaged chip.
Fig. 1 Device fabrication details. a–h Growth and patterning process steps (see “Methods” section for details). For clarity, panel e shows device cross-section turned out-of-plane by 90 degrees with respect to drawings shown in panels a–d, f–h. Scanning electron microscopy (SEM) images of (i) patterned bottom electrodes, (j) partially planarized bottom electrodes through chemical-mechanical polishing and etch-back, and (k) a fragment of completed crossbar array. l SEM image of the full 64 × 64 memristor crossbar array. Bottom left and bottom right insets show, correspondingly, material layers at the device cross-section with corresponding thicknesses in nanometers, and the packaged chip.CC BY · PubMed Central
Fig. 2 Device characterization results. a Representative I–V curves, measured with quasi-static DC voltage sweeps, for the 36 formed devices of the 6 × 6 subarrays located in the center of the crossbar. For clarity, the curve for one particular device is highlighted. b Retention results for 10 different devices with data for each device shown with a specific color. The tests for each device are performed 9 times with randomly chosen initial conductance. The evolution of the conductance was measured at 400 s intervals while continuously baking the crossbar circuit at 100 °C. c The standard deviation of the absolute conductance change normalized to Gmax = 62.5 µS, i.e., 100% × |Ginitial − Gfinal|/Gmax, as a function of the time interval for several ranges of initial conductances. Similar to panel b, the top axis corresponds to the measured retention data at 100 °C for 500 devices, with each device tested at 6 different initial states, while the bottom axis shows extrapolated results. For panels b and c, the bottom axes show extrapolated time at room temperature (RT) assuming 1.1 eV activation energy (see “Methods” section for details). d The switching endurance results for a crossbar device. The data are obtained by repeatedly applying alternative polarity sequences of 1-ms voltage pulses. The absolute amplitude of pulse in each sequence is initially 0.8 V and then ramped up with 0.1 V steps until the device reaches the extreme (i.e., on or off) state. Inset is a zoomed-in portion of the main panel, showing typical continuous switching during the endurance test. The device is switched about 105 times between its extreme states during the experiment. e Measured evolution of conductance upon application of increasing amplitude voltage pulses. All parameters of the utilized pulse sequences are similar to those shown in Fig. 3c inset, except for the 50 mV incremental step. f–h Extracted statistics of switching threshold voltages, defined as the smallest absolute voltage at which device conductance, measured at 0.25 V, change by 20%, shown as (f) histogram and (g, h) voltage maps for (g) set and (h) reset transitions. The conductances are measured at 0.1 V for panels b–d.
Fig. 2 Device characterization results. a Representative I–V curves, measured with quasi-static DC voltage sweeps, for the 36 formed devices of the 6 × 6 subarrays located in the center of the crossbar. For clarity, the curve for one particular device is highlighted. b Retention results for 10 different devices with data for each device shown with a specific color. The tests for each device are performed 9 times with randomly chosen initial conductance. The evolution of the conductance was measured at 400 s intervals while continuously baking the crossbar circuit at 100 °C. c The standard deviation of the absolute conductance change normalized to Gmax = 62.5 µS, i.e., 100% × |Ginitial − Gfinal|/Gmax, as a function of the time interval for several ranges of initial conductances. Similar to panel b, the top axis corresponds to the measured retention data at 100 °C for 500 devices, with each device tested at 6 different initial states, while the bottom axis shows extrapolated results. For panels b and c, the bottom axes show extrapolated time at room temperature (RT) assuming 1.1 eV activation energy (see “Methods” section for details). d The switching endurance results for a crossbar device. The data are obtained by repeatedly applying alternative polarity sequences of 1-ms voltage pulses. The absolute amplitude of pulse in each sequence is initially 0.8 V and then ramped up with 0.1 V steps until the device reaches the extreme (i.e., on or off) state. Inset is a zoomed-in portion of the main panel, showing typical continuous switching during the endurance test. The device is switched about 105 times between its extreme states during the experiment. e Measured evolution of conductance upon application of increasing amplitude voltage pulses. All parameters of the utilized pulse sequences are similar to those shown in Fig. 3c inset, except for the 50 mV incremental step. f–h Extracted statistics of switching threshold voltages, defined as the smallest absolute voltage at which device conductance, measured at 0.25 V, change by 20%, shown as (f) histogram and (g, h) voltage maps for (g) set and (h) reset transitions. The conductances are measured at 0.1 V for panels b–d.CC BY · PubMed Central
Fig. 3 Conductance tuning results. a Forming and high-precision tuning to 20 µS, 10 µS, 100 µS, and 8 µS target conductances of a crossbar device, with 1% relative precision. The inset shows the applied sequence of pulses during conductance tuning. Pulse sequences parameters are similar to those of panel c, except for the utilized 50 mV incremental step. b Example of device tuning with a 1% relative error to different conductance levels equally spaced from 3 µS to 45 µS. c–e Programming Einstein image to the 64 × 64 crossbar array with a 5% relative error. c Tuning statistics. Inset shows details of the write-verify pulse sequence. d The target device conductances in the range of 10 µS to 100 µS corresponding to the gray-scale quantized image and (e) their actual measured values after completing automated tuning. The relative tuning error is defined as 100% × |Itarget(0.25 V) − Iactual(0.25 V) | /Itarget(0.25 V). All conductances are specified at 0.25 V. Einstein image copyright by Yousuf Karsh.
Fig. 3 Conductance tuning results. a Forming and high-precision tuning to 20 µS, 10 µS, 100 µS, and 8 µS target conductances of a crossbar device, with 1% relative precision. The inset shows the applied sequence of pulses during conductance tuning. Pulse sequences parameters are similar to those of panel c, except for the utilized 50 mV incremental step. b Example of device tuning with a 1% relative error to different conductance levels equally spaced from 3 µS to 45 µS. c–e Programming Einstein image to the 64 × 64 crossbar array with a 5% relative error. c Tuning statistics. Inset shows details of the write-verify pulse sequence. d The target device conductances in the range of 10 µS to 100 µS corresponding to the gray-scale quantized image and (e) their actual measured values after completing automated tuning. The relative tuning error is defined as 100% × |Itarget(0.25 V) − Iactual(0.25 V) | /Itarget(0.25 V). All conductances are specified at 0.25 V. Einstein image copyright by Yousuf Karsh.CC BY · PubMed Central
Fig. 4 Experimental results for pattern classification. a A portion of the crossbar circuit is utilized in a 64 × 10 single-layer perceptron MNIST image classification experiment. b Examples of target and (c) actual conductances after tuning with a 1% relative error. d Measured classification fidelity and its comparison with simulation results as a function of weight import accuracy. In each simulation trial, the weights are selected randomly from a range of target_value × [1 − tuning_error, 1 + tuning_error]. e, f Measured output currents for all ten outputs over the 10-s interval for patterns ‘7’ and ‘9’ (shown in the corresponding insets) for the experiment with a 1% relative tuning error. The currents are measured, one row at a time, by simultaneously applying input voltages on all 64 columns and grounding 10 specified rows. See Fig. 3 caption for the definition of relative tuning error.
Fig. 4 Experimental results for pattern classification. a A portion of the crossbar circuit is utilized in a 64 × 10 single-layer perceptron MNIST image classification experiment. b Examples of target and (c) actual conductances after tuning with a 1% relative error. d Measured classification fidelity and its comparison with simulation results as a function of weight import accuracy. In each simulation trial, the weights are selected randomly from a range of target_value × [1 − tuning_error, 1 + tuning_error]. e, f Measured output currents for all ten outputs over the 10-s interval for patterns ‘7’ and ‘9’ (shown in the corresponding insets) for the experiment with a 1% relative tuning error. The currents are measured, one row at a time, by simultaneously applying input voltages on all 64 columns and grounding 10 specified rows. See Fig. 3 caption for the definition of relative tuning error.CC BY · PubMed Central
Fig. 5 Modeling of ex-situ trained MLP classifier. a The block diagram for the distributed mixed-signal implementation of 784-64-10 multilayer perceptron (MLP) classifier with 64 × 64 crossbar circuits. Programming circuitry is omitted for clarity. b Modeled relative tuning error after 10 rounds of tuning for the first layer of MLP network as a function of device-to-device variations when using four different conductance tuning approaches. See Section “Modeling of Advanced Systems” for more details of the tuning approaches. c Simulated accuracy of MLP classifier as a function of device-to-device variations when using baseline and the most advanced tuning approach after 10 rounds of tuning. Inset shows zoom-in for the high classification accuracy portion of the graph. The shown numbers roughly correspond to the device variations observed in the experiment. The box plot shows the statistics over 10 different runs of initial conductances. For simplicity, memristors’ static I–V nonlinearities and noise are neglected, and ideal peripheral circuits are assumed in simulations. See Fig. 3 caption for the definition of relative tuning error.
Fig. 5 Modeling of ex-situ trained MLP classifier. a The block diagram for the distributed mixed-signal implementation of 784-64-10 multilayer perceptron (MLP) classifier with 64 × 64 crossbar circuits. Programming circuitry is omitted for clarity. b Modeled relative tuning error after 10 rounds of tuning for the first layer of MLP network as a function of device-to-device variations when using four different conductance tuning approaches. See Section “Modeling of Advanced Systems” for more details of the tuning approaches. c Simulated accuracy of MLP classifier as a function of device-to-device variations when using baseline and the most advanced tuning approach after 10 rounds of tuning. Inset shows zoom-in for the high classification accuracy portion of the graph. The shown numbers roughly correspond to the device variations observed in the experiment. The box plot shows the statistics over 10 different runs of initial conductances. For simplicity, memristors’ static I–V nonlinearities and noise are neglected, and ideal peripheral circuits are assumed in simulations. See Fig. 3 caption for the definition of relative tuning error.CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 1 Not valid · 10 Not checked · 60 Own results and interpretations

44 more claims

Create a free account to see all 78 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 13
  • It may state two separate claims as one. 20

Create a free account to see these claims.

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?

Citations

Cites

Show all 58 references

Research works printed but not linked: 1

  1. Burr, G. W. et al. Neuromorphic computing using nonvolatile memory. Adv. Phys. 2, 89–124 (2017).

Cited by

See all 265

Metadata

Authors
Kim, H., Mahmoodi, M. R., Nili, H., & Strukov, D. B.
Affiliations
University of California, Santa Barbara (US)
Corresponding author
Strukov, D. B.
Venue
Nature Communications
Published
2021-08-31
Volume
12
Issue
1
Pages
5198-5198
ISSN
2041-1723 (ISSN-L)
OpenAlex
W3197387328
Cited by
265
References
58
Citations by year
2026: 40 · 2025: 59 · 2024: 55 · 2023: 47 · 2022: 47 · 2021: 11
FWCI
11.76
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, other-oa)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • eScholarship (California Digital Library) (submittedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (1740352)
  • Semiconductor Research Corporation (JUMP CRISP)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Ferroelectric and Negative Capacitance Devices
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
Keywords
memristors, metal-oxide memristors, non-volatile memory, neuromorphic computing, in-memory computing, multilayer perceptron, MNIST classification, conductance tuning, foundry-compatible fabrication

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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: 4K-memristor analog-grade passive crossbar circuit. https://app.cochat.ai/papers/10.1038/s41467-021-25455-0
The paper
Kim, H., Mahmoodi, M. R., Nili, H., & Strukov, D. B. (2021). 4K-memristor analog-grade passive crossbar circuit. Nature Communications. https://doi.org/10.1038/s41467-021-25455-0

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.