Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 37 of the 37 references it cites were found in a registry or an index; 4 of the 22 sources its claims cite were checked for support.

Framework for In-Memory Computing Based on Memristor and Memcapacitor for On-Chip Training

Open access
Singh, Ankur, Lee, Byung-Geun · 2023 · IEEE Access

Abstract

Memristive crossbar arrays have gained considerable attention from researchers to perform analog in-memory vector-matrix multiplications in machine learning accelerators with low power and constant computational time. This work introduces a comprehensive framework for co-designing the software and hardware for deep neural networks (DNN) based on memristive and memcapacitive crossbars while considering various non-idealities. The model takes into account device-level factors, including conductance variation, cycle-to-cycle variation, device-to-device variation, peripheral circuits for error/weight gradient computation, and high tolerance. The overall neural network performance is thoroughly assessed by integrating these elements into a unified DNN training process. The proposed framework is implemented using a hybrid approach with Python and PyTorch. Performance evaluation was conducted using a simplified 8-layer VGG network on a measured 128×128 array with weight resolution. Remarkably, the memristive and memcapacitive crossbar arrays achieved outstanding training accuracies of 90.02% and 91.03%, respectively, for the CIFAR-10 dataset. Additionally, detailed hardware estimation for both mem-elements devices is provided, enabling meaningful comparisons with prior works.

Overview

CoChat Summary

## TL;DR

The paper presents a Python/PyTorch NeuroSim-based framework for evaluating on-chip training in memristor and memcapacitor in-memory computing systems while accounting for device, array, and peripheral non-idealities. Using a simplified 8-layer VGG on CIFAR-10 with a 128×128 array, it reports training accuracies of 90.02% for memristor and 91.03% for memcapacitor.

## Key findings

  • The proposed framework was tested on a simplified 8-layer VGG network with a 128×128 RRAM array, and achieved training accuracies of 90.02% (memristor) and 91.03% (memcapacitor) on CIFAR-10.
  • The paper reports estimated training dynamic energy per epoch of 108.36 J and training latency per epoch of 104.31 sec.
  • In the reported hardware results, the memristor and memcapacitor designs achieved training throughput of 1.51 TOPS and 1.54 TOPS, respectively, with training energy efficiency of 2.10 TOPS/W and 2.32 TOPS/W.
  • The paper states that training accuracy is more strongly affected by line resistance than by device-to-device or cycle-to-cycle variation, and that device-to-device and cycle-to-cycle variation were well managed.

## Methods

  • The authors built a hybrid Python + PyTorch CUDA framework extending NeuroSim for compute-in-memory on-chip training, with hardware-extracted models for TiOx memristor and Si memcapacitor devices.
  • They evaluated the framework on a simplified 8-layer VGG network (6 convolutional layers and 2 fully connected layers) using the CIFAR-10 dataset, and also report a ResNet-34 experiment.
  • The analysis incorporated non-idealities such as conductance/capacitance variation, nonlinearity, asymmetry, sneak paths, line resistance, peripheral circuits, and ADC quantization, with hardware configuration including 128×128 crossbars and 5-bit ADC precision.

## Limitations

  • The paper explicitly says it does not specify the exact on-chip storage capacity required for intermediate results or routing to multiple arrays.
  • The authors state they have not developed a comprehensive analysis of offloading performance and power overheads yet; they note this work is ongoing.
  • They also state that they have not incorporated a read noise model yet and intend to do so in future work.
  • Evident from the text: the evaluation is centered on a simplified 8-layer VGG and a CIFAR-10 benchmark, with the paper itself noting that future iterations should include a wider range of networks and more complex datasets.

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

Integrity signals

Registry

Crossref · 10.1109/access.2023.3324375

Cites

37 of 37 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1 The architecture of the convolutional layer and fully connected layer integrated with memristor and memcapacitor crossbars, accompanied by the flowchart of the proposed framework.
FIGURE 1 The architecture of the convolutional layer and fully connected layer integrated with memristor and memcapacitor crossbars, accompanied by the flowchart of the proposed framework.From the PDF
Image 2
From the PDF
FIGURE 2 The simplified 8-layer VGG network comprises 6 convolutional layers and 2 fully connected layers.
FIGURE 2 The simplified 8-layer VGG network comprises 6 convolutional layers and 2 fully connected layers.From the PDF
FIGURE 3 Crossbar structure defined in the simulator using (a) memristor device, and (b) memcapacitor devices. Measured memductance data of the (c) memristor device with TiOx material, (d) memcapacitor device with Si material.
FIGURE 3 Crossbar structure defined in the simulator using (a) memristor device, and (b) memcapacitor devices. Measured memductance data of the (c) memristor device with TiOx material, (d) memcapacitor device with Si material.From the PDF
Image 5
From the PDF
FIGURE 4 Device structure (a) TiOx-based memristor device (b) Sibased memcapacitor device.
FIGURE 4 Device structure (a) TiOx-based memristor device (b) Sibased memcapacitor device.From the PDF
FIGURE 5 Distribution of the TiOx memristor conductance in the HRS and LRS.
FIGURE 5 Distribution of the TiOx memristor conductance in the HRS and LRS.From the PDF
FIGURE 6 Training accuracy comparison of non-idealities (a) different device-to-device memductance variation (b) different cycle-to-cycle variation.
FIGURE 6 Training accuracy comparison of non-idealities (a) different device-to-device memductance variation (b) different cycle-to-cycle variation.From the PDF
FIGURE 7 The accuracy is influenced by the line resistance and the sneak paths.
FIGURE 7 The accuracy is influenced by the line resistance and the sneak paths.From the PDF
FIGURE 8 Transistor level used in the framework (a) current sense amplifier (CSA), (b) voltage sense amplifier (VSA), (c) level shifter, and (d) successive approximation register (SAR) ADC.
FIGURE 8 Transistor level used in the framework (a) current sense amplifier (CSA), (b) voltage sense amplifier (VSA), (c) level shifter, and (d) successive approximation register (SAR) ADC.From the PDF
Image 11
From the PDF
FIGURE 9 The framework was trained using the CIFAR-10 dataset for 256 epochs. Training throughput (TOPS) of VMM with (a) memristor device (b) memcapacitor device. Training Energy Efficiency (TOPS/W) of VMM with (c) memristor device (d) memcapacitor device.
FIGURE 9 The framework was trained using the CIFAR-10 dataset for 256 epochs. Training throughput (TOPS) of VMM with (a) memristor device (b) memcapacitor device. Training Energy Efficiency (TOPS/W) of VMM with (c) memristor device (d) memcapacitor device.From the PDF
FIGURE 10 The training accuracy of the VMM accelerators with (a) memristor device, (b) memcapacitor device.
FIGURE 10 The training accuracy of the VMM accelerators with (a) memristor device, (b) memcapacitor device.From the PDF
Image 14
From the PDF
Image 15
From the PDF

Verification

Claims the paper makes

4 Valid · 1 Not valid · 15 Not checked · 37 Own results and interpretations

23 more claims

Create a free account to see all 57 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 16
  • Probably not a substantial claim. 15

Create a free account to see these claims.

Self-check

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

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 37 references

Research works printed but not linked: 1

  1. LeCun, 1998, The MNIST Database of Handwritten Digits

Cited by

See all 19

Metadata

Authors
Singh, A., & Lee, B.-G.
Affiliations
Gwangju Institute of Science and Technology (KR)
Venue
IEEE Access
Published
2023
Volume
11
Pages
112590-112599
ISSN
2169-3536 (ISSN-L)
OpenAlex
W4387623795
Cited by
19
References
37
Citations by year
2026: 5 · 2025: 9 · 2024: 5
FWCI
1.41
Type
journal article
Language
en
Open access
Locations
  • IEEE Access (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Research Foundation of Korea (NRF) Grant through the Korean Government (2021R1A2C22013480)
  • Nano-Material Technology Development Program through NRF
  • Ministry of Science and ICT (NRF-2022M3H4A1A01009658)
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 › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
in-memory computing, memristors, memcapacitor, memristor crossbar, on-chip learning, deep neural network accelerators, conductance variation, cycle-to-cycle variation, device-to-device variation, VGG, CIFAR-10, hardware-software co-design

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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: Framework for In-Memory Computing Based on Memristor and Memcapacitor for On-Chip Training. https://app.cochat.ai/papers/10.1109/access.2023.3324375
The paper
Singh, A., & Lee, B.-G. (2023). Framework for In-Memory Computing Based on Memristor and Memcapacitor for On-Chip Training. IEEE Access. https://doi.org/10.1109/access.2023.3324375

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.