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

The DOI matches its Crossref record; 83 of the 83 references it cites were found in a registry or an index; 15 of the 38 sources its claims cite were checked for support.

Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators

Open access

Abstract

Analog in-memory computing-a promising approach for energy-efficient acceleration of deep learning workloads-computes matrix-vector multiplications but only approximately, due to nonidealities that often are non-deterministic or nonlinear. This can adversely impact the achievable inference accuracy. Here, we develop an hardware-aware retraining approach to systematically examine the accuracy of analog in-memory computing across multiple network topologies, and investigate sensitivity and robustness to a broad set of nonidealities. By introducing a realistic crossbar model, we improve significantly on earlier retraining approaches. We show that many larger-scale deep neural networks-including convnets, recurrent networks, and transformers-can in fact be successfully retrained to show iso-accuracy with the floating point implementation. Our results further suggest that nonidealities that add noise to the inputs or outputs, not the weights, have the largest impact on accuracy, and that recurrent networks are particularly robust to all nonidealities.

Overview

CoChat Summary

TL;DR

The paper develops a hardware-aware retraining framework for analog in-memory computing (AIMC) and shows that many large DNNs, including CNNs, RNNs, and transformers, can be brought to iso-accuracy or near-iso-accuracy on a realistic PCM-based crossbar model. Across 11 benchmarks, five reached the paper’s 99% iso-accuracy target, and the average accuracy gap fell from 11.3% to 1.1% at 1 h after programming.

Key findings

  • The benchmark covered 11 medium- to large-scale DNNs spanning image classification, speech-to-text, and NLP, with model sizes from 0.3M to 108M parameters, depths up to 121 layers, and crossbar utilization from 4% to 86%.
  • With hardware-aware (HWA) retraining, the normalized accuracies at 1 h after programming were all above 97%, and 5/11 workloads reached the paper’s 99% iso-accuracy target.
  • Among the tested nonidealities, those that add noise to inputs or outputs had the largest impact on DNN accuracy, including ADC/DAC resolution, additive output noise, and S-shaped ADC nonlinearity.
  • CNNs were the most sensitive topology and RNNs the least sensitive; some RNNs and transformers remained near iso-accuracy even after long-term PCM drift, while HWA-trained models for Speech-SWB300 and LSTM-PTB stayed iso-accurate out to 1 year.

Methods

  • The authors built a standardized AIMC crossbar simulation model for PCM-based hardware, incorporating dynamic-range limits, programming errors, drift, read noise, output noise, DAC/ADC quantization, and IR-drop; they used a 512 × 512 tile size and 8-bit quantization by default.
  • They HWA-retrained 11 DNNs starting from FP-trained models, using SGD with noise injection and learned peripheral scaling factors for inputs and outputs; evaluation was done at delays of 1 s, 1 h, 1 day, and 1 year after programming.
  • They performed sensitivity analyses by stress-testing individual nonidealities, ranking layers by sensitivity in large CNNs, and measuring MVM error as relative deviation from ideal matrix-vector multiplication.

Limitations

  • The authors state that the results are simulated and that the larger-scale DNN iso-accuracy results remain to be verified in hardware in future.
  • They note that the model uses simplifications and abstractions, including Gaussian noise assumptions and an approximate IR-drop calculation, and therefore cannot replace ultimate hardware verification.
  • Evidence from the text: some results depend on a PCM-specific calibration, so transferability to non-PCM devices is not direct, although the authors say the framework can be adapted.

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

Integrity signals

Registry

Crossref · 10.1038/s41467-023-40770-4

Cites

83 of 83 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Illustration of the HWA training approach. In our hardware-aware (HWA) training setup, DNNs are first trained in 32-bit floating point (FP32), then retrained in a hardware-aware manner, by adding nonidealities and noise sources into the forward path and using SGD to improve the robustness to such generic nonidealities. HWA training is only performed once—no specific device or chip characteristics, such as failure maps, are taken into account during HWA training, so resulting models remain widely deployable. This HWA-trained model is then programmed onto AIMC multiple times (here in simulation) and DNN accuracy is evaluated over time, taking into account conductance drift of PCM devices and read noise accumulation31.
Fig. 1 Illustration of the HWA training approach. In our hardware-aware (HWA) training setup, DNNs are first trained in 32-bit floating point (FP32), then retrained in a hardware-aware manner, by adding nonidealities and noise sources into the forward path and using SGD to improve the robustness to such generic nonidealities. HWA training is only performed once—no specific device or chip characteristics, such as failure maps, are taken into account during HWA training, so resulting models remain widely deployable. This HWA-trained model is then programmed onto AIMC multiple times (here in simulation) and DNN accuracy is evaluated over time, taking into account conductance drift of PCM devices and read noise accumulation31.CC BY · PubMed Central
Fig. 2 Illustration of the AIMC crossbar-model abstraction. Our analog in-memory computing (AIMC) crossbar model (using the nonideal matrix-vector multiplication (MVM) of Eqs. (1) and (2) together with hardware-calibrated PCM conductance noise and drift Eqs. (7)–(14) in “Methods”) assumes that each array or “tile” approximates MVM y~≈Wx, where digital inputs are converted with a digital-to-analog converter (DAC) to voltages, and current is integrated while weights are represented as conductances. Analog outputs are converted back from physical units to floating point (FP), using column-wise parallel analog-to-digital converters (ADCs), output scale vector γ and bias vector β. Analog weight, input, and output ranges remain fixed; digital scales are used to map the FP weight values to the analog weights (ie. normalized conductances) of the crossbar and scale the ADC output ticks per column appropriately for subsequent (digital) layers. Negative weights are programmed onto a different conductance for current subtraction in the evaluation phase, and output noise is fully represented (see Eqs. (1) and (2)).
Fig. 2 Illustration of the AIMC crossbar-model abstraction. Our analog in-memory computing (AIMC) crossbar model (using the nonideal matrix-vector multiplication (MVM) of Eqs. (1) and (2) together with hardware-calibrated PCM conductance noise and drift Eqs. (7)–(14) in “Methods”) assumes that each array or “tile” approximates MVM y~≈Wx, where digital inputs are converted with a digital-to-analog converter (DAC) to voltages, and current is integrated while weights are represented as conductances. Analog outputs are converted back from physical units to floating point (FP), using column-wise parallel analog-to-digital converters (ADCs), output scale vector γ and bias vector β. Analog weight, input, and output ranges remain fixed; digital scales are used to map the FP weight values to the analog weights (ie. normalized conductances) of the crossbar and scale the ADC output ticks per column appropriately for subsequent (digital) layers. Negative weights are programmed onto a different conductance for current subtraction in the evaluation phase, and output noise is fully represented (see Eqs. (1) and (2)).CC BY · PubMed Central
Fig. 3 Nonidealities of the AIMC crossbar model. A Correlations between analog in-memory computing (AIMC) outputs—for matrix-vector multiplications (MVMs) performed between Gaussian random weight matrices and uniform random inputs—and the ideal (FP32) expected results reveal significant deviations, due primarily to weight-programming errors and PCM conductance drift (shown here without any mean-drift compensation36). B Short-term noise sources induce cycle-to-cycle noise for repeated MVM calculations even with the same programmed weight matrix. C “IR-drops” due to finite wire resistance result from input-position dependency of the accumulated AIMC column-currents. An expected 0 output—the correct result when a linearly-graded weight matrix (ranging from −1 to 1 in order) is read with a constant input on all rows—can actually deviate drastically due to the degree of ordering of weights here shown from 0 (completely unordered, typical case) to 1 (fully ordered, extreme case). D The MVM error ϵM^(*) Eq. (20) of our standard PCM-based AIMC inference model (Fig. 2; dotted lines) ≈15% roughly corresponds to fixed-point quantized digital (solid lines) at ~4 bits. E Correlations of MVM deviation, yi~−yi, vs. desired MVM output yi = wijxj illustrate the importance of proper input scaling α, for wij~N(0,0.246) and xj~N(0,1). Red dots mimic the weight-to-activation correlations that SGD learning will produce, using x~j=ρwkj+(1−ρ)xj, while blue dots represent the uncorrelated component for comparison, (ρ = 0.0). Low α = 1 leads to input clipping (a, e, i) and ϵM^(*) exceeding 35% (gray text). Intermediate α values can still lead to saturated outputs for correlated inputs, even without input clipping (f, j); excessive α values reduce clipping but increase ϵM^(*) dramatically (d, h, l). We optimize α during hardware-aware (HWA) training, then keep it fixed during AIMC inference, minimizing ϵM^(*) regardless of input correlation (c, g, k).
Fig. 3 Nonidealities of the AIMC crossbar model. A Correlations between analog in-memory computing (AIMC) outputs—for matrix-vector multiplications (MVMs) performed between Gaussian random weight matrices and uniform random inputs—and the ideal (FP32) expected results reveal significant deviations, due primarily to weight-programming errors and PCM conductance drift (shown here without any mean-drift compensation36). B Short-term noise sources induce cycle-to-cycle noise for repeated MVM calculations even with the same programmed weight matrix. C “IR-drops” due to finite wire resistance result from input-position dependency of the accumulated AIMC column-currents. An expected 0 output—the correct result when a linearly-graded weight matrix (ranging from −1 to 1 in order) is read with a constant input on all rows—can actually deviate drastically due to the degree of ordering of weights here shown from 0 (completely unordered, typical case) to 1 (fully ordered, extreme case). D The MVM error ϵM^(*) Eq. (20) of our standard PCM-based AIMC inference model (Fig. 2; dotted lines) ≈15% roughly corresponds to fixed-point quantized digital (solid lines) at ~4 bits. E Correlations of MVM deviation, yi~−yi, vs. desired MVM output yi = wijxj illustrate the importance of proper input scaling α, for wij~N(0,0.246) and xj~N(0,1). Red dots mimic the weight-to-activation correlations that SGD learning will produce, using x~j=ρwkj+(1−ρ)xj, while blue dots represent the uncorrelated component for comparison, (ρ = 0.0). Low α = 1 leads to input clipping (a, e, i) and ϵM^(*) exceeding 35% (gray text). Intermediate α values can still lead to saturated outputs for correlated inputs, even without input clipping (f, j); excessive α values reduce clipping but increase ϵM^(*) dramatically (d, h, l). We optimize α during hardware-aware (HWA) training, then keep it fixed during AIMC inference, minimizing ϵM^(*) regardless of input correlation (c, g, k).CC BY · PubMed Central
Fig. 4 Comparison of AIMC nonidealities. Comparison of the relative impact of various analog in-memory computing (AIMC) nonidealities on DNN accuracy. A AIMC deviations (y~−y) from the ideal matrix-vector multiplication (MVM) output (y) are shown, for uncorrelated (blue dots) and weakly correlated (ρ = 0.05) random activations (red dots), as a single nonideality is increased until standard MVM error reaches ϵM^(*)=20%. All other parameters remain fixed to our standard crossbar model (ϵM^(*)=15%, Fig. 3). For instance, IR-drop needs to be scaled 11.7 × to incur ϵM^(*)=20%. Even at constant ϵM^(*)=20%, MVM deviations are structured differently and thus the impact on DNN accuracy can vary significantly. B Grid shows loss in normalized accuracy (A^1h) over the base HWA-trained model at 1 h after programming when boosting a given nonideality to ϵM^(*)=20%. Thus 0% means no accuracy impact despite the amplified nonideality, whereas 100% means a drop to chance level. For the HMM, LSTM sensitivity is reported for a portion of the training set (instead of the benchmark set) directly on the LSTM output without the hidden Markov model to speed up computations. For the transformer models, only one GLUE task is evaluated (MRPC).
Fig. 4 Comparison of AIMC nonidealities. Comparison of the relative impact of various analog in-memory computing (AIMC) nonidealities on DNN accuracy. A AIMC deviations (y~−y) from the ideal matrix-vector multiplication (MVM) output (y) are shown, for uncorrelated (blue dots) and weakly correlated (ρ = 0.05) random activations (red dots), as a single nonideality is increased until standard MVM error reaches ϵM^(*)=20%. All other parameters remain fixed to our standard crossbar model (ϵM^(*)=15%, Fig. 3). For instance, IR-drop needs to be scaled 11.7 × to incur ϵM^(*)=20%. Even at constant ϵM^(*)=20%, MVM deviations are structured differently and thus the impact on DNN accuracy can vary significantly. B Grid shows loss in normalized accuracy (A^1h) over the base HWA-trained model at 1 h after programming when boosting a given nonideality to ϵM^(*)=20%. Thus 0% means no accuracy impact despite the amplified nonideality, whereas 100% means a drop to chance level. For the HMM, LSTM sensitivity is reported for a portion of the training set (instead of the benchmark set) directly on the LSTM output without the hidden Markov model to speed up computations. For the transformer models, only one GLUE task is evaluated (MRPC).CC BY · PubMed Central
Fig. 5 Specifications of AIMC nonidealities. Tolerances of individual analog in-memory computing (AIMC) nonidealities across DNNs. A As a single nonideality parameter is increased from the our standard setting, accuracy A^1h eventually drops to 99% (compared to accuracy of the standard AIMC crossbar model). Four nonidealities are shown, with DNN line-colors matching the text-label color in (B). B Grid shows x*, the threshold value at which that particular nonideality produces A^1h=99% (DNN’s curve crosses dotted line in (A)). For instance, reducing DAC precision from 8-bit down to 5-bit, while maintaining all other parameters from the standard AIMC crossbar model, causes exactly 1% additional accuracy loss in the LSTM-PTB model. Text-label colors at the top match the lines in (A); grid colors reflect relative sensitivity index rs=(x^(*)−minx^(*))/(maxx^(*)−minx^(*)), with min and max values taken across all DNNs. rs = 1 (red) indicates the most-sensitive and rs = 0 (green) the least-sensitive DNN. RNNs are generally observed to be more robust to AIMC nonidealities than CNNs, even with the limited hyper-parameter tuning available for RNN-T due to its large number of MVM FLOPS and parameters.
Fig. 5 Specifications of AIMC nonidealities. Tolerances of individual analog in-memory computing (AIMC) nonidealities across DNNs. A As a single nonideality parameter is increased from the our standard setting, accuracy A^1h eventually drops to 99% (compared to accuracy of the standard AIMC crossbar model). Four nonidealities are shown, with DNN line-colors matching the text-label color in (B). B Grid shows x*, the threshold value at which that particular nonideality produces A^1h=99% (DNN’s curve crosses dotted line in (A)). For instance, reducing DAC precision from 8-bit down to 5-bit, while maintaining all other parameters from the standard AIMC crossbar model, causes exactly 1% additional accuracy loss in the LSTM-PTB model. Text-label colors at the top match the lines in (A); grid colors reflect relative sensitivity index rs=(x^(*)−minx^(*))/(maxx^(*)−minx^(*)), with min and max values taken across all DNNs. rs = 1 (red) indicates the most-sensitive and rs = 0 (green) the least-sensitive DNN. RNNs are generally observed to be more robust to AIMC nonidealities than CNNs, even with the limited hyper-parameter tuning available for RNN-T due to its large number of MVM FLOPS and parameters.CC BY · PubMed Central
Fig. 6 Compactness of conductance values distributions. The hardware-aware (HWA) training reduces matrix-vector multiplication (MVM) error by creating more compact conductance distributions. A MVM error decreases as constructed conductance distributions, produced by a generalized normal distribution Eq. (4), are made more compact by increasing β. Example distributions in red at top show β = 1 (Laplace distribution), β = 2 (normal distribution), and even more compact distributions for higher β. “SEM̀” indicates standard error of the mean. B Data from (A) is replotted as a function of the (excess) kurtosis of the distribution. According to the definition of excess kurtosis, a normal distribution (that is β = 2 in A) has a value of 0, and positive or negative values for longer tail distributions (i.e., β < 2) or more compact distributions (ie. β > 2), respectively. Note that longer tail distributions (large kurtosis) lead to higher MVM error, while more compact distributions (lower kurtosis) reduce MVM error (C) Kurtosis of the conductance values per layer, comparing HWA-trained models (solid bars), to FP32 weight data scaled by the overall absolute maximum weight (hashed bars). Column-wise scaling, and the tuning of both weights and scaling parameters during HWA training, help lead to significantly more compact distributions with smaller kurtosis values. Box plot shows first quartile and third quartile with a line at the median. The whiskers extend 1.5× the interquartile range. Outlier points are those past the end of the whiskers.
Fig. 6 Compactness of conductance values distributions. The hardware-aware (HWA) training reduces matrix-vector multiplication (MVM) error by creating more compact conductance distributions. A MVM error decreases as constructed conductance distributions, produced by a generalized normal distribution Eq. (4), are made more compact by increasing β. Example distributions in red at top show β = 1 (Laplace distribution), β = 2 (normal distribution), and even more compact distributions for higher β. “SEM̀” indicates standard error of the mean. B Data from (A) is replotted as a function of the (excess) kurtosis of the distribution. According to the definition of excess kurtosis, a normal distribution (that is β = 2 in A) has a value of 0, and positive or negative values for longer tail distributions (i.e., β < 2) or more compact distributions (ie. β > 2), respectively. Note that longer tail distributions (large kurtosis) lead to higher MVM error, while more compact distributions (lower kurtosis) reduce MVM error (C) Kurtosis of the conductance values per layer, comparing HWA-trained models (solid bars), to FP32 weight data scaled by the overall absolute maximum weight (hashed bars). Column-wise scaling, and the tuning of both weights and scaling parameters during HWA training, help lead to significantly more compact distributions with smaller kurtosis values. Box plot shows first quartile and third quartile with a line at the median. The whiskers extend 1.5× the interquartile range. Outlier points are those past the end of the whiskers.CC BY · PubMed Central
Fig. 7 Layer-wise noise impact for ImageNet CNNs. A Bar-charts reveal the relative impact that different DNN layers have on AIMC accuracy when the PCM conductances in just that layer are made very noisy (overall PCM noise scale set to 15), while all other layers see only minimal PCM noise (overall noise scale set to 0). The height of each bar-segment, arranged in sequential DNN layer order, corresponds to the relative impact of that layer; colors simply delineate layer boundaries. Note that ResNet-50 and WideResNet-50 have very similar graphs since their layers only differ in width. B Accuracy A*^1h as the most critical layers in these four CNNs are exempted from PCM noise, plotted as the fraction of noise-exempt layers is increased, in order from most-sensitive to least-sensitive. C Corresponding cumulative fraction of weight-parameters that are PCM-noise-exempted. For ResNet-50, ResNet-18, and DenseNet-121 reducing PCM noise in just a few layers (dotted vertical lines) allows an AIMC crossbar model to achieve iso-accuracy (dashed horizontal line). However, for WideResNet, even non-PCM nonidealities would need to be improved since the residual system noise already causes a significant ( ≈ 2%) drop in accuracy.
Fig. 7 Layer-wise noise impact for ImageNet CNNs. A Bar-charts reveal the relative impact that different DNN layers have on AIMC accuracy when the PCM conductances in just that layer are made very noisy (overall PCM noise scale set to 15), while all other layers see only minimal PCM noise (overall noise scale set to 0). The height of each bar-segment, arranged in sequential DNN layer order, corresponds to the relative impact of that layer; colors simply delineate layer boundaries. Note that ResNet-50 and WideResNet-50 have very similar graphs since their layers only differ in width. B Accuracy A*^1h as the most critical layers in these four CNNs are exempted from PCM noise, plotted as the fraction of noise-exempt layers is increased, in order from most-sensitive to least-sensitive. C Corresponding cumulative fraction of weight-parameters that are PCM-noise-exempted. For ResNet-50, ResNet-18, and DenseNet-121 reducing PCM noise in just a few layers (dotted vertical lines) allows an AIMC crossbar model to achieve iso-accuracy (dashed horizontal line). However, for WideResNet, even non-PCM nonidealities would need to be improved since the residual system noise already causes a significant ( ≈ 2%) drop in accuracy.CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 3 Not valid · 6 Not checked · 81 Own results and interpretations

64 more claims

Create a free account to see all 100 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 15
  • It may state two separate claims as one. 17
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 81 references

Research works printed but not linked: 8

  1. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R. & Bengio, Y. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research 1, 6869–6898 (2017).
  2. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014).
  3. Goodfellow, I., Warde-Farley, D., Mirza, M., Courville, A. & Bengio, Y. Maxout networks. In International conference on machine learning 1319–1327 (PMLR, 2013).
  4. Noh, H., You, T., Mun, J. & Han, B. Regularizing deep neural networks by noise: its interpretation and optimization. in Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, (Red Hook, NY, USA) 5115–5124 (Curran Associates Inc., 2017).
  5. Rakin, A. S., He, Z. & Fan, D. Parametric noise injection: trainable randomness to improve deep neural network robustness against adversarial attack. in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 588–597 (IEEE, 2018).
  6. Wan, L., Zeiler, M., Zhang, S., Le Cun, Y. & Fergus, R. Regularization of neural networks using drop connect. in Proceedings of the 30th International Conference on Machine Learning (eds Dasgupta, S. & McAllester, D.) Vol. 28 of Proceedings of Machine Learning Research, 1058–1066 (PMLR, 2013).
  7. Glorot, X. & Bengio, Y. Understanding the difficulty of training deep feedforward neural networks. in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics 249–256 (JMLR Workshop and Conference Proceedings, 2010).
  8. Godfrey, J. J. & Holliman, E. Switchboard-1 release 2 ldc97s62. 926, 927 (1997).

Cited by

See all 213

Metadata

Authors
Rasch, M. J., Mackin, C., Le Gallo, M., Chen, A., Fasoli, A., Odermatt, F., Li, N., Nandakumar, S. R., Narayanan, P., Tsai, H., Burr, G. W., Sebastian, A., & Narayanan, V.
Affiliations
  • IBM (United States) (US)
  • IBM Research - Almaden (US)
  • IBM Research - Zurich (CH)
Corresponding author
Rasch, Malte J.
Venue
Nature Communications
Published
2023-08-30
Volume
14
Issue
1
Pages
5282-5282
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4386293192
Cited by
213
References
81
Citations by year
2026: 83 · 2025: 70 · 2024: 46 · 2023: 14
FWCI
15.35
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
Rensselaer Polytechnic Institute
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
Keywords
analog in-memory computing, deep learning inference, non-idealities, inference accuracy, energy-efficient accelerators, recurrent neural networks, transformer models, convolutional neural networks, matrix-vector multiplication

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).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

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: Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators. https://app.cochat.ai/papers/10.1038/s41467-023-40770-4
The paper
Rasch, M. J., Mackin, C., Le Gallo, M., Chen, A., Fasoli, A., Odermatt, F., Li, N., Nandakumar, S. R., Narayanan, P., Tsai, H., Burr, G. W., Sebastian, A., & Narayanan, V. (2023). Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators. Nature Communications. https://doi.org/10.1038/s41467-023-40770-4

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.