CoChat

Gradient descent-based programming of analog in-memory computing cores

Open access
Büchel, J., Vasilopoulos, A., Kersting, B., Odermatt, F., Brew, K., Ok, I., Choi, S., Saraf, I., Chan, V., Philip, T., Saulnier, N., Narayanan, V., Gallo, M. Le, Sebastian, A. · 2022 · 2022 International Electron Devices Meeting (IEDM)

The DOI matches its Crossref record; 1 of the 1 references it cites were found in a registry or an index.

Abstract

The precise programming of crossbar arrays of unit-cells is crucial for obtaining high matrix-vector-multiplication (MVM) accuracy in analog in-memory computing (AIMC) cores. We propose a radically different approach based on directly minimizing the MVM error using gradient descent with synthetic random input data. Our method significantly reduces the MVM error compared with conventional unit-cell by unit-cell iterative programming. It also eliminates the need for high-resolution analog-to-digital converters (ADCs) to read the small unit-cell conductance during programming. Our method improves the experimental inference accuracy of ResNet-9 implemented on two phase-change memory (PCM)-based AIMC cores by 1.26%.

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/iedm45625.2022.10019486

Cites

1 of 1 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 (a) In standard iterative programming, conductance values of the crossbar are read out ( 𝐺 ) and pulses ( 𝑈 ) proportional to the difference between the readout and target conductances ( 𝐺 ) are used to update the crossbar conductances ( 𝐺 ). (b) Gradient descent-based programming (GDP) uses gradients of the MVM error to generate programming pulses that iteratively bring the crossbar conductances closer to their targets. (c) Pseudocode of GDP.
Fig. 1 (a) In standard iterative programming, conductance values of the crossbar are read out ( 𝐺 ) and pulses ( 𝑈 ) proportional to the difference between the readout and target conductances ( 𝐺 ) are used to update the crossbar conductances ( 𝐺 ). (b) Gradient descent-based programming (GDP) uses gradients of the MVM error to generate programming pulses that iteratively bring the crossbar conductances closer to their targets. (c) Pseudocode of GDP.CC-BY · From the PDF
Image 2
CC-BY · From the PDF
Fig. 5 Comparison between the iterative programming algorithm (a) and GDP (b). GDP reduces the per unit-cell weight error and total MVM error.
Fig. 5 Comparison between the iterative programming algorithm (a) and GDP (b). GDP reduces the per unit-cell weight error and total MVM error.CC-BY · From the PDF
Fig. 7 Programming algorithm when using two devices per polarity. If the weight fits on just one device, we choose the one with highest SET conductance for programming and RESET the other. If both devices cannot accommodate the weight, we do not program the weight and leave both devices in SET state. Finally, if both devices are needed to fit the weight, we choose the device with lowest SET conductance for programming. The conductance of a single device is denoted 𝑔 .
Fig. 7 Programming algorithm when using two devices per polarity. If the weight fits on just one device, we choose the one with highest SET conductance for programming and RESET the other. If both devices cannot accommodate the weight, we do not program the weight and leave both devices in SET state. Finally, if both devices are needed to fit the weight, we choose the device with lowest SET conductance for programming. The conductance of a single device is denoted 𝑔 .CC-BY · From the PDF
Image 5
CC-BY · From the PDF
Image 6
CC-BY · From the PDF
Fig. 6 We compute the weight error (y-axis) between the target conductances ( G ) and the readout conductances ( 𝐺 ), as well as the estimated conductances ( 𝐺 ) and compare iterative programming (a) to GDP (b). Iterative programming leads to a lower distance between 𝐺 and 𝐺 than GDP, but higher distance between 𝐺 and 𝐺 . Hence, GDP finds a set of readout weights that actually produces a larger error w.r.t. the target weights, but still yields lower MVM error. Purposefully programming the weights 'away' from the target weights indicates that GDP accounts for small nonlinearities in the tile.
Fig. 6 We compute the weight error (y-axis) between the target conductances ( G ) and the readout conductances ( 𝐺 ), as well as the estimated conductances ( 𝐺 ) and compare iterative programming (a) to GDP (b). Iterative programming leads to a lower distance between 𝐺 and 𝐺 than GDP, but higher distance between 𝐺 and 𝐺 . Hence, GDP finds a set of readout weights that actually produces a larger error w.r.t. the target weights, but still yields lower MVM error. Purposefully programming the weights 'away' from the target weights indicates that GDP accounts for small nonlinearities in the tile.CC-BY · From the PDF
Fig. 8 Illustration of 𝜖 and 𝜖 for single-device (SD) and twodevice (TD) GDP for 500 iterations. When GDP is used with two devices, the gap between the total and nonlinear error is closed. GDP with only one device reaches the two-device iterative baseline.
Fig. 8 Illustration of 𝜖 and 𝜖 for single-device (SD) and twodevice (TD) GDP for 500 iterations. When GDP is used with two devices, the gap between the total and nonlinear error is closed. GDP with only one device reaches the two-device iterative baseline.CC-BY · From the PDF
Image 9
CC-BY · From the PDF
Image 10
CC-BY · From the PDF
Image 11
CC-BY · From the PDF
Image 12
CC-BY · From the PDF
Image 13
CC-BY · From the PDF
Image 14
CC-BY · From the PDF
Fig. 14 GDP begins to outperform the baseline (black dashed line) at a batch size of 64 and performs best for batch sizes greater or equal to 256.
Fig. 14 GDP begins to outperform the baseline (black dashed line) at a batch size of 64 and performs best for batch sizes greater or equal to 256.CC-BY · From the PDF
Fig. 16 Decreasing the MVM error has a direct effect on downstream tasks such as neural network inference. (a) Single-device (SD) and twodevice (TD) GDP significantly boost inference accuracy on a scaled down version of ResNet-9. This increase in performance is a result of reducing the weight error (b), which is also reflected in the per-layer total error (c). For each method, we programmed the network weights and ran inference on the experimental platform one time on the full 10k images of the CIFAR-10 test dataset.
Fig. 16 Decreasing the MVM error has a direct effect on downstream tasks such as neural network inference. (a) Single-device (SD) and twodevice (TD) GDP significantly boost inference accuracy on a scaled down version of ResNet-9. This increase in performance is a result of reducing the weight error (b), which is also reflected in the per-layer total error (c). For each method, we programmed the network weights and ran inference on the experimental platform one time on the full 10k images of the CIFAR-10 test dataset.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 1 references

Research works printed but not linked: 7

  1. Lanza, 2022, Science, 376, 1066
  2. Yu, 2021, IEEE Circuits Syst. Mag, 21, 31
  3. Nandakumar, 2020, IEDM Tech. Dig., 29.4.1
  4. Papandreou, 2011, Proc. ISCAS, 329
  5. Yao, 2020, Nature, 577, 641
  6. Khaddam-Aljameh, 2022, JSSC, 57, 1027
  7. Gallo, 2022, Neuromorphic Comput. Eng, 2, 014009

Cited by

See all 19

Metadata

Authors
Büchel, J., Vasilopoulos, A., Kersting, B., Odermatt, F., Brew, K., Ok, I., Choi, S., Saraf, I., Chan, V., Philip, T., Saulnier, N., Narayanan, V., Gallo, M. L., & Sebastian, A.
Affiliations
  • IBM Research - Zurich (CH)
  • IBM Research - Albany (US)
  • IBM Research – Albany,NY,USA
  • IBM (United States) (US)
Venue
2022 International Electron Devices Meeting (IEDM)
Publisher
IEEE
Published
2022-12-03
Pages
33.1.1-33.1.4
ISSN
0163-1918 (ISSN-L)
OpenAlex
W4317793291
Cited by
18
References
1
Citations by year
2026: 3 · 2025: 3 · 2024: 7 · 2023: 5
FWCI
6.07
Type
conference paper
Language
en
Open access
Locations
  • Technical digest - International Electron Devices Meeting/Technical digest (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
  • Zenodo (CERN European Organization for Nuclear Research) (cc-by)
Funders
Horizon Europe
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 › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
Keywords
analog in-memory computing, crossbar arrays, matrix-vector multiplication, phase-change memory, inference accuracy, analog-to-digital converter resolution

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: Gradient descent-based programming of analog in-memory computing cores. https://app.cochat.ai/papers/10.1109/iedm45625.2022.10019486
The paper
Büchel, J., Vasilopoulos, A., Kersting, B., Odermatt, F., Brew, K., Ok, I., Choi, S., Saraf, I., Chan, V., Philip, T., Saulnier, N., Narayanan, V., Gallo, M. L., & Sebastian, A. (2022). Gradient descent-based programming of analog in-memory computing cores. 2022 International Electron Devices Meeting (IEDM). https://doi.org/10.1109/iedm45625.2022.10019486

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.