Resistive Crossbars as Approximate Hardware Building Blocks for Machine Learning: Opportunities and Challenges
The DOI matches its Crossref record; 173 of the 178 references it cites were found in a registry or an index.
Abstract
Traditional computing systems based on the von Neumann architecture are fundamentally bottlenecked by data transfers between processors and memory. The emergence of data-intensive workloads, such as machine learning (ML), creates an urgent need to address this bottleneck by designing computing platforms that utilize the principle of colocated memory and processing units. Such an approach, known as “in-memory computing,” can potentially eliminate data movement costs by computing inside the memory array itself. Crossbars based on resistive nonvolatile memory (NVM) devices have shown immense promise in serving as the building blocks of in-memory computing systems for ML workloads. This is because their high density can lead to higher on-chip storage capacity, while they can also perform massively parallel, in situ matrix-vector multiplication (MVM) operations, thereby accelerating the main computational kernel of ML workloads. However, resistive crossbar-based analog computing is inherently approximate due to the device- and circuit-level nonidealities. Furthermore, the area and energy costs of peripheral circuits for conversions between the analog and digital domains can greatly diminish the intrinsic efficiency of crossbar-based MVM computation. We present a comprehensive overview of the emerging paradigm of computing using NVM crossbars for accelerating ML workloads. We describe the design principles of resistive crossbars, including the devices and associated circuits that constitute them. We discuss intrinsic approximations arising from the device and circuit characteristics and study their functional impact on the MVM operation. Next, we present an overview of spatial architectures that exploit the high storage density of NVM crossbars. Furthermore, we elaborate on software frameworks that effectively capture device-circuit-architecture characteristics to evaluate the performance of large-scale deep neural networks (DNNs) using resistive crossbar-based hardware. Finally, we discuss open challenges and future research directions that need to be explored in order to realize the vision of resistive crossbars as the building blocks of future computing platforms.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
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?
In the citing paper
“However, the disadvantage of time-encoding schemes is the reduction of computation speed and hardware overhead required for partial sums computation96.”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [96]
In the citing paper
“The resolution of ADC required to represent all possible outputs of the VMM operation depends on input precision (DAC resolution), number of crossbar rows , and precision of the weights cells (conductance resolution), and can be calculated as 96.”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [96]
In the citing paper
“ADC sharing requires additional multiplexers and sample-and-hold circuits and also increases latency96 (i.e. more time is required to process each input pattern, thus reducing the throughput of the ANN).”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [96]
In the citing paper
“In binarized networks, ADC can be replaced by a 1-bit comparator96 or ADC-like multi-level sense amplifier158.”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [96]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 1×
“Deep neural networks (DNNs), in particular, have had enormous success in several ML tasks, such as image classification [2], [3], object detection [4], and natural language processing [5].”
BackgroundImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 2×
“[2] A.”
BackgroundPyTorch: An Imperative Style, High-Performance Deep Learning Library
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473 · cited 1×
“One approach to perform this is functionally abstract the nonideal MVM operation and integrate it into existing ML frameworks (such as PyTorch [127], Tensor-flow [128], and Caffe [129]).”
MethodGirshick, R. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.169 · Cited by 28,105
TensorFlow: A system for large-scale machine learning
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., et al. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1605.08695 · Cited by 11,822
Caffe: Convolutional Architecture for Fast Feature Embedding
Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., & Darrell, T. (2014) · Proceedings of the 22nd ACM international conference on Multimedia · doi:10.1145/2647868.2654889 · Cited by 11,077 · cited 1×
“One approach to perform this is functionally abstract the nonideal MVM operation and integrate it into existing ML frameworks (such as PyTorch [127], Tensor-flow [128], and Caffe [129]).”
MethodMastering the game of Go without human knowledge
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al. (2017) · Nature · doi:10.1038/nature24270 · Cited by 9,084 · cited 1×
“Rapid advances in machine learning (ML) and artificial intelligence (AI) have touched human lives in an unprecedented way as the application space has proliferated over the last decade [1].”
BackgroundAnisotropic Superexchange Interaction and Weak Ferromagnetism
Moriya, T. (1960) · Physical Review · doi:10.1103/physrev.120.91 · Cited by 7,234 · cited 1×
“While the stability of the Neel wall is aided by the Dzyaloshinskii–Moriya interaction [67], [68], the Bloch walls need external magnetic fields for proper operation [69].”
MethodCurrent-driven excitation of magnetic multilayers
Slonczewski, J. C. (1996) · Journal of Magnetism and Magnetic Materials · doi:10.1016/0304-8853(96)00062-5 · Cited by 6,825
A thermodynamic theory of “weak” ferromagnetism of antiferromagnetics
Dzyaloshinsky, I. (1958) · Journal of Physics and Chemistry of Solids · doi:10.1016/0022-3697(58)90076-3 · Cited by 6,259 · cited 1×
“While the stability of the Neel wall is aided by the Dzyaloshinskii–Moriya interaction [67], [68], the Bloch walls need external magnetic fields for proper operation [69].”
Method
Research works printed but not linked: 3
- agarwal, 2020, CrossSim
- han, 2015, Learning both weights and connections for efficient neural network, Proc Adv Neural Inf Process Syst, 1135
- 2015, 2015 ITRS 2 0 Official Publication zip
Cited by
Items 101–140 of 140
CODEX: Stochastic Encoding Method to Relax Resistive Crossbar Accelerator Design RequirementsPreprint
Amirsoleimani, A. (2021) · doi:10.36227/techrxiv.14485215 · Cited by 1
Lee, C., & Kim, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2101.07864
“According to the type of approaches to model crossbar array [1], analytical inference model [10, 11] depicts the model as matrix and its linear operation.”
BackgroundMethodSahu, U., Sisodia, N., Sharda, J., Muduli, P. K., & Bhowmik, D. (2021) · doi:10.36227/techrxiv.14763498.v1
2021 Index Proceedings of the IEEE Vol. 107-109
(2021) · Proceedings of the IEEE · doi:10.1109/jproc.2021.3135169
CODEX: Stochastic Encoding Method to Relax Resistive Crossbar Accelerator Design RequirementsPreprint
Amirsoleimani, A., Liu, T., Xu, J., Alibart, F., Beilliard, Y., Ecoffey, S., Drouin, D., & Genov, R. (2021) · doi:10.36227/techrxiv.14485215.v1
Sahu, U., Sisodia, N., Sharda, J., Muduli, P. K., & Bhowmik, D. (2021) · doi:10.36227/techrxiv.14763498
2022 Index Proceedings of the IEEE Vol. 108-110
(2022) · Proceedings of the IEEE · doi:10.1109/jproc.2023.3234419
SEVDA: Singular Value Decomposition Based Parallel Write Scheme for Memristive CNN Accelerators
Al-Shaarawy, A., Genov, R., & Amirsoleimani, A. (2023) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas46773.2023.10182097
“Since device conductances cannot be negative, weights must be stored in a sign-magnitude form, where polarity is represented by two separate devices [13], as illustrated in figure 1b. FC layers are the simplest operation to map onto the crossbar, as they can be represented as a simple matrix-vector…”
BackgroundIntroduction to Artificial Neural Networks (ANN) and Spiking Neural Networks (SNN)
Bhowmik, D. (2024) · Springer tracts in electrical and electronics engineering · doi:10.1007/978-981-97-4445-9_3 · Cited by 1
Why Spintronics-Based Neuromorphic Computing?
Bhowmik, D. (2024) · Springer tracts in electrical and electronics engineering · doi:10.1007/978-981-97-4445-9_1
The Ferromagnetic Domain-Wall Synapse Device
Bhowmik, D. (2024) · Springer tracts in electrical and electronics engineering · doi:10.1007/978-981-97-4445-9_4
Unearthing the Potential of Spiking Neural Networks
Chowdhury, S. S., Kosta, A. K., Sharma, D., Apolinario, M. P. E., & Roy, K. (2024) · Design, Automation, and Test in Europe · doi:10.23919/date58400.2024.10546699 · Cited by 1
“Note, the power consumption and area can be dominated by the high precision ADCs [48]. optimized SNNs have the potential to result in lightweight and efficient architectures, making them well-suited for edge applications.”
BackgroundSelf-citationHarnessing Valley-Spin Hall Effect in WSe2 for Energy-Efficient Logic-Memory Integration
Gupta, S. K., & Cho, K. (2024) · Spintronics · doi:10.1117/12.3028292
“To that end, several works (such as [33]) have explored implementing in-memory matrix-vector multiplication - a dominant kernel in AI workloads - and have established energy efficiency and performance benefits of in-memory computing.”
Gupta, S., Yadav, M. S., Tavva, V. K., & Rawat, B. (2024) · International Semiconductor Conference · doi:10.1109/cas62834.2024.10736761
“Despite recent works have reported various memory circuit/architecture topologies for improving reading accuracy and latency, limited investigation has been conducted into the energy efficiency of 1T-1M device [10].”
BackgroundEnhancement of Synaptic Properties using Ta2O5/ZnO/Ta2O5 Tri-layer Device
Yeom, C.-M., Kim, S.-H., Kim, Y.-B., Kim, D., Roh, K., Kim, Y., Kim, Y.-G., Kwon, H.-M., & Lee, H. (2025) · IEEE International Conference on Microelectronic Test Structures (ICMTS) · doi:10.1109/icmts63811.2025.11068932
Design for Reliability of Multi-Bit Operations in RRAM-Based SIMPLY Logic-in-Memory Circuits
Zanotti, T., Borellini, E., Vatalaro, M., Maccaronio, V., Pavan, P., Rose, R. D., & Puglisi, F. M. (2025) · IEEE International Conference on IC Design and Technology (ICICDT) · doi:10.1109/icicdt65192.2025.11078043
“In-memory computing with Resistive Random Access Memory (RRAM) devices represents a promising technology for implementing the next generation of ultra-low power edge computing solutions [1], [2].”
Integrated Electroforming of Memristor Cells in Crossbar Arrays
Heinen, R., Zoche, J., Wunderlich, R., & Heinen, S. J. (2025) · IEEE Nordic Circuits and Systems Conference (NorCAS) · doi:10.1109/norcas66540.2025.11231282
“This process leverages Ohm’s law and Kirchhoff’s current law to perform multiply and accumulate operations in parallel across all rows and columns [3].”
Ren, T., Bi, W., Wang, L., & Zhao, Y. (2025) · 2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD) · doi:10.1109/iccad66269.2025.11240930
“In the CIM architecture, to provide theoretically complete precision guarantees from the analog MVM, the ADC must have a unique level for each possible output of the MVM operation, ensuring that no information is lost during digitization [23].”
Kim, D., Sharma, T., Mukherjee, A., Mukherjee, M., & Roy, K. (2025) · International Symposium on Low Power Electronics and Design (ISLPED) · doi:10.1109/islped65674.2025.11261802 · Cited by 1
“A typical crossbar structure consists of a memory array of 1-transistor, 1-resistor (eNVM 1T-1R) configuration bit-cells, along with peripheral components such as digital-to-analog converters (DACs), transimpedance amplifiers (TIAs), analog-to-digital converters (ADCs), and shift-add circuits [16].”
Self-citationYan, Y., Li, S., Zhang, W., Si, X., & Yoshioka, K. (2025) · IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia) · doi:10.1109/icce-asia67487.2025.11263505
“2(c) shows, the peripheral analog-to-digital converter (ADC) is a well-known bottleneck in analog CIM systems [12].”
Житяев, И. Л., Kartel, M. S., Житяева, Ю. Ю., & Смирнов, В. А. (2025) · Nanoindustry Russia · doi:10.22184/1993-8578.2025.18.11s.480.485
Exploration of Neural Networks and Future Acceleration Through Hardware
Luo, Y. (2024) · International Conference on Applied Machine Learning (ICAML) · doi:10.1109/icaml64299.2024.00059
“Ref [8] overview of accelerating machine learning workloads in in-memory computing systems using resistive nonvolatile memory crossbars.”
Abarra, Y., Misawa, N., Matsui, C., & Takeuchi, K. (2026) · Japanese Journal of Applied Physics · doi:10.35848/1347-4065/ae4880
Jadia, H. K., Ghosh, A., Ali, M. H., Uddin, S. F., Sathish, N., Mandal, S. K., Raut, N., Mulaosmanovic, H., Dünkel, S., Beyer, S., et al. (2026) · Neuromorphic Computing and Engineering · doi:10.1088/2634-4386/ae4cc5 · Cited by 3
Self-citationBayesian neuromorphic computing based on resistive memory
Dalgaty, T. (2020) · theses.fr (ABES) · doi:10.70675/e5b4f3e2z44a5z4c41zbac6z4c933e86b101 · Cited by 1
da Cunha, A. C. W. (2023) · theses.fr (ABES) · doi:10.70675/138d4d1dz87daz4754z9cf4za4514aff5c49
Neuromorphic in-memory learning with analog integrated circuits and nanoscale memristive devices
Garg, N. (2024) · theses.fr (ABES) · doi:10.70675/a5f34c0ezcc3bz42c8z8a52z529a545221ca · Cited by 2
Viollet, V., Prenat, G., Muktash, A. A., Popon, C., Anghel, L., & Lecoq, X. (2026) · IEEE Latin American Test Symposium (LATS) · doi:10.1109/lats70329.2026.11480313
“A-IMC faces limitations due to peripheral circuitry, especially analog-to-digital converters (ADC), which add significant energy, latency, and area overheads that could even outweigh its benefits [13].”
Sadashiva, A., Holla, A., & Bhowmik, D. (2026) · Lecture notes in networks and systems · doi:10.1007/978-981-96-9983-4_12
Zuo, Y., Zexin, F., Wang, C., Wu, Y., Huang, J., & Ma, Y. (2026) · International Symposium on Quality Electronic Design (ISQED) · doi:10.1109/isqed69900.2026.11534775
“As discussed in [37]–[39], RRAM suffers from significantly higher latency (1–2 orders of magnitude) compared to SRAM at the same technology node and has limited endurance (approximately 10 6 – 10 9 conservative writes), which severely impacts their lifetime and constrains their applicability to…”
Dyn-A-Map: Dynamic AI-Model Mapping for robust and accurate MRAM-based analog edge computing
Chowdhury, M. A., Zand, R., & Salehi, S. (2026) · npj Unconventional Computing · doi:10.1038/s44335-026-00072-6
Farias, M., Martins, W. S., & Kung, H. T. (2026) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas66217.2026.11562844
“Crossbar nonidealities are quantified by the NF [29]–[31], (cid:31)(cid:31)(cid:31)(cid:31) (cid:31)(cid:31)(cid:31)(cid:31) where i 0 is the expected output and ∆ i is the amount of current that was deviated due to nonidealities.”
Memristors for Advanced Computation and Memory Applications
Ahmad, G., Khalid, M., & Wan, M. (2026) · Springer eBooks · doi:10.1007/978-981-96-8867-8_18-1
Energy‐Aware Perturbation Optimization for Memristor‐Array Convolutional Neural Networks
Xu, A., Lu, L., Wu, Y., Guo, Y., Li, Y., Wang, Y., & Li, Q. (2026) · Advanced Intelligent Systems · doi:10.1002/aisy.70477 · Cited by 1
Liu, Y., Sun, L. C., Zeng, R., Huang, M. C. Y., & Wu, H. X. (2026) · International Symposium on Low Power Electronics and Design (ISLPED) · doi:10.1145/3816440.3818580
“A common implementation uses resistive crossbar arrays [8–13 , 20– 25 Despite this, CMOS resistive crossbar designs are burdened by nonidealities that degrade their accuracy and reliability.”
Anonymous, Reddy, G. V. S., Tulapurkar, A. A., & Bhowmik, D. (2026) · Physical Review Applied · doi:10.1103/6958-v2np
Wafer-scale SOT-MRAM for analog crossbar array applications
Liu, S., Rogers, V., Hu, C.-Y., Song, M.-Y., Bao, X., & Incorvia, J. (2026) · Science Advances · doi:10.1126/sciadv.aee6952
Gökgöz, B. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7385571
ADC-Free Current-Domain Compute-In-Memory Processors for Intelligent Image Sensing
Gupte, K. K., Chen, D. J., & Wang, C. (2026) · IEEE International Midwest Symposium on Circuits and Systems (MWSCAS) · doi:10.1109/mwscas67364.2026.11680770
“Unlike conventional crossbar implementations, where current magnitude depends directly on conductance values [10], the output current in the proposed design is largely unaffected as long as the switching thresholds are maintained.”
Xu, L., Huang, M., Funni, S. D., Cha, J. J., & Yildiz, B. (2026) · ACS Nano · doi:10.1021/acsnano.6c09861
Metadata
- Authors
- Chakraborty, I., Ali, M., Ankit, A., Jain, S., Roy, S., Sridharan, S., Agrawal, A., Raghunathan, A., & Roy, K.
- Affiliations
- Purdue University West Lafayette (US)
- IBM (United States) (US)
- ORCID
- Venue
- Proceedings of the IEEE
- Published
- 2020-12
- Volume
- 108
- Issue
- 12
- Pages
- 2276-2310
- ISSN
- 0018-9219 (ISSN-L)
- 1558-2256
- OpenAlex
- W3042468132
- Cited by
- 140
- References
- 178
- Citations by year
- 2026: 18 · 2025: 31 · 2024: 27 · 2023: 20 · 2022: 22 · 2021: 18 · 2020: 3
- FWCI
- 5.41
- Type
- journal article
- Language
- en
- Open access
- closed
- Licence
- Locations
- Proceedings of the IEEE (publishedVersion)
- Topics
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Keywords
- in-memory computing, machine learning hardware acceleration, matrix-vector multiplication, analog computing, approximate computing, analog-to-digital conversion, spatial architecture, deep neural network inference
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
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.