CONV-SRAM: An Energy-Efficient SRAM With In-Memory Dot-Product Computation for Low-Power Convolutional Neural Networks
Open accessThe DOI matches its Crossref record; 25 of the 29 references it cites were found in a registry or an index.
Abstract
This paper presents an energy-efficient static random access memory (SRAM) with embedded dot-product computation capability, for binary-weight convolutional neural networks. A 10T bit-cell-based SRAM array is used to store the 1-b filter weights. The array implements dot-product as a weighted average of the bitline voltages, which are proportional to the digital input values. Local integrating analog-to-digital converters compute the digital convolution outputs, corresponding to each filter. We have successfully demonstrated functionality (>98% accuracy) with the 10 000 test images in the MNIST hand-written digit recognition data set, using 6-b inputs/outputs. Compared to conventional full-digital implementations using small bitwidths, we achieve similar or better energy efficiency, by reducing data transfer, due to the highly parallel in-memory analog computations.
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
“Among charge-based memories, SRAM cells (a bi-stable transistor structure typically made of two CMOS inverters connected back-to-back which retains a charge concentration, see Fig. 3a for an example of the structure of a crossbar array of 6T SRAM) have been widely used for VMM75–77.”
From Hardware implementation of memristor-based artificial neural networks (2024) · cites this paper as [76]
Citations
Cites
ImageNet 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 1×
“Convolutional neural networks (CNNs) provide state-of-theart results in a wide variety of AI/ML applications, ranging from image classification [3] to speech recognition [1].”
BackgroundGradient-based learning applied to document recognition
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998) · Proceedings of the IEEE · doi:10.1109/5.726791 · Cited by 58,805 · cited 5×
“However, for the MNIST dataset with LeNet-5 CNN, we estimate [22] those to have only small contributions to the overall energy-efficiency per MAV operation, due to the high parallelism supported by our in-memory approach.”
MethodInfluentialHinton, G., Deng, L., Yu, D., Dahl, G., Mohamed, A.-R., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T., et al. (2012) · IEEE Signal Processing Magazine · doi:10.1109/msp.2012.2205597 · Cited by 10,260
DeepFace: Closing the Gap to Human-Level Performance in Face Verification
Taigman, Y., Yang, M., Ranzato, M., & Wolf, L. (2014) · 2014 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2014.220 · Cited by 6,542 · cited 1×
“Facial recognition [2] enables using our faces to get access to devices in a more intuitive manner, instead of traditional passcodes.”
BackgroundXNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
Rastegari, M., Ordonez, V., Redmon, J., & Farhadi, A. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46493-0_32 · Cited by 4,408 · cited 2×
“[6] proposed a binary-weight-network (BWN), where the filter weights (wi s) can be trained to be +1/−1 (with a common scaling factor per 3-D filter: α).”
BackgroundMethodEfficient Processing of Deep Neural Networks: A Tutorial and Survey
Sze, V., Chen, Y.-H., Yang, T.-J., & Emer, J. S. (2017) · Proceedings of the IEEE · doi:10.1109/jproc.2017.2761740 · Cited by 4,215 · cited 1×
“In the conventional all-digital implementation of CNNs [4], [5], [9], with the memory and the processing elements being physically separate, reading wi s and the partial sums from the on-chip SRAMs lead to a lot of data movement per computation [10] and, hence, make them energy-hungry.”
BackgroundEyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks
Chen, Y.-H., Krishna, T., Emer, J. S., & Sze, V. (2017) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2016.2616357 · Cited by 3,247 · cited 12×
“CNNs typically consist of a cascade of convolutional (CONV) and fully-connected (FC) layers (Fig.”
BackgroundMethodInfluentialBinaryConnect: Training Deep Neural Networks with binary weights during propagations
Courbariaux, M., Bengio, Y., & David, J.-P. (2015) · arXiv · doi:10.48550/arxiv.1511.00363 · Cited by 2,180 · cited 1×
“BWN performs better than binary connect [7], which does not incorporate the scaling factor of α per filter, and also binarized neural networks [8], where both weights and activations are constrained to ±1.”
Background1.1 Computing's energy problem (and what we can do about it)
Horowitz, M. (2014) · 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC) · doi:10.1109/isscc.2014.6757323 · Cited by 2,069 · cited 1×
“to access data from memory can be much higher than the energy needed for a compute operation with that data [11].”
BackgroundHubara, I., Soudry, D., & Yaniv, R. E. (2016) · arXiv · doi:10.48550/arxiv.1602.02505 · Cited by 1,667 · cited 1×
“BWN performs better than binary connect [7], which does not incorporate the scaling factor of α per filter, and also binarized neural networks [8], where both weights and activations are constrained to ±1.”
BackgroundIn-Memory Computation of a Machine-Learning Classifier in a Standard 6T SRAM Array
Zhang, J., Wang, Z., & Verma, N. (2017) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2016.2642198 · Cited by 532 · cited 21×
“signals have less variations typically, compared to those arising from local Vt mismatch in the pMOS devices [13].”
BackgroundMethodResultInfluentialMoons, B., Uytterhoeven, R., Dehaene, W., & Verhelst, M. (2017) · 2017 IEEE International Solid-State Circuits Conference (ISSCC) · doi:10.1109/isscc.2017.7870353 · Cited by 479 · cited 1×
“Recent hardware implementations [5], [9], [16], [19]–[21] for NNs have focused on reduced bit-precisions to achieve”
BackgroundBiswas, A., & Chandrakasan, A. P. (2018) · 2018 IEEE International Solid - State Circuits Conference - (ISSCC) · doi:10.1109/isscc.2018.8310397 · Cited by 337 · cited 1×
“To address this problem, we present an SRAM-embedded convolution computation architecture [12], conceptually shown in Fig.”
MethodSelf-citationLee, J., Kim, C., Kang, S., Shin, D., Kim, S., & Yoo, H.-J. (2018) · 2018 IEEE International Solid - State Circuits Conference - (ISSCC) · doi:10.1109/isscc.2018.8310262 · Cited by 303 · cited 3×
“Our numbers are also comparable to the energy efficiency of [20] (not quoted for MNIST), which uses 1-b for weights.”
BackgroundResultKhwa, W.-S., Chen, J.-J., Li, J.-F., Si, X., Yang, E.-Y., Sun, X., Liu, R., Chen, P.-Y., Li, Q., Yu, S., et al. (2018) · 2018 IEEE International Solid - State Circuits Conference - (ISSCC) · doi:10.1109/isscc.2018.8310401 · Cited by 260 · cited 4×
“Finally, this paper supports multi-bit resolution for the inputs and outputs of the dot products, compared to [13] (output: 1-b) and [16] and [17] (both input–output: 1-b).”
BackgroundMethodInfluentialA Multi-Functional In-Memory Inference Processor Using a Standard 6T SRAM Array
Kang, M., Gonugondla, S. K., Patil, A., & Shanbhag, N. R. (2018) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2017.2782087 · Cited by 224 · cited 6×
“In the conventional approach (with 6T SRAM bit-cells) [13], [14], [17], where multiple word-lines (WLs) are activated for the same bitline, there might be a situation where one of the accessed bit-cells in that column is in pseudo-write mode (Fig.”
BackgroundMethodResultInfluentialAn Energy-Efficient Precision-Scalable ConvNet Processor in 40-nm CMOS
Moons, B., & Verhelst, M. (2017) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2016.2636225 · Cited by 198 · cited 7×
“Basics of a typical CNN for a classification problem, showing the structure for the CONV and FC layers [4], [5].”
BackgroundMethodResultInfluentialA 42pJ/decision 3.12TOPS/W robust in-memory machine learning classifier with on-chip training
Gonugondla, S. K., Kang, M., & Shanbhag, N. (2018) · 2018 IEEE International Solid - State Circuits Conference - (ISSCC) · doi:10.1109/isscc.2018.8310398 · Cited by 195 · cited 2×
“[15] proposed an on-chip training to compensate for chip-to-chip variations.”
BackgroundMethodWhatmough, P. N., Lee, S. K., Lee, H., Rama, S., Brooks, D., & Wei, G.-Y. (2017) · 2017 IEEE International Solid-State Circuits Conference (ISSCC) · doi:10.1109/isscc.2017.7870351 · Cited by 164 · cited 5×
“Recent hardware implementations [5], [9], [16], [19]–[21] for NNs have focused on reduced bit-precisions to achieve”
BackgroundResultInfluentialAndo, K., Ueyoshi, K., Orimo, K., Yonekawa, H., Sato, S., Nakahara, H., Takamaeda-Yamazaki, S., Ikebe, M., Asai, T., Kuroda, T., et al. (2018) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2017.2778702 · Cited by 160 · cited 8×
“When compared to a near-memory approach [16], which uses only 1-b for IFMP/OFMP, we still achieve 8.”
BackgroundMethodResultInfluentialMoons, B., Bankman, D. J., Yang, L., Murmann, B., & Verhelst, M. (2018) · Custom Integrated Circuits Conference · doi:10.1109/cicc.2018.8357071 · Cited by 104
Paul, N., Lee, S. K., Lee, H., Rama, S., David, B., & Wei, G.-Y. (2017) · IEEE Conference Proceedings · Cited by 77
Energy-Efficient Reconfigurable SRAM: Reducing Read Power Through Data Statistics
Duan, C., Gotterba, A. J., Sinangil, M. E., & Chandrakasan, A. P. (2017) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2017.2731814 · Cited by 21 · cited 1×
“The results indicate that the proposed energy-efficient, SRAM-embedded dot-product computation architecture could enable low power ML applications (e.g. “always-ON” sensing) for “smart” devices in the Internet-of-Everything.”
BackgroundSelf-citationBiswas, A., & Chandrakasan, A. P. (2016) · doi:10.1109/esscirc.2016.7598334 · Cited by 12 · cited 2×
“Hence, in future work, data-dependent memory architectures e.g. 8T SRAMs, [23], [24] can be used to store and access the inputs/outputs.”
BackgroundSelf-citationEnergy-efficient smart embedded memory design for IoT and AI
Biswas, A. (2018) · DSpace@MIT (Massachusetts Institute of Technology) · Cited by 5
Self-citation
Research works printed but not linked: 1
- whatmough, 2017, A 28 nm SoC with a 1.2 GHz 568 nJ/prediction sparse deep-neural-network engine with >0.1 timing error rate tolerance for IoT applications, IEEE Int Solid-State Circuits Conf (ISSCC) Dig Tech Papers, 242
Cited by
Memory devices and applications for in-memory computing
Sebastian, A., Le Gallo, M., Khaddam-Aljameh, R., & Eleftheriou, E. (2020) · Nature Nanotechnology · doi:10.1038/s41565-020-0655-z · Cited by 2,289
Edge Computing with Artificial Intelligence: A Machine Learning Perspective
Hua, H., Li, Y., Wang, T., Dong, N., Li, W., & Cao, J. (2023) · ACM Computing Surveys · doi:10.1145/3555802 · Cited by 563
“A static random access memory (SRAM) [102] is able to reduce memory data throughput, and it combines parallel CNNs to enable simultaneous access to different memory blocks.”
BackgroundMethodInfluentialHardware implementation of memristor-based artificial neural networks
Aguirre, F., Sebastian, A., Le Gallo, M., Song, W., Wang, T., Yang, J. J., Lu, W., Chang, M.-F., Ielmini, D., Yang, Y., et al. (2024) · Nature Communications · doi:10.1038/s41467-024-45670-9 · Cited by 564
“The proposed circuit comprises 4 kb synapses fabricated in a 65 nm CMOS processand reported anenergy ef fi ciencyof 55.8 TOPS per W. Incases where x is non-binary, one approach is to employ capacitors in addition to the SRAM cells 76,77,79 , involving a three-step process.”
InfluentialIn-Memory Computing: Advances and Prospects
Verma, N., Jia, H., Valavi, H., Tang, Y., Ozatay, M., Chen, L.-Y., Zhang, B., & Deaville, P. (2019) · IEEE Solid-State Circuits Magazine · doi:10.1109/mssc.2019.2922889 · Cited by 473
“, [13] requires adopting a 10-T bit cell, and [14] requires multiple memory-operation cycles).”
BackgroundXNOR-SRAM: In-Memory Computing SRAM Macro for Binary/Ternary Deep Neural Networks
Yin, S., Jiang, Z., Seo, J.-S., & Seok, M. (2020) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2019.2963616 · Cited by 429
“For example, CONV-SRAM [20] integrates digital-toanalog converters (DACs) for analog wordlines, binary weights stored in SRAM, and analog-to-digital converters (ADCs) to convert the in-memory computation results back to digital values.”
BackgroundMethod
Metadata
- Authors
- Biswas, A., & Chandrakasan, A. P.
- Affiliations
- Texas Instruments (United States) (US)
- Massachusetts Institute of Technology (US)
- Venue
- IEEE Journal of Solid-State Circuits
- Published
- 2019-01
- Volume
- 54
- Issue
- 1
- Pages
- 217-230
- ISSN
- 0018-9200 (ISSN-L)
- 1558-173X
- OpenAlex
- W2904299207
- Cited by
- 414
- References
- 29
- Citations by year
- 2026: 36 · 2025: 57 · 2024: 75 · 2023: 75 · 2022: 74 · 2021: 59 · 2020: 24 · 2019: 13
- FWCI
- 9.17
- Type
- journal article
- Language
- en
- Open access
- green
- hdl.handle.net
- Licence
- Locations
- IEEE Journal of Solid-State Circuits (publishedVersion)
- DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc-sa)
- Funders
- Intel Corporation
- 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 › Engineering › Electrical and Electronic Engineering › Analog and Mixed-Signal Circuit Design
- Keywords
- SRAM, in-memory computing, dot product computation, energy efficiency, low-power neural network hardware, analog-to-digital converter, MNIST classification
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.