Embedded Intelligence on FPGA: Survey, Applications and Challenges
Open accessThe DOI matches its Crossref record; 90 of the 93 references it cites were found in a registry or an index; 36 of the 140 sources its claims cite were checked for support.
Abstract
Embedded intelligence (EI) is an emerging research field and has the objective to incorporate machine learning algorithms and intelligent decision-making capabilities into mobile and embedded devices or systems. There are several challenges to be addressed to realize efficient EI implementations in hardware such as the need for: (1) high computational processing; (2) low power consumption (or high energy efficiency); and (3) scalability to accommodate different network sizes and topologies. In recent years, an emerging hardware technology which has demonstrated strong potential and capabilities for EI implementations is the FPGA (field programmable gate array) technology. This paper presents an overview and review of embedded intelligence on FPGA with a focus on applications, platforms and challenges. There are four main classification and thematic descriptors which are reviewed and discussed in this paper for EI: (1) EI techniques including machine learning and neural networks, deep learning, expert systems, fuzzy intelligence, swarm intelligence, self-organizing map (SOM) and extreme learning; (2) applications for EI including object detection and recognition, indoor localization and surveillance monitoring, and other EI applications; (3) hardware and platforms for EI; and (4) challenges for EI. The paper aims to introduce interested researchers to this area and motivate the development of practical FPGA solutions for EI deployment.
Overview
CoChat Summary
TL;DR
This review surveys embedded intelligence (EI) on FPGA across techniques, applications, hardware platforms, and challenges, arguing that FPGA platforms are well suited for EI due to parallelism, flexibility, and energy efficiency. It synthesizes many cited accelerator studies and reports speed, power, and accuracy comparisons against CPU and GPU implementations.
Key findings
- FPGA is presented as a strong EI hardware candidate because it offers parallel computation, lower power consumption, short turnaround time versus ASIC, and reconfigurability; one comparison states FPGA can reach 10 times faster than CPU platforms, with the lowest level having same performance as GPU, and another states FPGA is 6.5 times more energy efficient than GPU.
- Cited accelerator results include CoNNA speeding up at least 3.08 times and up to 14.1 times faster than the MIT Eyeriss accelerator; an Arria 10 FPGA inference implementation being 6 times faster than GPU and 44 times faster than CPU; and a YOLO V2 Arria 10 design achieving 3.3 times more power efficiency than Titan X GPU and 418 times more power efficiency than CPU.
- Application-specific results include facial expression recognition achieving about 60.3% recognition efficiency at 200 MHz for 400 faces/second; a smart chair sitting posture ANN achieving 97.78% floating-point accuracy and 97.43% fixed-point accuracy with 7.35 mW dynamic power; and a millimeter-wave radar CNN acceleration being 30.42% faster than NVIDIA GPU with a 0.27% accuracy reduction to 87.81%.
- The paper identifies major EI challenges as high computational processing, low power consumption/high energy efficiency, and scalability, plus FPGA-specific issues such as reconfiguration time, programming difficulty, multiple-FPGA communication, limited on-chip memory, lack of built-in cache, and bandwidth typically two generations older than high-end GPU.
Methods
- Design: narrative survey/review of embedded intelligence on FPGA, organized around four classification and thematic descriptors: EI techniques, EI applications, EI hardware and platforms, and EI challenges.
- Data/sample: reviewed works and references spanning EI techniques such as machine learning and neural networks, deep learning, expert systems, fuzzy intelligence, swarm intelligence, SOM and extreme learning; applications including object detection/recognition, indoor localization and surveillance monitoring; and FPGA platforms from Xilinx, Intel/Altera, and Achronix.
- Main analysis: qualitative synthesis and tabular summaries, including Table 1 for classification/thematic descriptors and Table 2 for FPGA hardware platforms; no new empirical experiments or primary dataset are reported.
Limitations
- The paper does not state limitations of the review itself.
- Evident from the text: it is a narrative survey without a stated systematic search protocol, inclusion/exclusion criteria, or quantitative meta-analysis; reported speed, power, and accuracy figures come from heterogeneous cited studies and platforms, so they are not directly comparable.
- The paper discusses FPGA challenges such as reconfiguration time, programming cost, on-chip memory limits, and bandwidth constraints, but these are framed as challenges for FPGA-based EI rather than limitations of the review.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
28 Valid · 32 Not valid · 97 Not checked · 43 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“To have high accuracy and efficiency of the LBS for indoor environments, fingerprint-based indoor localization is introduced by using scalable deep neural network (DNN) architecture implement together with denoising auto-encoder (SDNNLoc) based on FPGA [68].”
What the sources say
- Disputedcites [68] Xie, T., Jiang, H., Zhao, X., and Zhang, C. (2019). A Wi-Fi-Based wireless indoor position sensing system with multipath interference mitigation. Sensors, 19. sourceExperimental — accuracy not yet measured
The Introduction identifies the paper’s proposed system as RSSI-assisted time-difference-of-arrival (TDoA) positioning, and the Conclusions describe an improved cross-correlation method for mitigating multipath interference. The paper discusses fingerprinting and neural networks as other approaches, but it does not introduce the claimed FPGA-based SDNNLoc architecture with a denoising autoencoder. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
Every source the paper cites for this claim was checked and supports it.
In the paper
“The result of implementing RCCM and AddNet was the ability to save resources by up to 50% more than a standard 8-bit quantized network with similar accuracy [34].”
What the sources say
- Supportedcites [34] Faraone, 2019, AddNet: Deep Neural Networks Using FPGA-Optimized Multipliers, IEEE Trans. Very Large Scale Integr. Syst. source
The paper’s opening summary (p. 1) reports up to 50% resource savings over traditional 8-bit quantized networks. It also says RCCM implementations other than the one with the lowest resource requirements achieve at least similar accuracy to an 8-bit design. (LLM, conf=0.92)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Both forward and backward propagation on a homogenous FPGA system were able to improve the training by using a DarkFPGA accelerator with batch level parallelism.”
What the sources say
- Supportedcites [32] Luo, 2020, Towards efficient deep neural network training by FPGA-based batch-level parallelism, J. Semicond. source
The Conclusion states that DarkFPGA uses batch-level parallelism to accelerate both forward and backward propagation on a homogeneous FPGA system. Section 8.2 also reports improved training performance from performing the entire process on the FPGA. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This approach gives better accuracy than the network that constrain weight parameters to have binary value.”
What the sources say
- Partly supportedcites [34] Faraone, 2019, AddNet: Deep Neural Networks Using FPGA-Optimized Multipliers, IEEE Trans. Very Large Scale Integr. Syst. sourceExperimental — accuracy not yet measured
The Accuracy Versus Area section (p. 12) reports that AddNet improves accuracy over 4-bit and 6-bit fixed-point implementations. It does not report an accuracy comparison with a network whose weights are constrained to binary values, so the specific comparator in the claim is unsupported. (LLM, conf=0.82)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“No overfitting phenomenon is observed for the model, as determined by the dropout that followed the layers in the neural network.”
What the sources say
- Supportedcites [59] 2019, FPGA platform applied for facial expression recognition system using convolutional neural networks, Procedia Comput. Sci. source
Section 4.1 (Testing) explicitly says the model did not exhibit overfitting and attributes this to dropout following the neural-network layers. The claim closely paraphrases that passage. (LLM, conf=0.95)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
51 more claims · 4 with their cited sources checked
Create a free account to see all 200 claims and to verify the 18 references still open.
- Probably not a substantial claim. 91
- It may state two separate claims as one. 81
- The quote only partly states this claim. 1
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Compared to 32-bit floating-point, a framework of 16-bit fixed-point multiplication reduces the power-consuming by 6.2 times, but with a reduction in accuracy of 0.5%.”
From Face Recognition Based on Deep Learning and FPGA for Ethnicity Identification (2022) · cites this paper as [43]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“As compared to the standard network, a reduction in parameters by 11.2 times was achieved using this fire module technique with no impact on the accuracy of classification [43].”
From Face Recognition Based on Deep Learning and FPGA for Ethnicity Identification (2022) · cites this paper as [43]
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 6. Challenges
“For example, a GPU Titan X uses GDDR5 but Arria 10 only uses DDR3.”
§ 5. FPGA Hardware and Platforms
“| Arria 10 | GT1150 | Intel | 1150 | 1518 | DDR4 |”
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 1×
“In the paper, it is shown that by running at the clock frequency of 125MHz it was able to become seven times more energy efficient than the LSTM reference for similar performance [81].”
BackgroundFaster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Ren, S., He, K., Girshick, R., & Sun, J. (2017) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2577031 · Cited by 73,472 · cited 1×
“The algorithm applies anchor boxes to change object detection with different aspect ratios, and go through numerous pooling layers to acquire the classifier image [64].”
MethodMobileNetV2: Inverted Residuals and Linear Bottlenecks
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00474 · Cited by 28,755 · cited 2×
“The VGG-16 model per-trained is applied on MobileNet architecture to extract the feature maps.”
MethodParticle swarm optimization: An overview
Poli, R., Kennedy, J., & Blackwell, T. (2007) · Swarm Intelligence · doi:10.1007/s11721-007-0002-0 · Cited by 21,101
Xception: Deep Learning with Depthwise Separable Convolutions
Chollet, F. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.195 · Cited by 20,380 · cited 1×
“In the research [24], a depthwise separable convolution neural network on FPGA is being reviewed due to its ability to reduce the number of model parameters and improve calculation speeds, which makes it widely used in mobile edge computing applications such as Xception [25] and MobileNet [26].”
MethodEyeriss: 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 2×
“Based on the analysis of the research, CoNNA is able to speed up at least 3.08 times and up to 14.1 times faster than the MIT Eyeriss Accelerator [20].”
BackgroundInfluentialAnt colony optimization theory: A survey
Dorigo, M., & Blum, C. (2005) · Theoretical Computer Science · doi:10.1016/j.tcs.2005.05.020 · Cited by 2,295 · cited 1×
“For the ACO method, a graph is used to represent the solution space [49].”
MethodMobileNetV2: Inverted Residuals and Linear BottlenecksPreprint
Sandler, M., Howard, A., Zhu, M., et al. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1801.04381 · Cited by 2,255 · cited 2×
“The VGG-16 model per-trained is applied on MobileNet architecture to extract the feature maps.”
MethodThiNet: A Filter Level Pruning Method for Deep Neural Network Compression
Luo, J.-H., Wu, J., & Lin, W. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.541 · Cited by 1,832 · cited 1×
“The coarse-grained reconfigurable architecture is applied in CoNNA, which allows immediate selection of the CNN network [19].”
MethodThe Horus WLAN location determination system
Youssef, M., & Agrawala, A. (2005) · Proceedings of the 3rd international conference on Mobile systems, applications, and services · doi:10.1145/1067170.1067193 · Cited by 1,653 · cited 2×
“The performance evaluation is focused on the runtime of the SDNNLoc and compared with two other baseline systems, which are Horus [71] and DeepFi [72].”
Method
Research works printed but not linked: 3
- Duan, 2006, Fuzzy-PID control of stepping motor, Comput. Simul., 2, 290
- Eberchart, J.K.R. (December, January 27). Particle Swarm Optimization. Proceedings of the IEEE International Conference on Neural Networks, Perth, Australia.
- Xilinx (2019, April 24). Xilinx Reports Record Revenues Exceeding $3 Billion For Fiscal. Available online: https://www.xilinx.com/news/press/2019/xilinx-reports-record-revenues-exceeding-3-billion-for-fiscal-2019.html.
Cited by
Lightweight Deep Learning for Resource-Constrained Environments: A Survey
Liu, H.-I., Galindo, M., Xie, H., Wong, L.-K., Shuai, H.-H., Li, Y.-H., & Cheng, W.-H. (2024) · ACM Computing Surveys · doi:10.1145/3657282 · Cited by 259
“For more references on FPGA AI applications, consult References [148, 170].”
BackgroundUsing acoustic indices in ecology: Guidance on study design, analyses and interpretation
Bradfer‐Lawrence, T., Desjonqueres, C., Eldridge, A., Johnston, A., & Metcalf, O. (2023) · Methods in Ecology and Evolution · doi:10.1111/2041-210x.14194 · Cited by 106
Machine Learning for the Control and Monitoring of Electric Machine Drives: Advances and Trends
Zhang, S., Wallscheid, O., & Porrmann, M. (2023) · IEEE Open Journal of Industry Applications · doi:10.1109/ojia.2023.3284717 · Cited by 122
“of the weights is reduced, resulting in reduced model size and faster computation at the cost of minimal reductions in predictive accuracy [79].”
BackgroundA review of AI edge devices and lightweight CNN and LLM deployment
Sun, K., Wang, X., Miao, X., & Zhao, Q. (2024) · Neurocomputing · doi:10.1016/j.neucom.2024.128791 · Cited by 103
Disclosing Edge Intelligence: A Systematic Meta-Survey
Barbuto, V., Savaglio, C., Chen, M., & Fortino, G. (2023) · Big Data and Cognitive Computing · doi:10.3390/bdcc7010044 · Cited by 80
“However, they refer to specific domains (e.g., robotics [18] or cyber-security [19]) or technology (e.g., embedded intelligence for FPGA [20]), to narrow contributions (e.g., a novel version of an algorithm or an optimized model [21,22]), and to broader surveys/review papers outlining definitions,…”
Metadata
- Authors
- Seng, K. P., Lee, P. J., & Ang, L. M.
- Affiliations
- UNSW Sydney (AU)
- University of the Sunshine Coast (AU)
- Corresponding author
- Seng, Kah Phooi
- Venue
- Electronics
- Publisher
- MDPI AG
- Published
- 2021-04-08
- Volume
- 10
- Issue
- 8
- Pages
- 895
- ISSN
- 2079-9292 (ISSN-L)
- OpenAlex
- W3148609646
- Cited by
- 105
- References
- 93
- Citations by year
- 2026: 14 · 2025: 18 · 2024: 25 · 2023: 19 · 2022: 21 · 2021: 8
- FWCI
- 11.97
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Electronics (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Hardware and Architecture › Embedded Systems Design Techniques
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Embedded Systems and FPGA Applications
- Physical Sciences › Engineering › Control and Systems Engineering › Embedded Systems and FPGA Design
- Keywords
- embedded intelligence, FPGA, low power consumption, energy efficiency, scalability, artificial neural networks, deep learning, expert systems, swarm intelligence, self-organizing maps, object detection and recognition, indoor positioning, surveillance monitoring, embedded devices
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 2026-10-06, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-06 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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.