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

The DOI matches its Crossref record; 20 of the 20 references it cites were found in a registry or an index.

Deployment of Low-Bit Quantized Neural Networks on FPGA-Based Edge Devices Using FINN Framework

Open access
Akkuş, Batuhan · 2026 · SSRN Electronic Journal

Abstract

The increasing demand for deploying machine learn-ing (ML) algorithms on edge devices has driven the need for hardware-efficient and low-power solutions. Although GPU-based systems offer high accuracy and fast inference, they are impractical for edge deployment due to their high power con-sumption, hardware requirements, large footprint, and thermal constraints. In contrast, field programmable gate arrays (FPGAs) provide reconfigurable hardware architectures, enabling energy-efficient and resource-aware deployment tailored to specific appli-cation needs. This paper investigates the deployment of quantized neural networks (QNNs) on FPGAs using the FINN framework and Brevitas library. A custom convolutional neural network (CNN) has been trained on the German Traffic Sign Recognition Benchmark (GTSRB) dataset and quantized to different bit-width levels using the quantization-aware training (QAT) method. The models’ inference accuracy, power consumption, hardware utilization, and thermal performance were evaluated. The binary version (W1A1) of the CNV light model achieved an accuracy of 95.12% with a power consuming of 1.64 W. This study demonstrates that FPGAs, combined with quantized neural networks, enable scalable and energy-efficient ML inference on edge devices.

Overview

Integrity signals

Registry

Crossref · 10.2139/ssrn.6906368This paper is marked retracted.

Cites

20 of 20 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Verification

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Citations

Cites

Show all 10 references

Research works printed but not linked: 11

  1. Tensil, 2022, Tensil AI Documentation Version
  2. R L Ria Cheruvu, 2024, Intel solution white paper: Optimizing large language models with the openvino� toolkit
  3. R C Elisabeth Pongratz, 2019, Performance evaluation of mathworks hdl coder as a vendor independent dfe generation
  4. Y Umuroglu, 2017, Finn github repository
  5. M Lechner, 2019, Resconn: Resourceefficient fpga-accelerated cnn for traffic sign classification, 2019 Tenth International Green and Sustainable Computing Conference (IGSC), 1
  6. M Tas�c�?, 2022, Nicelenmis� derin O � g?renme ag?lar? ?Ic�in fpga tabanl? h?zland?r?c? tasar?m?, Bal?kesir U � niversitesi Fen Bilimleri En-stitu�su� Elektrik-elektronik Mu�hendislig?i Anabilim Dal?
  7. G D S Korol, 2024, Optimizing for adaptive cnns on fpgas: A multi-level approach
  8. Fpga Intel, 2023, SoC Design Example User Guide ID: 768979
  9. A Pappalardo, 2019, Xilinx/brevitas: Release v0
  10. Amd, 2018, Zynq-7000 SoC Data Sheet: Overview DS190 (v1.11.1)
  11. 2021, Vivado Design Suite User Guide Power Analysis and Optimization (UG907)

Metadata

Authors
Akkuş, B.
Corresponding author
Akkuş, Batuhan
Venue
SSRN Electronic Journal
Publisher
Elsevier BV
Published
2026-06-09
ISSN
1556-5068 (ISSN-L)
OpenAlex
W7164036175
Cited by
0
References
10
Type
posted content
Language
en
Status
Retracted
Open access
Locations
SSRN Electronic Journal (publishedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Hardware and Architecture › Embedded Systems Design Techniques
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Embedded Systems and FPGA Applications
Keywords
low-bit quantization, neural network quantization, FPGA, edge computing, brevitas, quantization-aware training, convolutional neural networks, energy-efficient inference, power consumption, thermal performance, hardware resource utilization, traffic sign recognition

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.