Ioannis Kouretas
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Ioannis Kouretas has a total of 38 co-authors. Here are the top co-authors by number of shared publications:
- Vassilis Paliouras
Electrical and Computer Engineering Department, University of Patras, Patras, Greece
26 shared publications - Konstantinos Peppas
Department of Informatics and Communications, University of Peloponnese, 22100 Tripoli, Greece
6 shared publications - Anastasios G. Skrivanos
Department of Informatics and Communications, University of Peloponnese, 22131 Tripoli, Greece
4 shared publications Mark G. Arnold
XLNS Research, United States
4 shared publicationsV. Paliouras
Electrical and Computer Engineering Dept., University of Patras, Patras, 26 504, Greece
4 shared publicationsAnastasios Skrivanos
University of Peloponnese, Greece
3 shared publications
Latest Publications
- Open Access
- 17 August 2026
- Published by MDPI AG in Electronics
- Vol. 15 (16) , 3663
Hardware Aspects of Machine Learning-Based Cardiac Fibrillation Diagnosis
This paper presents the hardware aspects of a Hjorth-parameter deep neural network (DNN)-based cardiac fibrillation diagnosis pipeline targeting low-power Internet of Medical Things (IoMT) devices and ASIC implementations. Building on earlier edge-to-cloud studies where Hjorth activity, mobility, and complexity were computed in software on microcontrollers and classified by a floating-point DNN, we introduce a fully synthesizable fixed-point hardware module that computes these parameters in real time, together with a quantized neural network (QNN) operating directly on the resulting fixed-point features. The Hjorth block includes derivative generation, accumulator banks, variance computation, and hardware divider and square-root units. Using the Shandong Provincial Hospital Database (SPHD) as in our previous work, we evaluate the impact of end-to-end fixed-point quantization on the AF-related arrhythmia detection performance for bit widths between 6 and 16 bits. For 10–12-bit configurations, the quantized pipeline achieves accuracy of approximately 93.7% and an area under the ROC curve (AUC) above 0.97, closely matching the floating-point baseline while significantly reducing the arithmetic complexity and memory footprint. ASIC synthesis in a 28 nm CMOS standard-cell library shows that the complete atrial detector core, integrating the Hjorth extractor and the hardware fully connected QNN, occupies on the order of and dissipates about 3 mW, with a critical-path delay of 7.08 ns. For the optimal 10–12-bit operating points, the synthesized core achieves per-inference energy in the range of 18.6–19.0 nJ (18,600–19,000 pJ) per classification, confirming its suitability for integration into wearable and IoMT ECG monitoring nodes. These results demonstrate that co-designed fixed-point Hjorth hardware and quantized DNNs can deliver a favorable trade-off between diagnostic performance, area, power, latency, and per-inference energy compared with existing MCU-, FPGA-, and ASIC-based ECG classifiers.- 17 August 2026
- Published by MDPI AG in Electronics
- Vol. 15 (16) , 3663
Open AccessHardware Aspects of Machine Learning-Based Cardiac Fibrillation Diagnosis
This paper presents the hardware aspects of a Hjorth-parameter deep neural network (DNN)-based cardiac fibrillation diagnosis pipeline targeting low-power Internet of Medical Things (IoMT) devices and ASIC implementations. Building on earlier edge-to-cloud studies where Hjorth activity, mobility, and complexity were computed in software on microcontrollers and classified by a floating-point DNN, we introduce a fully synthesizable fixed-point hardware module that computes these parameters in real time, together with a quantized neural network (QNN) operating directly on the resulting fixed-point features. The Hjorth block includes derivative generation, accumulator banks, variance computation, and hardware divider and square-root units. Using the Shandong Provincial Hospital Database (SPHD) as in our previous work, we evaluate the impact of end-to-end fixed-point quantization on the AF-related arrhythmia detection performance for bit widths between 6 and 16 bits. For 10–12-bit configurations, the quantized pipeline achieves accuracy of approximately 93.7% and an area under the ROC curve (AUC) above 0.97, closely matching the floating-point baseline while significantly reducing the arithmetic complexity and memory footprint. ASIC synthesis in a 28 nm CMOS standard-cell library shows that the complete atrial detector core, integrating the Hjorth extractor and the hardware fully connected QNN, occupies on the order of and dissipates about 3 mW, with a critical-path delay of 7.08 ns. For the optimal 10–12-bit operating points, the synthesized core achieves per-inference energy in the range of 18.6–19.0 nJ (18,600–19,000 pJ) per classification, confirming its suitability for integration into wearable and IoMT ECG monitoring nodes. These results demonstrate that co-designed fixed-point Hjorth hardware and quantized DNNs can deliver a favorable trade-off between diagnostic performance, area, power, latency, and per-inference energy compared with existing MCU-, FPGA-, and ASIC-based ECG classifiers. - Open Access
- 10 July 2026
- Published by MDPI AG in Preprints.org
Hardware Aspects of Machine Learning-Based Cardiac Fibrillation Diagnosis
This paper presents the hardware aspects of a Hjorth-parameter and deep neural network 2 (DNN)–based cardiac fibrillation diagnosis pipeline targeting low-power Internet of Medi- 3 cal Things (IoMT) devices and ASIC implementations. Building on earlier edge-to-cloud 4 studies where Hjorth activity, mobility, and complexity were computed in software on 5 microcontrollers and classified by a floating-point DNN, we introduce a fully synthesizable 6 fixed-point hardware module that computes these parameters in real time, together with a 7 quantized neural network (QNN) operating directly on the resulting fixed-point features. 8 The Hjorth block includes derivative generation, accumulator banks, variance computa- 9 tion, and hardware divider and square-root units. Using the Shandong Provincial Hospital 10 Database (SPHD) and the same 1-minute ECG windowing protocol as in our previous work, 11 we evaluate the impact of end-to-end fixed-point quantization on AF-related arrhythmia 12 detection performance for bit-widths between 6 and 16 bits. For 10–12-bit configurations, 13 the quantized pipeline achieves accuracy of approximately 93.7% and area under the ROC 14 curve (AUC) above 0.97, closely matching the floating-point baseline while significantly 15 reducing arithmetic complexity and memory footprint. ASIC synthesis in a 28 nm CMOS 16 standard-cell library shows that the complete atrial_detector core, integrating the Hjorth 17 extractor and the hardware fully connected QNN, occupies on the order of 2 × 104 µm2 and 18 dissipates about 3 mW, with a critical-path delay of 7.08 ns, confirming its suitability for 19 integration into wearable and IoMT ECG monitoring nodes. These results demonstrate that 20 co-designed fixed-point Hjorth hardware and quantized DNNs can deliver a favourable 21 trade-off between diagnostic performance, area, power, and latency compared with existing 22 MCU-, FPGA-, and ASIC-based ECG classifiers.
- 10 July 2026
- Published by MDPI AG in Preprints.org
Open AccessHardware Aspects of Machine Learning-Based Cardiac Fibrillation Diagnosis
This paper presents the hardware aspects of a Hjorth-parameter and deep neural network 2 (DNN)–based cardiac fibrillation diagnosis pipeline targeting low-power Internet of Medi- 3 cal Things (IoMT) devices and ASIC implementations. Building on earlier edge-to-cloud 4 studies where Hjorth activity, mobility, and complexity were computed in software on 5 microcontrollers and classified by a floating-point DNN, we introduce a fully synthesizable 6 fixed-point hardware module that computes these parameters in real time, together with a 7 quantized neural network (QNN) operating directly on the resulting fixed-point features. 8 The Hjorth block includes derivative generation, accumulator banks, variance computa- 9 tion, and hardware divider and square-root units. Using the Shandong Provincial Hospital 10 Database (SPHD) and the same 1-minute ECG windowing protocol as in our previous work, 11 we evaluate the impact of end-to-end fixed-point quantization on AF-related arrhythmia 12 detection performance for bit-widths between 6 and 16 bits. For 10–12-bit configurations, 13 the quantized pipeline achieves accuracy of approximately 93.7% and area under the ROC 14 curve (AUC) above 0.97, closely matching the floating-point baseline while significantly 15 reducing arithmetic complexity and memory footprint. ASIC synthesis in a 28 nm CMOS 16 standard-cell library shows that the complete atrial_detector core, integrating the Hjorth 17 extractor and the hardware fully connected QNN, occupies on the order of 2 × 104 µm2 and 18 dissipates about 3 mW, with a critical-path delay of 7.08 ns, confirming its suitability for 19 integration into wearable and IoMT ECG monitoring nodes. These results demonstrate that 20 co-designed fixed-point Hjorth hardware and quantized DNNs can deliver a favourable 21 trade-off between diagnostic performance, area, power, and latency compared with existing 22 MCU-, FPGA-, and ASIC-based ECG classifiers.
- Open Access
- 13 May 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (10) , 4839
An IoT-Based Real-Time Monitoring and Alert System for Sea Turtle Nest Protection
This paper presents a low-cost Internet-of-Things (IoT) telemetry and alerting system for monitoring and protecting sea turtle nests. The proposed platform integrates temperature, humidity, vibration, ultrasonic proximity, and ambient light sensors into an autonomous sensing node based on the ESP8266 microcontroller. Measurements are transmitted wirelessly to a cloud backend for real-time visualization and rule-based alert generation. The system is designed to support continuous nest-level monitoring and rapid response to environmental and anthropogenic threats such as overheating, artificial light exposure during hatching, and physical disturbance. In contrast to approaches that require extensive historical datasets or machine-learning models, the proposed solution employs transparent threshold-based rules that provide reliable operation without training data. The platform emphasizes low cost, ease of deployment, and scalability, making it suitable for large-scale conservation deployments across multiple nesting sites. It provides conservation practitioners with actionable situational awareness that complements existing field-based monitoring and protection strategies.
- 13 May 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (10) , 4839
Open AccessAn IoT-Based Real-Time Monitoring and Alert System for Sea Turtle Nest Protection
This paper presents a low-cost Internet-of-Things (IoT) telemetry and alerting system for monitoring and protecting sea turtle nests. The proposed platform integrates temperature, humidity, vibration, ultrasonic proximity, and ambient light sensors into an autonomous sensing node based on the ESP8266 microcontroller. Measurements are transmitted wirelessly to a cloud backend for real-time visualization and rule-based alert generation. The system is designed to support continuous nest-level monitoring and rapid response to environmental and anthropogenic threats such as overheating, artificial light exposure during hatching, and physical disturbance. In contrast to approaches that require extensive historical datasets or machine-learning models, the proposed solution employs transparent threshold-based rules that provide reliable operation without training data. The platform emphasizes low cost, ease of deployment, and scalability, making it suitable for large-scale conservation deployments across multiple nesting sites. It provides conservation practitioners with actionable situational awareness that complements existing field-based monitoring and protection strategies.
- 8 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 14 (2) , 604-618
A 22-nm 4.92 TOPS/W End-to-End RNS DNN Accelerator for Edge-AI Devices
This work presents an end-to-end Residue Numbering System (RNS) Deep Neural Network (DNN) accelerator targeting edge-AI devices. The proposed architecture translates the intrinsic advantages of RNS-particularly the efficient implementation of multiply-add operations-into system-level power efficiency gains. This is made possible by the introduction of novel architectural features, such as the amortization of high-complexity RNS operations (base extension, activation, and scaling), and the integration of bespoke RNS low-power techniques, such as a clock-gating scheme that exploits the periodic usage of the high-complexity RNS units, and voltage scaling that exploits the RNS capability to achieve clock frequency constraints with smaller supply voltages. Tradeoffs between hardware performance metrics of RNS implementations for various operational scenarios and RNS bases have been systematically analyzed, and comparisons against a conventional positional binary implementation have been conducted, guiding the optimal selection of design space parameters. Silicon power measurements on the 22-nm FDSOI prototype chips underpin the theoretical analysis and simulation results, which show considerable benefits of RNS-based DNN processing. These prove that RNS can increase the maximum achievable frequency of arithmetic circuits and result in 1.34× more energy-efficient processing compared to conventional binary counterparts. Compared to the state-of-the-art RNS-based DNN accelerator, the proposed architecture is 9× more power efficient, reaching a peak power efficiency of 4.92 TOPS/W.
- 8 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 14 (2) , 604-618
A 22-nm 4.92 TOPS/W End-to-End RNS DNN Accelerator for Edge-AI Devices
This work presents an end-to-end Residue Numbering System (RNS) Deep Neural Network (DNN) accelerator targeting edge-AI devices. The proposed architecture translates the intrinsic advantages of RNS-particularly the efficient implementation of multiply-add operations-into system-level power efficiency gains. This is made possible by the introduction of novel architectural features, such as the amortization of high-complexity RNS operations (base extension, activation, and scaling), and the integration of bespoke RNS low-power techniques, such as a clock-gating scheme that exploits the periodic usage of the high-complexity RNS units, and voltage scaling that exploits the RNS capability to achieve clock frequency constraints with smaller supply voltages. Tradeoffs between hardware performance metrics of RNS implementations for various operational scenarios and RNS bases have been systematically analyzed, and comparisons against a conventional positional binary implementation have been conducted, guiding the optimal selection of design space parameters. Silicon power measurements on the 22-nm FDSOI prototype chips underpin the theoretical analysis and simulation results, which show considerable benefits of RNS-based DNN processing. These prove that RNS can increase the maximum achievable frequency of arithmetic circuits and result in 1.34× more energy-efficient processing compared to conventional binary counterparts. Compared to the state-of-the-art RNS-based DNN accelerator, the proposed architecture is 9× more power efficient, reaching a peak power efficiency of 4.92 TOPS/W.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 156544-156558
Efficient Majority-Logic Design of Modulo 2 n – 1 Parallel-Prefix Adders
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 156544-156558
Open AccessEfficient Majority-Logic Design of Modulo 2 n – 1 Parallel-Prefix Adders
- Open Access
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 205753-205763
Efficient Majority Logic Parallel-Prefix Adder Design
Emerging beyond CMOS technologies such as quantum-dot cellular automata (QCA), resistive RAM crossbars, spin-torque devices, and single-electron transistors implement only the three-input majority gate (MG) and the inverter as primitive operations. Therefore, efficient arithmetic designs are required for adder architectures implemented entirely in MG logic. This article introduces a methodology that maps any conventional parallel-prefix adder directly to an MG network, reformulating the sum computation so that the parallel-prefix tree is simplified. Based on the proposed technique, two topologies are introduced namely the Simplified Kogge–Stone (SKS) and Simplified Ladner–Fischer (SLF) adders. Both topologies utilise novel MG-based cells, and retain a critical-path depth of log2 n+3 MG delays, identical to their full-prefix counterparts. Analytical evaluation for 8- to 128-bit adders demonstrates that SKS reduces total MG count by 22.7%–38.5% compared with the best majority-logic Kogge–Stone counterparts, while SLF exhibits 5.9%–25.1% savings over majority-logic Ladner–Fischer designs, without increasing inverter overhead. The proposed methodology derives lower-complexity, high-throughput adders that are directly mapped on majority-gate nanotechnologies and offers a scalable procedure toward arithmetic building blocks for beyond-CMOS systems.
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 205753-205763
Open AccessEfficient Majority Logic Parallel-Prefix Adder Design
Emerging beyond CMOS technologies such as quantum-dot cellular automata (QCA), resistive RAM crossbars, spin-torque devices, and single-electron transistors implement only the three-input majority gate (MG) and the inverter as primitive operations. Therefore, efficient arithmetic designs are required for adder architectures implemented entirely in MG logic. This article introduces a methodology that maps any conventional parallel-prefix adder directly to an MG network, reformulating the sum computation so that the parallel-prefix tree is simplified. Based on the proposed technique, two topologies are introduced namely the Simplified Kogge–Stone (SKS) and Simplified Ladner–Fischer (SLF) adders. Both topologies utilise novel MG-based cells, and retain a critical-path depth of log2 n+3 MG delays, identical to their full-prefix counterparts. Analytical evaluation for 8- to 128-bit adders demonstrates that SKS reduces total MG count by 22.7%–38.5% compared with the best majority-logic Kogge–Stone counterparts, while SLF exhibits 5.9%–25.1% savings over majority-logic Ladner–Fischer designs, without increasing inverter overhead. The proposed methodology derives lower-complexity, high-throughput adders that are directly mapped on majority-gate nanotechnologies and offers a scalable procedure toward arithmetic building blocks for beyond-CMOS systems.
- Open Access
- 11 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 196374-196387
A Machine Learning-Based Heart Disease Detection Scheme for Predicting Spontaneous Termination of Atrial Fibrillation
This study introduces an electrocardiogram (ECG) detection system for predicting the spontaneous termination of atrial fibrillation (AF). The proposed system uses a two-stage approach: first, it extracts features from the ECG signals using Hjorth parameters (activity, mobility, and complexity), and second, it applies a deep neural network (DNN) classifier to analyze these features. The system was validated using clinical data obtained from the Shandong Provincial Hospital Database (SPHD), which includes recordings of various types of AF episodes, such as non-terminating, immediately-terminating, and soon-terminating AF. Experimental results indicate that the proposed system achieves an accuracy of approximately 94%, with precision and F-measure scores of 0.95, demonstrating robust performance in identifying AF termination events. To further assess the model’s effectiveness, we performed confusion matrix and receiver operating characteristic (ROC) analysis. These results confirm the potential of the proposed system for real-time deployment in portable consumer electronic (CE) devices, maintaining low implementation complexity while offering reliable AF detection.
- 11 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 196374-196387
Open AccessA Machine Learning-Based Heart Disease Detection Scheme for Predicting Spontaneous Termination of Atrial Fibrillation
This study introduces an electrocardiogram (ECG) detection system for predicting the spontaneous termination of atrial fibrillation (AF). The proposed system uses a two-stage approach: first, it extracts features from the ECG signals using Hjorth parameters (activity, mobility, and complexity), and second, it applies a deep neural network (DNN) classifier to analyze these features. The system was validated using clinical data obtained from the Shandong Provincial Hospital Database (SPHD), which includes recordings of various types of AF episodes, such as non-terminating, immediately-terminating, and soon-terminating AF. Experimental results indicate that the proposed system achieves an accuracy of approximately 94%, with precision and F-measure scores of 0.95, demonstrating robust performance in identifying AF termination events. To further assess the model’s effectiveness, we performed confusion matrix and receiver operating characteristic (ROC) analysis. These results confirm the potential of the proposed system for real-time deployment in portable consumer electronic (CE) devices, maintaining low implementation complexity while offering reliable AF detection.
- Open Access
- 6 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
A 22-nm 4.92 TOPS/W end-to-end RNS DNN Accelerator for Edge-AI Devices
This work presents an end-to-end Residue Numbering System (RNS) Deep Neural Network (DNN) accelerator targeting edge-AI devices. The developed architecture enables translating the advantages of RNS regarding the implementation of a single multiply-add operation into system-level power efficiency gains. This is made possible by the introduction of novel architectural features, such as the amortization of nontrivial RNS operations (base extension, activation and scaling), and the integration of bespoke RNS low-power techniques, such as a clock-gating scheme that exploits the periodic usage of the non-trivial RNS units, and voltage scaling that exploits RNS capability to achieve clock frequency constraints with smaller supply voltages. Systematic analysis of trade-offs between hardware performance metrics (area, power, throughput) of the RNS implementation for various operation scenarios and RNS bases, as well as comparisons against the conventional positional binary implementation, are conducted, guiding the optimal selection of design space parameters. Silicon power measurements on the 22-nm FDSOI prototype chips underpin the theoretical analysis and simulation results, which show considerable benefits of RNSbased DNN processing. These prove that RNS can not only increase the maximum achievable frequency of the arithmetic circuits, but also results in 1.33× more energy-efficient processing, compared to conventional binary counterparts. Compared to the state-of-the-art RNS-based DNN accelerator, the proposed architecture is shown to be 9× more power efficient, reaching a peak power efficiency of 4.92 TOPS/W.
- 6 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessA 22-nm 4.92 TOPS/W end-to-end RNS DNN Accelerator for Edge-AI Devices
This work presents an end-to-end Residue Numbering System (RNS) Deep Neural Network (DNN) accelerator targeting edge-AI devices. The developed architecture enables translating the advantages of RNS regarding the implementation of a single multiply-add operation into system-level power efficiency gains. This is made possible by the introduction of novel architectural features, such as the amortization of nontrivial RNS operations (base extension, activation and scaling), and the integration of bespoke RNS low-power techniques, such as a clock-gating scheme that exploits the periodic usage of the non-trivial RNS units, and voltage scaling that exploits RNS capability to achieve clock frequency constraints with smaller supply voltages. Systematic analysis of trade-offs between hardware performance metrics (area, power, throughput) of the RNS implementation for various operation scenarios and RNS bases, as well as comparisons against the conventional positional binary implementation, are conducted, guiding the optimal selection of design space parameters. Silicon power measurements on the 22-nm FDSOI prototype chips underpin the theoretical analysis and simulation results, which show considerable benefits of RNSbased DNN processing. These prove that RNS can not only increase the maximum achievable frequency of the arithmetic circuits, but also results in 1.33× more energy-efficient processing, compared to conventional binary counterparts. Compared to the state-of-the-art RNS-based DNN accelerator, the proposed architecture is shown to be 9× more power efficient, reaching a peak power efficiency of 4.92 TOPS/W.
- Open Access
- 23 April 2024
- Published by Wiley
Efficient Majority Logic Parallel-Prefix Adder Design
Beyond CMOS nanotechnology has been attracted interest by many researchers. The logical fundamental elements of many nanotechnologies are the majority, minority gates and inverters. The design of efficient adder systems and especially the parallel prefix adders is of very importance. In this paper efficient majority logic implementation of parallel prefix is introduced. The proposed methodology can be generalized to any parallel prefix structure design. Moreover, the proposed majority logic parallel prefix adder designs demonstrate decreased circuit complexity when compared to the literature.
- 23 April 2024
- Published by Wiley
Open AccessEfficient Majority Logic Parallel-Prefix Adder Design
Beyond CMOS nanotechnology has been attracted interest by many researchers. The logical fundamental elements of many nanotechnologies are the majority, minority gates and inverters. The design of efficient adder systems and especially the parallel prefix adders is of very importance. In this paper efficient majority logic implementation of parallel prefix is introduced. The proposed methodology can be generalized to any parallel prefix structure design. Moreover, the proposed majority logic parallel prefix adder designs demonstrate decreased circuit complexity when compared to the literature.
- 10 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 12 (2) , 667-683
A Multiplier-Free RNS-Based CNN Accelerator Exploiting Bit-Level Sparsity
In this work, a Residue Numbering System (RNS)-based Convolutional Neural Network (CNN) accelerator utilizing a multiplier-free distributed-arithmetic Processing Element (PE) is proposed. A method for maximizing the utilization of the arithmetic hardware resources is presented. It leads to an increase of the system's throughput, by exploiting bit-level sparsity within the weight vectors. The proposed PE design takes advantage of the properties of RNS and Canonical Signed Digit (CSD) encoding to achieve higher energy efficiency and effective processing rate, without requiring any compression mechanism or introducing any approximation. An extensive design space exploration for various parameters (RNS base, PE micro-architecture, encoding) using analytical models as well as experimental results from CNN benchmarks is conducted and the various trade-offs are analyzed. A complete end-to-end RNS accelerator is developed based on the proposed PE. The introduced accelerator is compared to traditional binary and RNS counterparts as well as to other state-of-the-art systems. Implementation results in a 22-nm process show that the proposed PE can lead to 1.85× and 1.54× more energy-efficient processing compared to binary and conventional RNS, respectively, with a 1.88× maximum increase of effective throughput for the employed benchmarks. Compared to a state-of-the-art, all-digital, RNS-based system, the proposed accelerator is 8.87× and 1.11× more energy- and area-efficient, respectively.
- 10 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 12 (2) , 667-683
A Multiplier-Free RNS-Based CNN Accelerator Exploiting Bit-Level Sparsity
In this work, a Residue Numbering System (RNS)-based Convolutional Neural Network (CNN) accelerator utilizing a multiplier-free distributed-arithmetic Processing Element (PE) is proposed. A method for maximizing the utilization of the arithmetic hardware resources is presented. It leads to an increase of the system's throughput, by exploiting bit-level sparsity within the weight vectors. The proposed PE design takes advantage of the properties of RNS and Canonical Signed Digit (CSD) encoding to achieve higher energy efficiency and effective processing rate, without requiring any compression mechanism or introducing any approximation. An extensive design space exploration for various parameters (RNS base, PE micro-architecture, encoding) using analytical models as well as experimental results from CNN benchmarks is conducted and the various trade-offs are analyzed. A complete end-to-end RNS accelerator is developed based on the proposed PE. The introduced accelerator is compared to traditional binary and RNS counterparts as well as to other state-of-the-art systems. Implementation results in a 22-nm process show that the proposed PE can lead to 1.85× and 1.54× more energy-efficient processing compared to binary and conventional RNS, respectively, with a 1.88× maximum increase of effective throughput for the employed benchmarks. Compared to a state-of-the-art, all-digital, RNS-based system, the proposed accelerator is 8.87× and 1.11× more energy- and area-efficient, respectively.