Democritus University of Thrace (DUTH)
Automated design of machine learning hardware accelerators
Abstract
dc:descriptionThis thesis presents a comprehensive framework for the automated design of hardware accelerators targeting Convolutional Neural Networks (CNNs), leveraging High-Level Synthesis (HLS) and high-level model descriptions in Python (via TensorFlow or PyTorch). The proposed methodology streamlines the development of CNN dataflow accelerators by enabling direct hardware generation from software models, significantly reducing design complexity and development time. The framework introduces a rich set of configurable optimization parameters, allowing exploration of different architectural designs with respect to performance, energy efficiency, and resource utilization. Beyond conventional design knobs, this work integrates, for the first time, hardware-level functional safety features, advanced buffering architectures, and support for diverse data representations including reduced-precision floating-point and fixed-point formats. Two specialized streaming convolution engines are introduced: LazyDCstream, optimized for dilated convolutions, and LeapConv, optimized for strided convolutions. LazyDCstream minimizes power consumption through efficient buffer reuse and ``lazy'' data movement that remains independent of dilation rate. LeapConv addresses inefficiencies in strided convolutions by decomposing them into unity-stride channels and merging them into a single hardware unit, thereby reducing redundant memory access with minimal area overhead. To support functional safety, the thesis introduces a low-overhead Algorithm-Based Fault Tolerant (ABFT) checker tailored for CNNs. The convolution checker predicts output checksums using only border pixels of the input, leveraging a novel convolutional invariance condition. This drastically reduces buffer and logic requirements, enabling efficient online fault detection. The framework is validated through the automatic synthesis of multiple CNN accelerators from real-world models. Experimental results demonstrate significant improvements in energy efficiency, configurability, and hardware utilization compared to state-of-the-art tools. Overall, this thesis delivers a scalable and practical solution for CNN hardware acceleration, addressing critical demands in performance, energy, and reliability for modern machine learning systems.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Filippas, Dionysios
- Φίλιππας, Διονύσιος
Subjects
dc:subject × 22- Επιταχυντές υλικού
- Μηχανική μάθηση
- Συνελικτικά νευρωνικά δίκτυα
- Σύνθεση υψηλού επιπέδου
- Αρχιτεκτονικές χαμηλής ισχύος
- Ασφαλή συστήματα
- Αριθμητική κινητής υποδιαστολής
- Εργαλεία αυτοματοποιημένης σχεδίασης υλικού
- Hardware accelerators
- Machine learning
- Convolutional neural networks
- High-level synthesis
- Low-power architecture
- Fault-tolerant systems
- Floating-point arithmetic
- Electronic design automation
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59369