Aristotle University Of Thessaloniki (AUTH)
Cross-layer approximation techniques for synthesis and optimization of machine learning accelerators
Abstract
dc:descriptionMachine learning has emerged as the pinnacle of modern technological progress, representing the forefront of state-of-the-art advancements, with its profound capacity to extract useful insights from real-world data, spanning various application domains. However, hardware deployment of ml algorithms on commodity hardware has become a challenging task, due to the increased demands for accuracy and performance. Domain-specific multiply and accumulate (mac)-based ml accelerators emerge as critical components, essential for alleviating the computational complexity of large ml models via parallelization of comprising mac operations. Still, we identify two main challenges with current ml deployment on accelerators. First, underlying circuitry suffers from runtime degradation phenomena, diminishing the reliability and potential performance of the application. Secondly, the massive parallelism causes elevated energy consumption, prohibiting the integration of ml algorithms on resource-constrained embedded systems. In this dissertation, we combat these challenges by utilizing cross-layer approximate computing (axc) principles. We propose automated approximation methodologies such that the energy efficiency of ml accelerators is improved, whilst their reliable operation is guaranteed. Our methodologies comply with given accuracy bounds, thus ensuring application quality, whilst targeting efficiency metrics via optimization methodologies. Specifically, we design aging- and variability-hardened arithmetic circuits for mac-based accelerators of improved performance, compared to the state-of-the-art. On the algorithmic level, we integrate compression methodologies in ml training and inference, such that the memory requirements and energy consumption of the deployed workloads are optimized. Overall, this dissertation promotes the automated application of approximate computing on several layers of ml accelerator design, achieving practical improvements on several important fronts, via coordinated and automated methodologies.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Balaskas, Konstantinos
- Μπαλάσκας, Κωνσταντίνος
Subjects
dc:subject × 26- Μηχανική μάθηση
- Προσεγγιστικός υπολογισμός
- Αυτοματοποιημένη βελτιστοποίηση
- Σύνθεση κυκλωμάτων
- Γήρανση κυκλωμάτων
- Κατασκευαστικές παρεκλίσεις
- Βαθιά νευρωνικά δίκτυα
- Αραίωση νευρωνικών δικτύων
- Κβάντωση νευρωνικών δικτύων
- Επιταχυντές υλικού
- Machine learning
- Approximate computing
- Deep neural networks
- Design automation
- Hardware accelerators
- Aging
- Process variations
- Pruning
- Quantization
- Circuit synthesis
- 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/57080