National Technical University of Athens (NTUA)
Υλοποίηση νευρομορφικών υπολογισμών με δικτυώματα μνημών μεταβλητής αντίστασης
Abstract
dc:descriptionIn recent decades, the global sales rate of integrated microelectronic circuits has rapidly increased, driven by the constant demand for “smart” portable electronic devices, electric and environmentally friendly cars, and embedded sensor systems capable of processing and recording data in real time. Although Moore’s Law continues to hold and the number of transistors per integrated circuit doubles every two years, technological limitations in scaling dimensions now make it imperative to search for innovative architectures and alternative computational approaches. In particular, end-user demands have focused on optimized computing system performance while reducing costs, achieving higher processing speeds, and drastically reducing energy consumption. These demands, combined with the exponentially increasing volume of data collected, transmitted, and processed daily, have led to the creation of a new technological era, known as the “Big Data Era”. This era, which emerged at the beginning of the second millennium, is characterized by massive and rapidly expanding data collections that require advanced storage, processing, and analysis methodologies. At the same time, artificial intelligence (AI) technologies have significantly contributed, facilitating the processing of the enormous volume of data collected daily.However, today's computers are based on the Von Neumann architecture, which requires the physical separation of the Central Processing Unit (CPU) and memory. This separation significantly reduces the speed of data processing and increases the power consumption. Although the use of Graphics Processing Units (GPU), which have multiple cores, can enhance the degree of parallelism when performing computational processes, however, when the GPU is used to implement neural networks, the synaptic weights are still stored in separate units, such as Static Random Access Memory (SRAM). An initial approach is to develop non-Von Neumann architectures, where the memory and computing units coexist on the same hardware substrate. Although large companies, such as IBM and Intel, have already invested in corresponding technologies, the implemented systems still lag behind biological systems in terms of energy efficiency. An attractive alternative for implementing neuromorphic systems with embedded memory properties, aiming to implement non-Von Neumann architectures and promote in-memory computing, is emerging memory technologies, known as Storage Class Memories (SCM). More specifically, Conductive Bridge Random Access Memories (CBRAM), which are part of the broader category of variable resistance memories, exhibit high write/erase speeds and high integration density. At the same time, they have good endurance in consecutive programming cycles and low power consumption. The operation of these memories, which were the subject of this research thesis, is based on the electrochemical phenomenon. More specifically, CBRAMs as two-terminal devices, consist of an electrochemically active electrode and an inert one, forming a metal - insulator - metal device. Under the application of external fields of appropriate polarity, metal ions produced by the electrochemically active electrode can be moved through the dielectric layer, towards the surface of the inert electrode. These ions form a conductive filament, connecting the two electrodes and inducing the low resistance state (LRS), while the local breakdown of the conductive channel creates the high resistance state (HRS). This thesis focuses on the study of the physical mechanisms that govern the operation of CBRAM devices while simultaneously presenting the individual steps for their development and electrical characterization. Specifically, these devices exhibit either volatile operation, which corresponds to a spontaneous return of the memory elements to the initial HRS state within a predetermined time interval (relaxation time), or non-volatile operation, during which the recording state is maintained stable without the need for an external power supply. The effect of the material of the passive electrode on the functional behavior of the devices was also systematically investigated, highlighting the decisive role of the passive electrode’s thermal conductivity coefficient in predicting and controlling the device operation. The ability of non-volatile devices to exhibit synaptic properties similar to those of the human brain was then examined, with emphasis laid on the characteristics of long-term synaptic plasticity. Also, their ability to be programmed into multiple, discrete conductance states is still under investigation. This feature is particularly important for their utilization as synaptic weights in artificial neural networks (ANNs), enabling the representation of intermediate weight values and, consequently, more accurate network training. Accordingly, volatile CBRAM devices are approached as candidate artificial neurons, offering a functional bridge between biological neural systems and Artificial Neural Networks (ANNs). In more detail, the development of crossbar structures with integrated non-volatile CBRAM memories constitutes a decisive step towards the implementation of non-Von Neumann architectures and the integration of processing and storage in the same computing system (in-memory computing). In this context, two different simulation models of CBRAM devices and by extension the corresponding Crossbar architectures were implemented, which were utilized for the modeling of various types of neural networks, such as deep neural networks (DNNs), recurrent neural networks (RNN) with reservoir computing (RC) and spiking neural networks (SNNs). Finally, a Crossbar architecture based on TiN (~25 nm) / SiO₂ (~60 nm) / Ag (~100 nm) is developed, and the corresponding peripheral circuit is designed for an application involving the recognition of the Iris dataset. The input data consist of four distinct features of petals and sepals from three flower species, achieving correct classification with a recognition accuracy of 86.66%. The complete circuit exhibits an energy requirement of approximately 48 mJ during the inference process, while the Crossbar architecture consumes about 1.15 mW. This approach demonstrates the capability of Crossbar-based architectures to perform supervised classification tasks directly at the hardware level, offering low power consumption and high integration density.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsioustas, Charalampos
- Τσιούστας, Χαράλαμπος
Subjects
dc:subject × 23- Τεχνητά νευρωνικά δίκτυα
- Νευρομορφικοί Υπολογισμοί
- Αρχιτεκτονικές αλληλοδιασταυρωμένων δομών
- Μνήμες αγώγιμης γέφυρας
- Προσομοίωση βιολογικών συνάψεων και νευρώνων
- Μεμρίστορ
- Μη-πτητικές και πτητικές μνήμες μεταβλητής αντίστασης
- Artificial neural networks
- Neuromorphic computing
- Crossbar architecture structures
- Conductive bridge RAMs
- Simulation of biological synapses and neurons
- Non-Von Neumann architecture
- In-memory computing
- Reservoir computing
- Memristor
- Non-volatile and volatile memristive memories
- Νανοτεχνολογία
- Επιστήμες Μηχανικού και Τεχνολογία
- Νανοεπιστήμη και Νανοτεχνολογία
- Nano-Technology
- Engineering and Technology
- Nanoscience and Nanotechnology
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60208