Matteo Galetta
Physics-Aware Compact Modeling of Analog Conductive-Metal-Oxide/HfOx ReRAM Device.
Rel. Carlo Ricciardi, Valeria Bragaglia. Politecnico di Torino, Corso di laurea magistrale in Nanotechnologies For Icts (Nanotecnologie Per Le Ict), 2024
|
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (15MB) | Preview |
Abstract: |
Over the last decade, standard computing architectures based on von Neumann paradigm struggled to manage Internet of Things and Artificial Intelligence (AI) workloads. The inherently inefficient data transfer between processing and storage units is not suited to deal with modern data-centric applications, which are growing in complexity and scale. To tackle the exponentially increasing power demand of Neural Networks computing tasks, new low-power hardware implementations are necessary. In-Memory computing has the potential to fulfill the energy requirements in the modern Information Technology field, enabling parallel data processing and reducing latency. Memristive devices within cross-point architectures turned out to be a promising solution to perform analog In-Memory Computing, allowing to map multi-level weights between Neural Networks layers. Especially for training in neuromorphic hardware, Resistive Random Access Memory (ReRAM) devices attracted significant attention, offering fast and low-power switching capabilities, high scalability and non-volatile data storage. The integration of ReRAMs in neuromorphic systems requires extensive optimization of several aspects of the technology, ranging from device materials and fabrication processes to physical/electrical features improvements. For these purposes, physical modeling becomes crucial to provide a detailed understanding and accurate predictivity of device performances. At IBM Research Europe, the Neuromorphic Devices and Systems group is conducting R&D projects concerning the development and the optimization of an innovative ReRAM technology, integrated in system-level crossbar arrays for AI accelerators. A robust compact model able to accurately capture the operation of ReRAM devices is essential to accelerate circuit-level simulations of memory arrays and neuromorphic hardware. Hereby this dissertation presents the development and the validation of a compact model for analog filamentary Conductive-Metal-Oxide/HfOx ReRAM IBM technology. The model integrates a physics-based approach to describe analytically the ion migration mechanisms causing restistive switching phenomena. Further analysis concerns the switching dynamics of the device, evaluating the time scales in which resistive switching occurs. The model validation is conducted against experimental data of electrical characterizations, demonstrating the model's accuracy, robustness and the capability to capture the analog behavior of the device. The model is designed to be computationally efficient, highlighting its potential contribution in simulations of next-generation computing architectures. |
---|---|
Relators: | Carlo Ricciardi, Valeria Bragaglia |
Academic year: | 2023/24 |
Publication type: | Electronic |
Number of Pages: | 83 |
Subjects: | |
Corso di laurea: | Corso di laurea magistrale in Nanotechnologies For Icts (Nanotecnologie Per Le Ict) |
Classe di laurea: | New organization > Master science > LM-29 - ELECTRONIC ENGINEERING |
Ente in cotutela: | IBM Research - Zurich (SVIZZERA) |
Aziende collaboratrici: | IBM Research-Zurich |
URI: | http://webthesis.biblio.polito.it/id/eprint/31785 |
Modify record (reserved for operators) |