The energy consumption of information and communication technology is increasing worldwide, and the rise of Artificial Intelligence (AI) leads to massive amounts of data that need to be processed and stored. Data-centric tasks, such as training and inference in AI models, impose a heavy computational burden on traditional digital computing systems based on the Von Neumann architecture. Motivated by these trends, the investigation of next-generation non-volatile computing and memory devices in which data and logic coexist is vital. The in-memory computing concept (IMC) is a promising paradigm shift where highly specialized tasks are executed within memory arrays, improving the efficiency of the data-intensive workloads, such as deep learning on the edge, by removing the data shuttling between computational and memory elements. Resistive random access memory (RRAM) is one of the most promising emerging memories, due to a broad range of attractive properties for IMC, such as non-volatile analog storage, low power consumption, and high memory density. 3D-Vertical RRAM (3D-VRRAM) is a promising option to achieve high memory cell capacity with low fabrication cost, toward large-scale integration competitive with state-of-the-art 3D NAND flash. This Doctoral dissertation presents the fabrication and characterization of planar 2D and 3D-VRRAM crossbar arrays (CBA) based on HfOx cells and capable of IMC with precise multilevel programming. Proof-of-concept open-loop and closed-loop IMC experiments are carried out on 3D-VRRAM, and to further support the application of IMC on real-life scenarios, the work also reports demonstration of relatively large-size image compression problems adopting 2D-RRAM. In addition, the dissertation presents results on the heterointegration between RRAM and spin-Hall nano-oscillators (SHNO), in the context of achieving a hybrid neural network. The RRAMs provide the non-volatile storage of synaptic weights, while SHNO act as fast oscillating neurons.
Il consumo energetico delle tecnologie dell’informazione e della comunicazione è in costante crescita a livello globale, e l’ascesa dell’Intelligenza Artificiale (AI) comporta la generazione di enormi quantità di dati che devono essere elaborati e memorizzati. Le attività data-centriche, come l’addestramento e l’inferenza nei modelli di AI, impongono un notevole carico computazionale sui sistemi digitali tradizionali basati sull’architettura di Von Neumann. Alla luce di queste tendenze, risulta di fondamentale importanza lo studio di dispositivi di calcolo e memoria non volatili di nuova generazione, in cui dati e logica coesistono. Il paradigma del calcolo in-memory (IMC) rappresenta un cambio di prospettiva promettente, in quanto consente l’esecuzione di compiti altamente specializzati direttamente all’interno delle matrici di memoria, migliorando l’efficienza delle elaborazioni data-intensive — come il deep learning in edge computing — grazie all’eliminazione del trasferimento di dati tra unità computazionali e unità di memoria. Tra le memorie emergenti, le memorie a switching resistivo (RRAM) si distinguono per un ampio ventaglio di proprietà vantaggiose ai fini dell’IMC, quali la capacità di immagazzinamento analogico non volatile, il basso consumo energetico e l’elevata densità. La tecnologia 3D-Vertical RRAM (3D-VRRAM) costituisce un’opzione promettente per ottenere un’elevata capacità di celle di memoria a costi di fabbricazione ridotti, favorendo un’integrazione su larga scala competitiva con le memorie flash NAND 3D allo stato dell’arte. La presente tesi di dottorato descrive la fabbricazione e la caratterizzazione di array crossbar planari 2D e 3D-Verticali basati su celle di HfOx, capaci di operazioni di IMC con programmazione multivello precisa. Sono riportati esperimenti di IMC in configurazione ad anello aperto e ad anello chiuso su 3D-VRRAM; inoltre, al fine di dimostrare l’applicabilità dell’IMC a scenari reali, viene presentata la risoluzione di problemi di compressione di immagini di dimensioni relativamente ampie tramite dispositivi 2D-RRAM. Infine, la tesi presenta risultati sull’eterointegrazione tra RRAM e nano-oscillatori spin-Hall (SHNO), nel contesto della realizzazione di una rete neurale ibrida. In questo schema, le RRAM forniscono la memorizzazione non volatile dei pesi sinaptici, mentre gli SHNO svolgono il ruolo di neuroni.
Fabrication and characterization of planar and 3D-Vertical resistive random-access memory arrays for in-memory computing
DAVIDE, BRIDAROLLI
2025
Abstract
The energy consumption of information and communication technology is increasing worldwide, and the rise of Artificial Intelligence (AI) leads to massive amounts of data that need to be processed and stored. Data-centric tasks, such as training and inference in AI models, impose a heavy computational burden on traditional digital computing systems based on the Von Neumann architecture. Motivated by these trends, the investigation of next-generation non-volatile computing and memory devices in which data and logic coexist is vital. The in-memory computing concept (IMC) is a promising paradigm shift where highly specialized tasks are executed within memory arrays, improving the efficiency of the data-intensive workloads, such as deep learning on the edge, by removing the data shuttling between computational and memory elements. Resistive random access memory (RRAM) is one of the most promising emerging memories, due to a broad range of attractive properties for IMC, such as non-volatile analog storage, low power consumption, and high memory density. 3D-Vertical RRAM (3D-VRRAM) is a promising option to achieve high memory cell capacity with low fabrication cost, toward large-scale integration competitive with state-of-the-art 3D NAND flash. This Doctoral dissertation presents the fabrication and characterization of planar 2D and 3D-VRRAM crossbar arrays (CBA) based on HfOx cells and capable of IMC with precise multilevel programming. Proof-of-concept open-loop and closed-loop IMC experiments are carried out on 3D-VRRAM, and to further support the application of IMC on real-life scenarios, the work also reports demonstration of relatively large-size image compression problems adopting 2D-RRAM. In addition, the dissertation presents results on the heterointegration between RRAM and spin-Hall nano-oscillators (SHNO), in the context of achieving a hybrid neural network. The RRAMs provide the non-volatile storage of synaptic weights, while SHNO act as fast oscillating neurons.| File | Dimensione | Formato | |
|---|---|---|---|
|
Bridarolli_phd_thesis_finale.pdf
accesso solo da BNCF e BNCR
Licenza:
Tutti i diritti riservati
Dimensione
98.41 MB
Formato
Adobe PDF
|
98.41 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/356407
URN:NBN:IT:POLIMI-356407