The continuous advancement of artificial intelligence (AI) and high-performance computing (HPC) has exposed the limitations of traditional von Neumann architectures, necessitating new computing paradigms to overcome power and memory bottlenecks, namely,the frequency wall and memory wall. This thesis work aims to study the OTS selector device characteristic in order to highlight a recent discovery and apply it for a high density cross point array for low power In-Memory Computing (IMC). The referred discovery is called polarity effect, and consists in modulation of the OTS threshold voltage switch typical of chalcogenide glasses. This effect allows this selector to act also as a memory device and would also simplify the fabrication process of already commercially available cross-point array technologies based on separate elements (Intel Optane 2017). By operating in the subthreshold regime, it is possible to have 5 orders of magnitude lower currents and to lower the computational power for MVM. The first chapter presents the main limitation in traditional computing, where Moore’s Law has dictated an exponential growth for the last 5 decades. The problem of continuous scaling has led to increased power dissipation and it has shown the physical limitation of this approach. Another limit of traditional computing is presented in the Von Neumann’s bottleneck, where the architectural limitation of separate processor and memory have hindered performance. Then some technologies, like storage class memory (SCM), neuromorphic computing and In-Memory Computing (IMC), are investigated as possible solutions. Among these, cross-point (XP) architectures provide an extremely efficient framework for matrix-vector multiplications (MVM), critical for AI workloads, like convolution operation. The second chapter of this work focuses on two-terminal emerging memory technologies, including Phase-Change Memory (PCM) which have been implemented in a commercially available product (Intel Optane 2017). Following, selector devices are presented, with the set of requirement and a comprehensive list of possible candidates. Particular attention is paid to OTS selectors based on chalcogenide glasses, which have emerged as a critical component for high-density XP arrays due to their abrupt switching behavior and compatibility with back-end-of-line (BEOL) integration. A comprehensive study of OTS device physics is conducted, analyzing conduction mechanisms, threshold switching behavior, and material properties. Various theoretical models, including Poole-Frenkel conduction, impact ionization, and field-assisted defect migration, are examined to provide insight into OTS switching dynamics. The third chapter outlines the measurement setup utilized for the device under test. It provides a detailed list of instruments used for both high- and low-frequency measurements. Additionally, the chapter explains the functioning of the filter designed to ensure high-precision measurements. Finally, it describes the system that facilitates communication between the instruments, to allow an increased measuring throughput. The fourth chapter explores the general characteristics of the OTS device in a 3D structure, with emphasis on the polarity effect. The primary objective is to conduct a quantitative analysis to determine how the OTS state can be controlled and to identify the optimal bias point for operating in the subthreshold regime during programming and reading operations. During the research process, novel behaviors of the electrical characteristics have been discovered and described through the dependencies with the control parameters. To show that the outcomes in subthreshold are applicable also to the more popular ON-conduction regime, a tight correlation between the two is reported. At the end, the collected results are analyzed to assess which model best fits the device’s electrical behavior. The fifth chapter dives into the characterization of the OTS device for IMC, with great attention on the reliability of the read operation, which is significantly affected by drift processes and read stress. Following it presents a series of experiments designed to mitigate and compensate these unwanted effects. Then, a concluding experiment integrates all the insights gained to demonstrate a refresh operation, ensuring ultra-stable read performance. The findings of this thesis show the potential of OTS-based Selector-Only Memory (SOM) to revolutionize memory and computing architectures by reducing power consumption, improving scalability, and enabling novel IMC solutions.
L’avanzamento continuo dell’intelligenza artificiale (AI) e del calcolo ad alte prestazioni (HPC) ha messo in evidenza i limiti delle architetture von Neumann tradizionali, richiedendo nuovi paradigmi di calcolo per superare i colli di bottiglia di potenza e memoria, ossia la frequency wall e la memory wall. Questo lavoro di tesi mira a studiare le caratteristiche del dispositivo selettore OTS per evidenziare una recente scoperta e applicarla in matrici cross-point (XP) ad alta densità per un’implementazione di computazione in memoria (IMC) a bassa potenza. La scoperta in questione è chiamata effetto di polarita’ e consiste nella modulazione della tensione di soglia (threshold voltage) di uno switch OTS tipico di vetri calcogenuri. Questo effetto consente al selettore di agire anche come memoria, semplificando quindi il processo di fabbricazione delle tecnologie cross-point basate su elementi separati (già disponibili in prodotti commerciali, come Intel Optane 2017). L’operazione in regime sottosoglia permette correnti inferiori di 5 ordini di grandezza e riduce la potenza di calcolo nelle moltiplicazioni matrice vettore (MVM). Il primo capitolo presenta la principale limitazione del calcolo tradizionale, in cui la Legge di Moore ha definito una crescita esponenziale negli ultimi 50 anni. Il problema della continua miniaturizzazione ha portato a un aumento della dissipazione di potenza e ha mostrato il limite fisico di questo approccio. Un altro limite del calcolo tradizionale è il Von Neumann bottleneck, in cui la separazione architetturale tra processore e memoria ostacola le prestazioni. Vengono quindi analizzate alcune tecnologie, come la storage class memory (SCM), il calcolo neuromorfico e l’in-memory computing (IMC), proposte come possibili soluzioni. Tra queste, le architetture cross-point (XP) offrono un framework estremamente efficiente per le MVM, fondamentali nei carichi di lavoro di AI (ad esempio nelle operazioni di convoluzione). Il secondo capitolo si focalizza sulle tecnologie emergenti a due terminali, tra cui le memorie a cambio di fase (phase-change memory, PCM), già implementate in prodotti commer- ciali (Intel Optane 2017). Successivamente, vengono introdotti i dispositivi selettori, con la lista dei requisiti e una panoramica delle possibili alternative. Particolare attenzione è rivolta ai selettori OTS (ovonic threshold switch) basati su vetri calcogenuri, che si sono imposti come componenti fondamentali per le matrici XP ad alta densità, grazie al loro comportamento di commutazione brusca e alla compatibilità con l’integrazione BEOL (back-end-of-line). Si analizzano in modo approfondito la fisica di funzionamento dei dispositivi OTS, i meccanismi di conduzione, il comportamento di threshold switching e le proprietà dei materiali. Vengono discussi vari modelli teorici, tra cui conduzione Poole-Frenkel, ionizzazione per impatto e migrazione di difetti assistita dal campo elettrico, per fornire una visione globale della dinamica di commutazione OTS. Il terzo capitolo delinea il setup di misura utilizzato per i dispositivi in esame. Elenca in dettaglio gli strumenti adoperati sia per misure ad alta frequenza, sia a bassa frequenza, illustrando inoltre il filtro progettato per garantire misure di elevata precisione. Infine, descrive il sistema di comunicazione tra gli strumenti, che permette di incrementare la produttività delle operazioni di misura. Il quarto capitolo illustra le caratteristiche generali del dispositivo OTS in configurazione 3D, con particolare enfasi sul effetto di polarita’. L’obiettivo principale è condurre un’analisi quantitativa per comprendere come controllare lo stato dell’OTS e identificare il punto di polarizzazione ottimale per operare in regime sottosoglia durante le fasi di programmazione e lettura. Successivamente, i risultati raccolti vengono analizzati per verificare quale modello teorico si adatti meglio al comportamento elettrico del dispositivo. Il quinto capitolo si concentra sulla caratterizzazione del dispositivo OTS per applicazioni IMC, con particolare attenzione all’affidabilità della lettura, spesso condizionata da processi di drift e stress di lettura. Viene poi illustrata una serie di esperimenti che mirano a mitigare e compensare tali fenomeni, culminando in un test finale che integra tutti i risultati ottenuti per dimostrare un’operazione di refresh e assicurare prestazioni di lettura estremamente stabili. I risultati di questa tesi evidenziano il potenziale delle memorie basate su OTS (Selector- Only Memory, SOM) di rivoluzionare le architetture di memoria e calcolo, consentendo di ridurre i consumi energetici, migliorare la scalabilità e abilitare nuove soluzioni IMC.
Characterization of 3D ovonic threshold switch devices for high density and low power in-memory computing
HU, GIUSEPPE YI CHENG
2023/2024
Abstract
The continuous advancement of artificial intelligence (AI) and high-performance computing (HPC) has exposed the limitations of traditional von Neumann architectures, necessitating new computing paradigms to overcome power and memory bottlenecks, namely,the frequency wall and memory wall. This thesis work aims to study the OTS selector device characteristic in order to highlight a recent discovery and apply it for a high density cross point array for low power In-Memory Computing (IMC). The referred discovery is called polarity effect, and consists in modulation of the OTS threshold voltage switch typical of chalcogenide glasses. This effect allows this selector to act also as a memory device and would also simplify the fabrication process of already commercially available cross-point array technologies based on separate elements (Intel Optane 2017). By operating in the subthreshold regime, it is possible to have 5 orders of magnitude lower currents and to lower the computational power for MVM. The first chapter presents the main limitation in traditional computing, where Moore’s Law has dictated an exponential growth for the last 5 decades. The problem of continuous scaling has led to increased power dissipation and it has shown the physical limitation of this approach. Another limit of traditional computing is presented in the Von Neumann’s bottleneck, where the architectural limitation of separate processor and memory have hindered performance. Then some technologies, like storage class memory (SCM), neuromorphic computing and In-Memory Computing (IMC), are investigated as possible solutions. Among these, cross-point (XP) architectures provide an extremely efficient framework for matrix-vector multiplications (MVM), critical for AI workloads, like convolution operation. The second chapter of this work focuses on two-terminal emerging memory technologies, including Phase-Change Memory (PCM) which have been implemented in a commercially available product (Intel Optane 2017). Following, selector devices are presented, with the set of requirement and a comprehensive list of possible candidates. Particular attention is paid to OTS selectors based on chalcogenide glasses, which have emerged as a critical component for high-density XP arrays due to their abrupt switching behavior and compatibility with back-end-of-line (BEOL) integration. A comprehensive study of OTS device physics is conducted, analyzing conduction mechanisms, threshold switching behavior, and material properties. Various theoretical models, including Poole-Frenkel conduction, impact ionization, and field-assisted defect migration, are examined to provide insight into OTS switching dynamics. The third chapter outlines the measurement setup utilized for the device under test. It provides a detailed list of instruments used for both high- and low-frequency measurements. Additionally, the chapter explains the functioning of the filter designed to ensure high-precision measurements. Finally, it describes the system that facilitates communication between the instruments, to allow an increased measuring throughput. The fourth chapter explores the general characteristics of the OTS device in a 3D structure, with emphasis on the polarity effect. The primary objective is to conduct a quantitative analysis to determine how the OTS state can be controlled and to identify the optimal bias point for operating in the subthreshold regime during programming and reading operations. During the research process, novel behaviors of the electrical characteristics have been discovered and described through the dependencies with the control parameters. To show that the outcomes in subthreshold are applicable also to the more popular ON-conduction regime, a tight correlation between the two is reported. At the end, the collected results are analyzed to assess which model best fits the device’s electrical behavior. The fifth chapter dives into the characterization of the OTS device for IMC, with great attention on the reliability of the read operation, which is significantly affected by drift processes and read stress. Following it presents a series of experiments designed to mitigate and compensate these unwanted effects. Then, a concluding experiment integrates all the insights gained to demonstrate a refresh operation, ensuring ultra-stable read performance. The findings of this thesis show the potential of OTS-based Selector-Only Memory (SOM) to revolutionize memory and computing architectures by reducing power consumption, improving scalability, and enabling novel IMC solutions.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Hu_Tesi.pdf
non accessibile
Descrizione: Tesi, 2025/04, Hu
Dimensione
34.7 MB
Formato
Adobe PDF
|
34.7 MB | Adobe PDF | Visualizza/Apri |
2025_04_Hu_executive summary.pdf
non accessibile
Descrizione: executive summary, 2025/04, Hu
Dimensione
4.08 MB
Formato
Adobe PDF
|
4.08 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/236303