According to Moore's Law, in the last decades the microelectronics industry has experienced an exponential growth in terms of performance and computing capabilities, thanks to the inexorable scaling of electronic components, reflected by the growing number of transistors on integrated circuits (IC) over the years. Such scaling path is nowadays coming to an end due to structural and physical limitations, such as the heat wall and the corresponding upper limit imposed to the clock frequency. Furthermore, traditional computing systems based on the von Neumann architecture are inherently disadvantaged by the physical separation of processing and memory units, since the data transfer between these two locations is much more slower than the CPU opertions. This performance gap between logic and memory constitutes a real bottleneck, called memory wall, and makes this kind of architecture generally inefficient in terms of time and energy consumption especially in applications that requires the collection and the processing of large amount of data, such as machine learning. In order to overcome these limitations, which have now become very stringent in the era of Internet of Things (IoT) and Big Data, the interest in new computing paradigms has increased in recent years, such as in-memory computing, neuromorphic computing and stochastic computing. In this context, emerging memory technologies called memristors, such as resistive switching random access memory (RRAM), phase change memory (PCM), ferroelectric memory (FeRAM) and spin-transfer torque magnetic memory (STT-MRAM) are being investigated in light of their non-volatility, area scalability, low current and fast operation, as well as compatibility with the CMOS process. Despite the great potential of these devices, an in-depth study is still required to overcome the numerous limitations that arise with their use in these fields, due to their stochastic nature. On the other hand, the inherent stochasticity of these devices becomes a strong point in favor of their use in security applications, such as the development of hardware primitives as Physical Unclonable Functions (PUFs). This thesis presents a study of the development process of a PUF, focusing on the figures of merit and on the essential characteristics that these kind of devices need to show in order to be competitive. The aim is to understand advantages and drawbacks of potential memristor-based implementations, proposing both analog and digital solutions. In particular, the work can be divided into two parts: the first part focuses on the feasibility study of a simple analog implementation starting from scratch, while the second part increases the complexity of the architectures in order to provide better performance. More specifically, the thesis is organized in 5 chapters, each of which addresses a different topic. Chapter 1 gives an overview of the current market scenario, addressing the problems that fuel the demand for new computing paradigms. This chapter also addresses the general hardware-based security concepts that provide best performance in the security field and the emerging memristive devices, potential candidates for these applications. Chapter 2 focuses on the definition, properties, types and figures of merit that characterize the Physical Unclonable Functions (PUF), with a general look to standard CMOS-based solutions. The aim of this chapter is to provide the background useful to analyze a PUF instance in terms of performance. Chapter 3 is entirely dedicated to a preliminary study of the development process of a possible analog implementation, exploiting a 1T1R architecture, starting from scratch. The chapter is strongly oriented towards the feasibility study of a possible solution, but also towards framing problems and requirements. Chapter 4 aims to solve some critical issues emerged in the feasibility study of an analog implementation, increasing the complexity of the architecture. In order to be competitive, a PUF solution must accomplish several tasks and this chapter shows pros and cons of possible new architectures. Chapter 5 proposes a digital alternative solution in order to overcome some critical limitations emerged during the study of the previous analog implementations. Starting from this point, the proposed implementation is analyzed in terms of performance metrics, resilience to ML-attacks and at architecture level.
In accordo con la Legge di Moore, negli ultimi decenni l'industria della microelettronica ha testimoniato una crescita esponenziale in termini di prestazioni e capacità di calcolo, grazie ad un sempre più marcato scaling dei componenti elettronici, riflesso dal crescente numero di transistor nei circuiti integrati (IC) nel corso degli anni. Questo trend di ridimensionamento dei componenti sta oggi volgendo al termine a causa di limitazioni strutturali e fisiche, come per esempio il così detto heat wall e il corrispondente limite superiore che viene imposto alla frequenza di clock per evitare una eccessiva dissipazione di calore. Inoltre, i sistemi di calcolo tradizionali basati sull'architettura von Neumann sono intrinsecamente penalizzati dalla separazione fisica delle unità di elaborazione e di memoria, poiché il trasferimento dei dati tra queste due posizioni è molto più lento delle operazioni della CPU. Questo gap prestazionale tra logica e memoria costituisce un vero e proprio collo di bottiglia, chiamato memory wall, e rende questo tipo di architettura generalmente inefficiente in termini di consumo energetico e di tempo soprattutto in applicazioni che richiedono la raccolta e l'elaborazione di grandi quantità di dati, come il machine learning. Al fine di superare queste limitazioni, divenute ormai molto stringenti nell'era dell'Internet of Things (IoT) e dei Big Data, negli ultimi anni nuovi paradigmi di calcolo hanno acquisito crescente interesse, come per esempio l'in-memory computing, il neuromorphic computing e lo stochastic computing. In questo contesto, le tecnologie di memoria emergenti note come memristori, quali “RRAM”, “PCM”, “FeRAM” e “STT-MRAM” sono oggetto di studio per l'impiego nella realizzazione di questi nuovi paradigmi di calcolo grazie alla loro non volatilità, elevata scalabilità dell'area, funzionamento veloce e a bassa corrente, nonché compatibilità con il processo CMOS. Nonostante il grande potenziale di questi dispositivi, è ancora necessario uno studio approfondito per superare le numerose limitazioni che si presentano con il loro utilizzo in questi campi, dovute alla loro natura stocastica. D'altra parte, la stocasticità intrinseca di questi dispositivi diventa un punto di forza per il loro utilizzo in applicazioni di sicurezza, come per esempio lo sviluppo di primitive hardware Physical Unclonable Functions (PUF). Questa tesi presenta uno studio del processo di sviluppo di una istanza PUF, concentrandosi sulle figure di merito e sulle caratteristiche essenziali che questo tipo di dispositivo deve mostrare per essere competitivo sul mercato. L'obiettivo dell'elaborato è quello di arrivare a comprendere vantaggi e svantaggi di potenziali implementazioni basate su memristori, proponendo soluzioni sia analogiche che digitali. In particolare, il lavoro può essere suddiviso in due macroaree: la prima si concentra sullo studio di fattibilità di una semplice implementazione analogica partendo dall'inizio, mentre la seconda parte si focalizza su concetti utili ad incrementare la complessità delle architetture al fine di fornire prestazioni migliori. Nello specifico, la tesi è organizzata in 5 capitoli, ognuno dei quali affronta un argomento diverso. Il Capitolo 1 offre una panoramica generale sull'attuale scenario di mercato, affrontando i problemi che alimentano la domanda di nuovi paradigmi computazionali. Questo capitolo affronta anche i concetti generali di primitive hardware per la sicurezza e offre un primo sguardo sui dispositivi memristivi emergenti, potenziali candidati per queste applicazioni. Il Capitolo 2 si concentra sulla definizione, sulle proprietà, sui tipi esistenti e sulle figure di merito che caratterizzano i PUF, con uno sguardo generale alle classiche implementazioni basate su tecnologia CMOS. Lo scopo di questo capitolo è fornire il background utile per analizzare un'istanza PUF in termini di prestazioni. Il Capitolo 3 è interamente dedicato allo studio preliminare del processo di sviluppo di una possibile implementazione analogica, sfruttando un'architettura 1T1R. Il capitolo è fortemente orientato allo studio di fattibilità di una possibile soluzione, ma anche all'inquadramento dei requisiti e dei problemi emergenti. Il Capitolo 4 si propone di risolvere alcune criticità emerse nello studio di fattibilità di un'implementazione analogica, aumentando la complessità dell'architettura. Per essere competitiva, una soluzione PUF deve svolgere diversi compiti e questo capitolo mostra pro e contro di possibili nuove architetture. Il Capitolo 5 propone una soluzione alternativa digitale al fine di superare alcune criticità emerse durante lo studio delle precedenti implementazioni analogiche. Partendo da questo presupposto fondamentale, l'implementazione proposta viene quindi analizzata in termini di prestazioni, resilienza agli attacchi di machine learning e a livello di architettura.
Development of a resistive memory array based physical unclonable function
Cattaneo, Lorenzo
2019/2020
Abstract
According to Moore's Law, in the last decades the microelectronics industry has experienced an exponential growth in terms of performance and computing capabilities, thanks to the inexorable scaling of electronic components, reflected by the growing number of transistors on integrated circuits (IC) over the years. Such scaling path is nowadays coming to an end due to structural and physical limitations, such as the heat wall and the corresponding upper limit imposed to the clock frequency. Furthermore, traditional computing systems based on the von Neumann architecture are inherently disadvantaged by the physical separation of processing and memory units, since the data transfer between these two locations is much more slower than the CPU opertions. This performance gap between logic and memory constitutes a real bottleneck, called memory wall, and makes this kind of architecture generally inefficient in terms of time and energy consumption especially in applications that requires the collection and the processing of large amount of data, such as machine learning. In order to overcome these limitations, which have now become very stringent in the era of Internet of Things (IoT) and Big Data, the interest in new computing paradigms has increased in recent years, such as in-memory computing, neuromorphic computing and stochastic computing. In this context, emerging memory technologies called memristors, such as resistive switching random access memory (RRAM), phase change memory (PCM), ferroelectric memory (FeRAM) and spin-transfer torque magnetic memory (STT-MRAM) are being investigated in light of their non-volatility, area scalability, low current and fast operation, as well as compatibility with the CMOS process. Despite the great potential of these devices, an in-depth study is still required to overcome the numerous limitations that arise with their use in these fields, due to their stochastic nature. On the other hand, the inherent stochasticity of these devices becomes a strong point in favor of their use in security applications, such as the development of hardware primitives as Physical Unclonable Functions (PUFs). This thesis presents a study of the development process of a PUF, focusing on the figures of merit and on the essential characteristics that these kind of devices need to show in order to be competitive. The aim is to understand advantages and drawbacks of potential memristor-based implementations, proposing both analog and digital solutions. In particular, the work can be divided into two parts: the first part focuses on the feasibility study of a simple analog implementation starting from scratch, while the second part increases the complexity of the architectures in order to provide better performance. More specifically, the thesis is organized in 5 chapters, each of which addresses a different topic. Chapter 1 gives an overview of the current market scenario, addressing the problems that fuel the demand for new computing paradigms. This chapter also addresses the general hardware-based security concepts that provide best performance in the security field and the emerging memristive devices, potential candidates for these applications. Chapter 2 focuses on the definition, properties, types and figures of merit that characterize the Physical Unclonable Functions (PUF), with a general look to standard CMOS-based solutions. The aim of this chapter is to provide the background useful to analyze a PUF instance in terms of performance. Chapter 3 is entirely dedicated to a preliminary study of the development process of a possible analog implementation, exploiting a 1T1R architecture, starting from scratch. The chapter is strongly oriented towards the feasibility study of a possible solution, but also towards framing problems and requirements. Chapter 4 aims to solve some critical issues emerged in the feasibility study of an analog implementation, increasing the complexity of the architecture. In order to be competitive, a PUF solution must accomplish several tasks and this chapter shows pros and cons of possible new architectures. Chapter 5 proposes a digital alternative solution in order to overcome some critical limitations emerged during the study of the previous analog implementations. Starting from this point, the proposed implementation is analyzed in terms of performance metrics, resilience to ML-attacks and at architecture level.File | Dimensione | Formato | |
---|---|---|---|
2021_04_Cattaneo.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
10.33 MB
Formato
Adobe PDF
|
10.33 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/10589/174076