The number of smart devices for the Internet of Things (IoT) is rapidly growing, and by 2025 almost 80 ZB of data per year will be generated by IoT devices alone, challenging the current cloud computing infrastructure. Thus, a shift to the edge computing paradigm, in which data are processed near their sources, is critical, but its implementation requires new energy efficient computing hardware. The approaching downscaling limit of transistor size implies the need for new nanoscale technologies and a departure from the conventional von Neumann architecture. Also, in-hardware security primitives need to be introduced at the silicon level. Among the possible technologies, emerging non-volatile memories (eNVMs) are very promising and enable the realization of in-memory computing paradigms, in which computation is executed directly inside the memory, therefore bypassing the slow and energy inefficient data exchange over a communication bus, i.e., the main bottleneck of von Neumann architectures. However, the intrinsic stochastic nature of eNVMs presents several challenges which can impact the circuit functionality and reliability. On the other hand, it can be exploited to implement hardware-level security primitives such as True Random Number Generators (TRNGs) and Physical Unclonable Functions (PUF). Thus, appropriate design tools and methodologies are needed to help circuit designers exploit eNVMs strengths while consciously addressing their limitations. The optimization of circuit simulation tools and the development of appropriate methodologies to analyze and improve innovative circuits based on eNVMs for computing and security applications is the goal of this Ph.D. thesis. Specifically, a physics-based Resistive RAM (RRAM) compact model (UniMORE compact model), was developed starting from a prototypical existing version and refined to include self-consistently the role of variability, thermal effects, and Random Telegraph Noise (RTN). In addition, a self-automated parameter extraction procedure is developed and included. Such procedure requires only the results of a few experiments that are commonly employed in the device characterization, and was validated both experimentally and on three RRAM technologies from the literature. The procedure allows quick model calibration and helps in determining the strengths and weaknesses of different RRAM technologies for a dependable device-circuit co-optimization. The calibrated compact model is used to analyze the performance and reliability trade-offs of different in-memory computing paradigms. Specifically, the results of circuits simulations of state-of-the-art Logic-in-Memory (LiM) circuits based on the material implication (IMPLY) logic and RRAM technology enabled the development of design procedures for optimizing their reliability, which are here discussed. Also, a novel smart IMPLY (SIMPLY) LiM architecture, which solves the circuit reliability issues of conventional IMPLY architectures, was developed. The reliability and performances of the SIMPLY architecture were thoroughly investigated considering different RRAM technologies and benchmarked on complex operations. Furthermore, the results of the study on RRAM-based low-bit precision neural networks (NNs) analog hardware accelerators are presented, highlighting specific reliability and performance trade-offs. Also, a novel hybrid in-memory computing hardware accelerator in which both SIMPLY and the analog vector matrix multiplication framework coexist on the same memory crossbar array is demonstrated. Finally, challenges and opportunities for RTN-based TRNG circuits are discussed, by exploiting the results of circuit simulations in which experimentally measured RTN data from different RRAM technologies are used.

L’enorme mole di dati prodotta dai dispositivi per l’Internet of Things richiede una trasformazione dell’attuale infrastruttura di cloud computing: lo spostamento di parte dell’elaborazione dove i dati vengono generati (edge computing). Per attuare questo cambio di paradigma, lo sviluppo di nuove soluzioni hardware per l’elaborazione dei dati più efficienti è fondamentale. Inoltre, il sopraggiungere del limite fisico di miniaturizzazione dei transistor implica la necessità di sviluppare nuovi nanodispositivi e nuove architetture di calcolo, che si scostano dalla tradizionale architettura di von Neumann. Inoltre, i sempre più stringenti requisiti di sicurezza richiedono l’introduzione di primitive di sicurezza direttamente a livello hardware. Le tecnologie emergenti nell’ambito delle memorie non volatili (eNVM) permettono l’implementazione di paradigmi di elaborazione in memoria, che eliminano il principale collo di bottiglia delle architetture di von Neumann, ovvero l’inefficiente scambio di dati tra la memoria e l’unità di elaborazione. Tuttavia, la natura stocastica dei dispositivi di eNVM influisce sulla funzionalità e sull'affidabilità di questi circuiti, complicandone la progettazione. D'altra parte, questi fenomeni stocastici possono essere sfruttati per implementare primitive di sicurezza a livello hardware. Pertanto, per sfruttare le potenzialità delle eNVM sono fondamentali nuovi strumenti e nuove metodologie di progettazione. L’obiettivo di questa tesi di dottorato è l'ottimizzazione di strumenti per la simulazione circuitale e lo sviluppo di metodologie appropriate per l’analisi ed il miglioramento di circuiti innovativi basati sulle eNVM per applicazioni di elaborazione e sicurezza. In particolare, viene proposto un modello compatto di RAM resistiva (RRAM) (modello compatto UniMORE), sviluppato a partire da una versione prototipale esistente e perfezionato per includere in modo autoconsistente il ruolo della variabilità, degli effetti termici e del Random Telegraph Noise (RTN). Inoltre, viene descritta una procedura di estrazione dei parametri semi-automatica che richiede solo i risultati di alcuni esperimenti comunemente impiegati durante la caratterizzazione dei dispositivi, e che è stata validata sia sperimentalmente che su tre tecnologie RRAM dalla letteratura. Grazie al modello compatto calibrato, vengono analizzate le prestazioni e l’affidabilità circuitale di diversi paradigmi di elaborazione in memoria. Mediante simulazioni circuitali di circuiti Logic-in-Memory (LiM) basati sulla logica di implicazione materiale (IMPLY) e sulla tecnologia RRAM, sono state sviluppate procedure di progettazione volte a massimizzarne l’affidabilità circuitale. Inoltre, è stata sviluppata una nuova architettura LiM chiamata smart IMPLY (SIMPLY), che risolve i problemi di affidabilità comunemente presenti nei circuiti convenzionali. L'affidabilità e le prestazioni dell'architettura SIMPLY sono state studiate dettagliatamente considerando diverse tecnologie RRAM e l’esecuzione di operazioni complesse. Inoltre, tramite simulazioni circuitali sono stati analizzati acceleratori hardware di reti neurali a bassa precisione, evidenziandone i compressi esistenti tra affidabilità ed elevate prestazioni. Viene anche presentato un nuovo acceleratore hardware che combina sullo stesso array di memoria due diversi paradigmi di elaborazione in memoria, ovvero SIMPLY e l’accelerazione in analogico del prodotto matrice vettore. Infine, vengono discusse le sfide e le opportunità per i circuiti True Random Number Generators basati su RTN, mediante i risultati di simulazioni circuitali che sfruttano segnali RTN misurati sperimentalmente da diverse tecnologie RRAM.

Circuiti innovativi ad alta efficienza energetica per l'elaborazione sicura in memoria basati su dispositivi di memoria resistivi / Tommaso Zanotti , 2022 Mar 25. 34. ciclo, Anno Accademico 2020/2021.

Circuiti innovativi ad alta efficienza energetica per l'elaborazione sicura in memoria basati su dispositivi di memoria resistivi

ZANOTTI, TOMMASO
2022

Abstract

The number of smart devices for the Internet of Things (IoT) is rapidly growing, and by 2025 almost 80 ZB of data per year will be generated by IoT devices alone, challenging the current cloud computing infrastructure. Thus, a shift to the edge computing paradigm, in which data are processed near their sources, is critical, but its implementation requires new energy efficient computing hardware. The approaching downscaling limit of transistor size implies the need for new nanoscale technologies and a departure from the conventional von Neumann architecture. Also, in-hardware security primitives need to be introduced at the silicon level. Among the possible technologies, emerging non-volatile memories (eNVMs) are very promising and enable the realization of in-memory computing paradigms, in which computation is executed directly inside the memory, therefore bypassing the slow and energy inefficient data exchange over a communication bus, i.e., the main bottleneck of von Neumann architectures. However, the intrinsic stochastic nature of eNVMs presents several challenges which can impact the circuit functionality and reliability. On the other hand, it can be exploited to implement hardware-level security primitives such as True Random Number Generators (TRNGs) and Physical Unclonable Functions (PUF). Thus, appropriate design tools and methodologies are needed to help circuit designers exploit eNVMs strengths while consciously addressing their limitations. The optimization of circuit simulation tools and the development of appropriate methodologies to analyze and improve innovative circuits based on eNVMs for computing and security applications is the goal of this Ph.D. thesis. Specifically, a physics-based Resistive RAM (RRAM) compact model (UniMORE compact model), was developed starting from a prototypical existing version and refined to include self-consistently the role of variability, thermal effects, and Random Telegraph Noise (RTN). In addition, a self-automated parameter extraction procedure is developed and included. Such procedure requires only the results of a few experiments that are commonly employed in the device characterization, and was validated both experimentally and on three RRAM technologies from the literature. The procedure allows quick model calibration and helps in determining the strengths and weaknesses of different RRAM technologies for a dependable device-circuit co-optimization. The calibrated compact model is used to analyze the performance and reliability trade-offs of different in-memory computing paradigms. Specifically, the results of circuits simulations of state-of-the-art Logic-in-Memory (LiM) circuits based on the material implication (IMPLY) logic and RRAM technology enabled the development of design procedures for optimizing their reliability, which are here discussed. Also, a novel smart IMPLY (SIMPLY) LiM architecture, which solves the circuit reliability issues of conventional IMPLY architectures, was developed. The reliability and performances of the SIMPLY architecture were thoroughly investigated considering different RRAM technologies and benchmarked on complex operations. Furthermore, the results of the study on RRAM-based low-bit precision neural networks (NNs) analog hardware accelerators are presented, highlighting specific reliability and performance trade-offs. Also, a novel hybrid in-memory computing hardware accelerator in which both SIMPLY and the analog vector matrix multiplication framework coexist on the same memory crossbar array is demonstrated. Finally, challenges and opportunities for RTN-based TRNG circuits are discussed, by exploiting the results of circuit simulations in which experimentally measured RTN data from different RRAM technologies are used.
Innovative Energy-Efficient Circuits Enabled by Resistive Memory devices for Secure In-Memory Computing
25-mar-2022
PAVAN, Paolo
PUGLISI, Francesco Maria
File in questo prodotto:
File Dimensione Formato  
PhD_Thesis_Tommaso_Zanotti.pdf

Open access

Descrizione: Tesi definitiva Zanotti Tommaso
Tipologia: Tesi di dottorato
Dimensione 9.41 MB
Formato Adobe PDF
9.41 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

Licenza Creative Commons
I metadati presenti in IRIS UNIMORE sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono rilasciati con licenza Attribuzione 4.0 Internazionale (CC BY 4.0), salvo diversa indicazione.
In caso di violazione di copyright, contattare Supporto Iris

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11380/1271184
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact