The information revolution has totally changed our human life and society. The widespread diffusion of internet and mobile technologies had an impact on the world that is probably more dramatic than the invention of steam engines and cars at the begin of XX century. This revolution has been possible thanks to the big efforts put by electronic companies on the development of integrated circuits, according to Moore's law. The MOSFET transistor, at the basis of computation in electronic processors, was shrunk from the 20 mm channel of 1975 to the 14 nm channel of 2014: There is no other technology in history that was able to accomplish such a dramatic improvement in such a short time. In the Turing machine approach, the ability to store information (memory function) is as important as the ability to process the information (logic function). For this reason, the development of logic devices (transistors) came along with the development of memory devices. Modern PCs and mobile devices make use of several types of memory, which differ in terms of speed, cost and data retention time. Among these, the non-volatile memory must retain data for years also when the power supply is switched off. The dominant non-volatile memory technology in the past decades up to the present day has been the Flash memory, in which the bit of information is stored as an electric charge in the floating gate of a MOSFET device. Nowadays, the Flash technology is facing several issue related to scaling, among which the most important are random telegraph noise, electrostatic control of oating gate and variability. For this reason, memory companies are trying to find alternative solutions to Flash memory. One of the most promising technology, which has already reached the industrial maturity, is the phase change memory (PCM). The PCM is a particular type of resistive memory, where the reversible phase transition of the active chalcogenide material, usually Ge2Sb2Te5 (GST), is used to store the logic bit of information. The two stable states in the memory correspond to the high-resistance amorphous phase (reset state) and the low-resistance crystalline phase (set state). PCM devices have been scaled to the 20-nm size, while low-power and nanosecond-switching operation has been demonstrated. On the other hand, the ultimate scaling of the PCM is still unclear, due to the impact of random telegraph noise, crystallization, and resistance drift. Nowadays, a deeper knowledge of the PCM physics is strongly requested to drive the development of the PCM technology in the years to come. This motivates the need for research activities, such as the ones described in this doctoral dissertation. The introductory chapter of this thesis provides an overview of the current non-volatile memory (NVM) scenario, subdividing the possible technology evolutions within an evolutionary scenario and a paradigm shift. The phase-change technology is then introduced, dealing with its history, the basic operation and the elementary physical description. This chapter reviews the current state-of-the-art in the physical comprehension of sub-threshold conduction, threshold switching, crystallization and structural relaxation, providing the basic elements needed for the comprehension of the following four chapters. Finally, the current perspectives of the PCM technology are discussed, with a quick glance on the so-called PCMS architecture, which is expected to solve the current PCM limitations in terms of size scaling by stacking the memory element and a cell selector made of another chalcogenide material. The second chapter is devoted to the study of structural relaxation-related phenomena in the amorphous phase of phase change memory devices. The chapter is particularly focused on the characterization and modeling of threshold voltage drift induced by structural relaxation in amorphous Ge2Sb2Te5. This effect leads to threshold voltage increase with time after program operation, which must be carefully controlled to avoid program/read failure in the memory device, especially in PCMS, where both the selector and the memory devices are affected by drift. Then, we show how it is possible to accelerate threshold voltage drift by mean of electrical pulses in the subthreshold region, which could represent an important tool from the application point view to limit the drift effect in the memory device. Finally, the study is extended to the modeling of resistance drift in the crystalline state of Ge-rich Ge-Sb-Te alloys for embedded non-volatile memory applications (ePCM). We show evidence of resistance drift and decay, which are attributed in our model to structural relaxation at the grain boundaries of the poly-crystalline state, and to grain boundaries coalescence respectively. The third chapter of this thesis deals with a detailed study of the retention capability in PCM on a large statistical scale. Such studies are fundamental in order to allow large arrays to properly satisfy the data retention requirements. A wide experimental characterization is then provided in the temperature range below 180C, presenting a detailed study of the cell-to-cell and cycle-to-cycle variability. The overall variability is interpreted through a compact Monte-Carlo model, able to explain both the cell-to-cell and the cycle-to-cycle variability contributions in terms of a pure gaussian spread in the activation energy for crystallization. Finally, the cycle-to-cycle variability is analyzed more in depth, allowing to subdivide the retention characteristics into three separate families namely: i) analog variability, ii) digital (binary) variability and iii) pseudo-repeatable characteristics. The fourth chapter is dedicated to the study of crystallization kinetics in PCM. Our work shows evidence of non-Arrhenius crystallization in GST directly in PCM devices, by comparing the thermally induced crystallization (thermal regime) with the electrically induced crystallization (pulsed regime). The non-Arrhenius crystallization, leading to different activation energies in the Arrhenius plot of crystallization time in the two regimes, is attributed to the fragile nature of GST glass and to the broke of Stoke-Einstein relation above glass transition. We propose a new experimental technique to study electrically induced crystallization down to the holding current. In this way, we were able to extend our study of crystallization kinetics, and to characterize set transition in a wide time range from 50 ns to 10 ms. Then, we model crystallization in PCM by a finite element approach, which is based on filamentary crystallization after threshold switching and on non-Arrhenius crystallization kinetics. Finally, we show evidence of electrically induced crystallization in the subthreshold regime, by performing continuous current stress experiments at low current of about 1 mA and for relatively long times, in the range of 103 s. The fifth chapter deals with a possible alternative application of PCM devices. We exploit the storage ability of PCM, together with additive crystallization and threshold switching, to perform boolean logic operation. In our approach, the PCM device is used as a state machine (memristor), where the state of the device can be changed by mean of electrical pulses applied to the cell. In our work we are able to accomplish a complete set of boolean logic operations, namely the NOT, the NAND and the NOR operation. When compared to standard CMOS logic, the PCM logic offers the advantages of logic-in-memory, of reconfigurable logic and zero static power dissipation, while is shows worse performance in terms of dynamic power consumption, switching time and endurance. This work paves the way for a new field of application for PCM, which together with neuromorphic computation makes this technology attractive for alternative way to compute information in the big data era of the present days.
La rivoluzione informatica ha completamente cambiato le nostre vite e la nostra societa'. La larga diffusione di internet e delle tecnologie mobili ha avuto un impatto sul mondo che molto probabilmente e' stato piu' importante dell'invenzione delle macchine a vapore del motore a scoppio di inizio '900. Questa rivoluzione e' stata resa possibile dal grande sforzo di investimenti messo dalle industrie di elettronica nello sviluppo dei circuiti integrati, in accordo con la legge di Moore. Il transistore MOSFET, che sta alla base del funzionamento dei moderni microprocessori, fu scalato dai 20 mm del 1975 ai 14 nm channel of 2014: Non c'e' nessun'altra tecnologia nella storia dell'uomo che e' stata in grado di ottenere un miglioramento cosi' marcato in un tempo cosi' breve. Nella macchina di Turing, la capacita' di immagazzinare i dati (funzione di memoria) e' altrettanto importante della capacita' di processare i dati (funzione logica). Per questa ragione, lo sviluppo dei dispositivi dedicati alla logica (transistor) e' andato di pari passo allo sviluppo dei dispositivi di memoria. I moderni PC e dispositivi mobili fanno uso di diversi tipi di memoria, che differiscono in termini di velocita', costo e tempo di ritenzione del dato. Tra questi, i dispositivi di memoria non volatili devono essere in grado di mantenere il dato in memoria anche quando l'alimentazione viene spenta. La tecnologia non volatile dominante negli ultimi decenni e' stata senza dubbio la memoria Flash, nel quale il bit di informazione e' immagazzinato come una carica elettrica in una floating gate di un transistore MOSFET. Al giorno d'oggi, la tecnologia Flash si sta scontrando con diversi limiti legati allo scaling, tra i quali i piu' importanti sono il random telegraph noise, la perdita del controllo elettrostatico del canale e della floating gate e la variabilita' su array. Per questa ragione, le industrie di semiconduttori stanno cercando di trovare soluzioni alternative alla tecnologia Flash. Una delle tecnologie piu' promettenti, che ha gia' raggiungo la maturita' industriale, e' la memoria a cambiamento di fase (PCM). La PCM e' un tipo particolare di memoria resistiva, in cui il cambio di fase reversibile del materiale attivo, generalmente la lega calcogenura Ge2Sb2Te5 (GST), e' usata per immagazzinare il bit di informazione. I due stati stabili nella memoria corrispondono allo stato amorfo alto-resistivo (stato di reset) e a allo stato cristallino basso-resistivo (stato di set). I dispositivi PCM sono stati scalati fino alla dimensione di 20 nm, ed e' stata dimostrato il funzionamento a bassa potenza e lo switching resistivo al nanosecondo. D'altro canto, il limite ultimo di scaling delle PCM non e' ancora chiaro, a causa dell'impatto del random telegraph noise, della cristallizzazione, e del drift della resistenza. Per queste ragioni, al giorno d'oggi e' richiesta una conoscenza piu' approfondita della sica delle PCM, al fine di guidare lo sviluppo della tecnologia negli anni a venire. Questo motiva la necessita' dell'attivita' di ricerca, come quella presentata in questo lavoro di dottorato. Il capitolo introduttivo di questa tesi fornisce una panoramica dello scenario corrente nel campo delle memorie non volatili (NVM), distinguendo le possibili evoluzioni della tecnologia in uno scenario evolutivo e in un cambio di paradigma. In questo capitolo vengono riepilogati gli attuali modelli fisici per la conduzione sottosoglia, lo switching a soglia, la cristallizzazione e il rilassamento strutturale, fornendo gli elementi di base per la comprensione dei seguenti quattro capitoli. A fine capitolo vengono discusse le attuali prospettive delle PCM, con una breve appendice alla architettura delle cosiddette PCMS, che sono pensate per risolvere il problema della limitazione della densit a di corrente del selettore, attraverso l'utilizzo di un selettore costituito da un altro materiale calcogenuro nella fase amorfa. Il secondo capitolo e' dedicato allo studio dei fenomeni legati al rilassamento strutturale nella fase amorfa dei dispositivi a cambiamento di fase. Il capitolo si concentra in particolare sulla caratterizzazione e la modellizzazione del drift della tensione di soglia indotto dal rilassamento strutturale nel GST amorfo. Tale effetto porta ad un incremento nel tempo della tensione di soglia dopo l'operazione di programmazione, che deve essere attentamente controllato per evitare fallimenti di lettura/programmazione nel dispositivo dei memoria, specialmente nelle PCMS, dove sia il selettore che il dispositivo di memoria sono soggette a drift. Poi, viene mostrato come sia possibile accelerare il drift della tensione di soglia attraverso l'applicazione di impulsi elettrici nella regione di sottosoglia, che puo' rivelarsi uno strumento importante dal punto di vista applicativo per limitare i fenomeni di drift nelle PCM. Infine, lo studio viene esteso al drift della resistenza dello stato cristallino in leghe Ge-Sb-Te ricche in germanio per applicazioni in memorie embedded non volatili (ePCM). Viene mostrata l'evidenza sperimentale di drift e diminuzione dell resistenza, che sono attribuiti rispettivamente al rilassamento strutturale ai bordi di grano dello stato poli-cristallino, e alla coalescenza dei bordi di grano. Il terzo capitolo di questa tesi fornisce uno studio dettagliato della capacita' di ritenzione delle PCM su un'ampia scala statistica. Tale studio e' fondamentale, in quanto permette ad ampi array di memoria di soddisfare le speciche di ritenzione del dato. Viene quindi presentata un'ampia caratterizzazione statistica nel range di temperatura al di sotto dei 180C, presentando uno studio dettagliato della variabilita' cella-cella e ciclo-ciclo. La variabilita' complessiva viene interpretata attraverso un modello Monte-Carlo compatto, in grado di imputare entrambi i contributi di variabilit a ad una dispersione puramente gaussiana dell'energia di attivazione per la cristallizzazione. Infine, viene analizzata piu' in dettaglio la variabilita' ciclo-ciclo, permettendo di classificare le caratteristiche di ritenzione in tre categorie: i) variabilita' analogica, ii) variabilita' digitale (binaria) e iii) caratteristiche pseudo-ripetibili. Il quarto capitolo e' dedicato allo studio della cinetica di cristallizzazione nelle PMC. Il nostro lavoro evidenzia la cristallizzazione non-Arrhenius nel GST direttamente nei dispositivi PCM, mettendo a confronto la cristallizzazione indotta dall'aumento della temperatura ambiente (regime termico) con la cristallizzazione dovuta all'applicazione di impulsi elettrici alla cella (regime impulsato). La cristallizzazione non-Arrhenius, che porta a diverse energie di attivazione nel diagramma Arrhenius dei tempi di cristallizzazione nei due regimi, e' attribuita alla natura fragile del GST vetroso e alla rottura della relazione di Stoke-Einstein per temperature superiori alla temperatura di transizione vetrosa. Viene quindi proposta una nuova tecnica sperimentale per studiare la cristallizzazione indotta da impulso elettrico fino alla corrente di holding. In questo modo, si riesce ad estendere lo studio della cinetica di cristallizzazione, e a caratterizzare la transizione di set in un ampio intervallo di tempo da 50 ns sino a 10 ms. Si propone quindi un modello a elementi finiti per la cristallizzazione nelle PCM, basato sulla cristallizzazione filamentare dopo lo switching a soglia e sulla cinetica non-Arrhenius dei tempi di cristallizzazione. Infine, viene mostrata l'evidenza sperimentale di cristallizzazione indotta da impulsi elettrici nella regione di sottosoglia, effettuando esperimenti di stress a corrente costante per basse correnti fino a 1 mA e per tempi relativamente lunghi, dell'ordine di 103 s. Il quinto capitolo si occupa di possibili applicazioni alternative dei dispositivi PCM. In particolare, viene mostrato come sia possibile sfruttare la capacita' di memoria delle PCM, unitamente all'additivita' della cristallizzazione allo switching a soglia, per svolgere operazioni logiche booleane. Nel nostro approccio, il dispositivo PCM e' utilizzato come una macchina di stato (memristore), in cui lo stato del dispositivo puo' essere modificato attraverso l'applicazione di impulsi elettrici alla cella. In questo lavoro si dimostra come svolgere un insieme completo di operazioni logiche, speci catamente l'operazione di NOT, di NAND e di NOR. In confronto alla logica standard CMOS, la logica PCM offre il vantaggio di unire le funzioni logiche e di memoria in un unico dispositivo, di mettere a disposizione una logica intrinsecamente riconfigurabile e di mostrare dissipazione di potenza statica nulla. Di contro, la logica PCM risulta peggiore in termini di velocita' di switching, di consumo di potenza dinamico a di durata in ciclatura. Questo lavoro apre la strada per un nuovo campo di applicazione della tecnologia PCM, che insieme all'applicazione in reti neuromorfiche fa di questa tecnologia un'attraente candidato per nuovi modi di processare l'informazione nell'era dei big data in cui siamo immersi.
Switching, reliability and novel functionalities in phase change devices
CIOCCHINI, NICOLA
Abstract
The information revolution has totally changed our human life and society. The widespread diffusion of internet and mobile technologies had an impact on the world that is probably more dramatic than the invention of steam engines and cars at the begin of XX century. This revolution has been possible thanks to the big efforts put by electronic companies on the development of integrated circuits, according to Moore's law. The MOSFET transistor, at the basis of computation in electronic processors, was shrunk from the 20 mm channel of 1975 to the 14 nm channel of 2014: There is no other technology in history that was able to accomplish such a dramatic improvement in such a short time. In the Turing machine approach, the ability to store information (memory function) is as important as the ability to process the information (logic function). For this reason, the development of logic devices (transistors) came along with the development of memory devices. Modern PCs and mobile devices make use of several types of memory, which differ in terms of speed, cost and data retention time. Among these, the non-volatile memory must retain data for years also when the power supply is switched off. The dominant non-volatile memory technology in the past decades up to the present day has been the Flash memory, in which the bit of information is stored as an electric charge in the floating gate of a MOSFET device. Nowadays, the Flash technology is facing several issue related to scaling, among which the most important are random telegraph noise, electrostatic control of oating gate and variability. For this reason, memory companies are trying to find alternative solutions to Flash memory. One of the most promising technology, which has already reached the industrial maturity, is the phase change memory (PCM). The PCM is a particular type of resistive memory, where the reversible phase transition of the active chalcogenide material, usually Ge2Sb2Te5 (GST), is used to store the logic bit of information. The two stable states in the memory correspond to the high-resistance amorphous phase (reset state) and the low-resistance crystalline phase (set state). PCM devices have been scaled to the 20-nm size, while low-power and nanosecond-switching operation has been demonstrated. On the other hand, the ultimate scaling of the PCM is still unclear, due to the impact of random telegraph noise, crystallization, and resistance drift. Nowadays, a deeper knowledge of the PCM physics is strongly requested to drive the development of the PCM technology in the years to come. This motivates the need for research activities, such as the ones described in this doctoral dissertation. The introductory chapter of this thesis provides an overview of the current non-volatile memory (NVM) scenario, subdividing the possible technology evolutions within an evolutionary scenario and a paradigm shift. The phase-change technology is then introduced, dealing with its history, the basic operation and the elementary physical description. This chapter reviews the current state-of-the-art in the physical comprehension of sub-threshold conduction, threshold switching, crystallization and structural relaxation, providing the basic elements needed for the comprehension of the following four chapters. Finally, the current perspectives of the PCM technology are discussed, with a quick glance on the so-called PCMS architecture, which is expected to solve the current PCM limitations in terms of size scaling by stacking the memory element and a cell selector made of another chalcogenide material. The second chapter is devoted to the study of structural relaxation-related phenomena in the amorphous phase of phase change memory devices. The chapter is particularly focused on the characterization and modeling of threshold voltage drift induced by structural relaxation in amorphous Ge2Sb2Te5. This effect leads to threshold voltage increase with time after program operation, which must be carefully controlled to avoid program/read failure in the memory device, especially in PCMS, where both the selector and the memory devices are affected by drift. Then, we show how it is possible to accelerate threshold voltage drift by mean of electrical pulses in the subthreshold region, which could represent an important tool from the application point view to limit the drift effect in the memory device. Finally, the study is extended to the modeling of resistance drift in the crystalline state of Ge-rich Ge-Sb-Te alloys for embedded non-volatile memory applications (ePCM). We show evidence of resistance drift and decay, which are attributed in our model to structural relaxation at the grain boundaries of the poly-crystalline state, and to grain boundaries coalescence respectively. The third chapter of this thesis deals with a detailed study of the retention capability in PCM on a large statistical scale. Such studies are fundamental in order to allow large arrays to properly satisfy the data retention requirements. A wide experimental characterization is then provided in the temperature range below 180C, presenting a detailed study of the cell-to-cell and cycle-to-cycle variability. The overall variability is interpreted through a compact Monte-Carlo model, able to explain both the cell-to-cell and the cycle-to-cycle variability contributions in terms of a pure gaussian spread in the activation energy for crystallization. Finally, the cycle-to-cycle variability is analyzed more in depth, allowing to subdivide the retention characteristics into three separate families namely: i) analog variability, ii) digital (binary) variability and iii) pseudo-repeatable characteristics. The fourth chapter is dedicated to the study of crystallization kinetics in PCM. Our work shows evidence of non-Arrhenius crystallization in GST directly in PCM devices, by comparing the thermally induced crystallization (thermal regime) with the electrically induced crystallization (pulsed regime). The non-Arrhenius crystallization, leading to different activation energies in the Arrhenius plot of crystallization time in the two regimes, is attributed to the fragile nature of GST glass and to the broke of Stoke-Einstein relation above glass transition. We propose a new experimental technique to study electrically induced crystallization down to the holding current. In this way, we were able to extend our study of crystallization kinetics, and to characterize set transition in a wide time range from 50 ns to 10 ms. Then, we model crystallization in PCM by a finite element approach, which is based on filamentary crystallization after threshold switching and on non-Arrhenius crystallization kinetics. Finally, we show evidence of electrically induced crystallization in the subthreshold regime, by performing continuous current stress experiments at low current of about 1 mA and for relatively long times, in the range of 103 s. The fifth chapter deals with a possible alternative application of PCM devices. We exploit the storage ability of PCM, together with additive crystallization and threshold switching, to perform boolean logic operation. In our approach, the PCM device is used as a state machine (memristor), where the state of the device can be changed by mean of electrical pulses applied to the cell. In our work we are able to accomplish a complete set of boolean logic operations, namely the NOT, the NAND and the NOR operation. When compared to standard CMOS logic, the PCM logic offers the advantages of logic-in-memory, of reconfigurable logic and zero static power dissipation, while is shows worse performance in terms of dynamic power consumption, switching time and endurance. This work paves the way for a new field of application for PCM, which together with neuromorphic computation makes this technology attractive for alternative way to compute information in the big data era of the present days.File | Dimensione | Formato | |
---|---|---|---|
PhD_Thesis_Nicola_Ciocchini.pdf
non accessibile
Descrizione: Thesis text
Dimensione
23.64 MB
Formato
Adobe PDF
|
23.64 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/10589/102947