The performance of computing systems has been increased exponentially for the last 5 decades, thanks to advances in microelectronic technology and architecture design. This was predicted in 1965 by Gordon Moore, who stated that the number of transistors on the chip should double every two years, thanks to the reduction of transistor area, which also enhances the operating frequency. Unfortunately, Moore's law is slowing down and will reach an end in the near future. One of the major scaling issues is the memory wall of von Neumann architectures, where computing units and memory are physically separated, therefore causing most of energy and computational time being spent in transferring the data between computing and memory units. New architectures that merge memory and computing are therefore highly desired. To face these challenges, emerging memory devices, such as resistive random access memory (RRAM), phase change memory (PCM) and spin-transfer-torque magnetic random access memory (STT-MRAM) have recently gained significant interest for their non-volatility, scalability, low current operation and compatibility with complementary metal-oxide-semiconductor (CMOS) process. Moreover, novel approaches aiming to radically subvert Von Neumann architecture blurring the distinction between computation and memory have also been subject of intensive research. Among these novel approaches, neuromorphic computing has rapidly attracted considerable attention for its ambitious objective to emulate brain ability to carry out extremely complex cognitive functions such as learning, recognition, inference, and decision making with an unrivaled energy efficiency due to spike-based information processing. Phase change memory (PCM) is one of major class of emerging memory that has attracted huge attention. The peculiar property of being able to emulate both synapse and neuron behavior makes them a suitable choice for designing and implementing brain-inspired system, thus encourages the researcher to investigate this field extensively. Moreover, the scalability, high switching speed and being non-volatile are make them more special for the further design. This master dissertation covers designing a memristor-based neuromorphic system based on the phase change memory (PCM) act as synapse, which is running a Spiking recurrent neural network (Spiking hopfield network). The system is used first, to elaborate the associative property of the hopfield network second, to show the capability of the system to solve Constraint satisfaction problem (CSP) such as Max-Cut, Graph coloring and more important one the Sudoku were chosen for this aim.
Le prestazioni dei sistemi informatici sono aumentate esponenzialmente negli ultimi 5 decenni, grazie ai progressi nella tecnologia microelettronica e nella progettazione dell'architettura. Questo è stato previsto nel 1965 da Gordon Moore, che ha dichiarato che il numero di transistor sul chip dovrebbe raddoppiare ogni due anni, grazie alla riduzione dell'area del transistor, che migliora anche la frequenza operativa. Sfortunatamente, la legge di Moore sta rallentando e finirà nel prossimo futuro. Uno dei principali problemi di ridimensionamento è la parete di memoria delle architetture di von Neumann, in cui le unità di calcolo e la memoria sono fisicamente separate, causando quindi la maggior parte dell'energia e del tempo di calcolo trascorsi nel trasferimento dei dati tra il calcolo e le unità di memoria. Le nuove architetture che uniscono memoria e informatica sono quindi altamente desiderate. Per affrontare queste sfide, i dispositivi di memoria emergenti, come la memoria ad accesso casuale resistivo (RRAM), la memoria a cambiamento di fase (PCM) e la memoria di accesso casuale magnetica a trasferimento di coppia (STT-MRAM) hanno recentemente guadagnato interesse significativo per la loro non volatilità , scalabilità, funzionamento a bassa corrente e compatibilità con il processo complementare di metallo-ossido-semiconduttore (CMOS). Inoltre, nuovi approcci che mirano a sovvertire radicalmente l'architettura di Von Neumann offuscando la distinzione tra calcolo e memoria sono stati anche oggetto di ricerca intensiva. Tra questi nuovi approcci, il calcolo neuromorfico ha rapidamente attirato notevole attenzione per il suo ambizioso obiettivo di emulare la capacità del cervello di svolgere funzioni cognitive estremamente complesse come l'apprendimento, il riconoscimento, l'inferenza e il processo decisionale con un'efficienza energetica senza rivali grazie all'elaborazione delle informazioni basata sugli spike . La memoria a cambiamento di fase (PCM) è una delle principali classi di memoria emergente che ha attirato grande attenzione. La proprietà peculiare di essere in grado di emulare sia il comportamento di sinapsi che di neuroni li rende una scelta adatta per progettare e implementare sistemi ispirati al cervello, incoraggiando così il ricercatore a indagare su questo campo in modo estensivo. Inoltre, la scalabilità, l'elevata velocità di commutazione e l'essere non volatile li rendono più speciali per l'ulteriore progettazione. Questa tesi di master copre la progettazione di un sistema neuromorfico basato su memristor basato sulla memoria di cambiamento di fase (PCM) agisce come sinapsi, che sta eseguendo una rete neurale ricorrente Spiking (rete Spiking hopfield). Il sistema viene utilizzato per primo, per elaborare la proprietà associativa della rete hopfield in secondo luogo, per mostrare la capacità del sistema di risolvere il problema di Costo del vincolo (CSP) come Max-Cut, colorazione del grafico e più importante il Sudoku è stato scelto per questo scopo.
Development and experimental demonstration of spiking neural network with phase change memory (PCM) for constraint satisfaction problem (CSP)
HASHEMKHANI, SHAHIN
2018/2019
Abstract
The performance of computing systems has been increased exponentially for the last 5 decades, thanks to advances in microelectronic technology and architecture design. This was predicted in 1965 by Gordon Moore, who stated that the number of transistors on the chip should double every two years, thanks to the reduction of transistor area, which also enhances the operating frequency. Unfortunately, Moore's law is slowing down and will reach an end in the near future. One of the major scaling issues is the memory wall of von Neumann architectures, where computing units and memory are physically separated, therefore causing most of energy and computational time being spent in transferring the data between computing and memory units. New architectures that merge memory and computing are therefore highly desired. To face these challenges, emerging memory devices, such as resistive random access memory (RRAM), phase change memory (PCM) and spin-transfer-torque magnetic random access memory (STT-MRAM) have recently gained significant interest for their non-volatility, scalability, low current operation and compatibility with complementary metal-oxide-semiconductor (CMOS) process. Moreover, novel approaches aiming to radically subvert Von Neumann architecture blurring the distinction between computation and memory have also been subject of intensive research. Among these novel approaches, neuromorphic computing has rapidly attracted considerable attention for its ambitious objective to emulate brain ability to carry out extremely complex cognitive functions such as learning, recognition, inference, and decision making with an unrivaled energy efficiency due to spike-based information processing. Phase change memory (PCM) is one of major class of emerging memory that has attracted huge attention. The peculiar property of being able to emulate both synapse and neuron behavior makes them a suitable choice for designing and implementing brain-inspired system, thus encourages the researcher to investigate this field extensively. Moreover, the scalability, high switching speed and being non-volatile are make them more special for the further design. This master dissertation covers designing a memristor-based neuromorphic system based on the phase change memory (PCM) act as synapse, which is running a Spiking recurrent neural network (Spiking hopfield network). The system is used first, to elaborate the associative property of the hopfield network second, to show the capability of the system to solve Constraint satisfaction problem (CSP) such as Max-Cut, Graph coloring and more important one the Sudoku were chosen for this aim.| File | Dimensione | Formato | |
|---|---|---|---|
|
874574-Thesis.pdf
non accessibile
Descrizione: Thesis text
Dimensione
40.5 MB
Formato
Adobe PDF
|
40.5 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/10589/148589