As neural networks and Artificial Intelligence (AI) models continue to grow in size and complexity, their execution is increasingly shifting toward cloud infrastructures, leading to high costs and environmental impact. This centralized approach is not well-suited to Edge AI, where computation must be performed locally under strict energy constraints. Neuromorphic engineering investigates alternative paradigms based on in-memory computing, event-driven communication, and analog/mixed-signal processing. Ferroelectric devices are promising candidates for non-volatile synaptic weight storage, but their integration with Complementary Metal–Oxide–Semiconductor (CMOS) circuits requires dedicated interfaces to translate their programmed state into controlled synaptic currents. This thesis investigates analog subthreshold and mixed-signal CMOS circuits for interfacing Ferroelectric Tunnel Junctions (FTJs) and Ferroelectric Capacitors (FeCaps) synaptic devices with a tunable and low-power neuron architecture, suitable for custom spiking-neural-network-based Edge AI systems. CMOS–FTJ synapses were simulated over a resistance range from 10 MΩ to 1 GΩ, while limiting the device voltage below 100 mV for safe readout. Non-regulated and regulated interfaces were analyzed, generating synaptic currents in the pA–nA range. CMOS–FeCap synapses demonstrated analog weight representation, although bottom-electrode voltage drift remained a key limitation for long-term analog storage. Variability and process corner analyses were used to guide transistor sizing and assess circuit robustness. The neuron circuit implements tunable leaky integrate-and-fire behavior, including adjustable gain, leakage, threshold, output spike width, and refractory period. A current mode comparator reduces the static current of the threshold-detection stage, lowering the energy per spike down to 100 pJ at a firing frequency of 300 Hz. The proposed synaptic and neuronal building blocks provide a foundation for in-memory spiking neural-network chips integrating ferroelectric devices with CMOS technology.
Con la crescita in dimensione e complessità delle reti neurali e dei modelli di Intelligenza Artificiale (AI), la loro esecuzione si sta spostando sempre più verso infrastrutture cloud, con conseguenti costi elevati e impatto ambientale. Questo approccio centralizzato non è adatto all’Edge AI, dove il calcolo deve avvenire localmente e sotto stringenti vincoli energetici. L’ingegneria neuromorfica studia paradigmi alternativi basati su in-memory computing, comunicazione event-driven ed elaborazione analogica/mixed-signal. I dispositivi ferroelettrici sono candidati promettenti per la memorizzazione non volatile dei pesi sinaptici, ma la loro integrazione con circuiti CMOS (Complementary Metal–Oxide– Semiconductor) richiede interfacce in grado di tradurre il loro stato in correnti sinaptiche. Questa tesi studia circuiti CMOS analogici sottosoglia e mixed-signal per interfacciare dispositivi sinaptici basati su Ferroelectric Tunnel Junctions (FTJ) e Ferroelectric Capacitors (FeCap) con un’architettura neuronale regolabile e a bassa potenza, destinata a sistemi Edge AI custom basati su reti neurali spiking. Sinapsi CMOS–FTJ sono state simulate su resistenze da 10 MΩ a 1 GΩ, mantenendo la tensione sotto 100 mV per una lettura sicura. Interfacce regolate e non regolate hanno generato correnti sinaptiche nell’intervallo pA–nA. Le sinapsi CMOS–FeCap hanno dimostrato di poter rappresentare pesi analogici, sebbene il drift della tensione all’elettrodo inferiore resti un limite per la memorizzazione a lungo termine. Analisi di variabilità e corner di processo hanno guidato il dimensionamento dei transistor e valutato la robustezza. Il circuito neuronale implementa un comportamento leaky integrate-and-fire, includendo guadagno, leakage, soglia, larghezza dello spike in uscita e periodo refrattario configurabili. Un comparatore in corrente riduce la corrente statica dello stadio di rilevamento della soglia, portando l’energia per spike a 100 pJ a una frequenza di firing di 300 Hz. I blocchi sinaptici e neuronali proposti forniscono una base per chip in-memory basati su reti neurali spiking, integrando dispositivi ferroelettrici con tecnologia CMOS.
Mixed-signal Cmos design with ferroelectrics for energy effiecient neuromorphic systems
Mazza, Matteo
2025/2026
Abstract
As neural networks and Artificial Intelligence (AI) models continue to grow in size and complexity, their execution is increasingly shifting toward cloud infrastructures, leading to high costs and environmental impact. This centralized approach is not well-suited to Edge AI, where computation must be performed locally under strict energy constraints. Neuromorphic engineering investigates alternative paradigms based on in-memory computing, event-driven communication, and analog/mixed-signal processing. Ferroelectric devices are promising candidates for non-volatile synaptic weight storage, but their integration with Complementary Metal–Oxide–Semiconductor (CMOS) circuits requires dedicated interfaces to translate their programmed state into controlled synaptic currents. This thesis investigates analog subthreshold and mixed-signal CMOS circuits for interfacing Ferroelectric Tunnel Junctions (FTJs) and Ferroelectric Capacitors (FeCaps) synaptic devices with a tunable and low-power neuron architecture, suitable for custom spiking-neural-network-based Edge AI systems. CMOS–FTJ synapses were simulated over a resistance range from 10 MΩ to 1 GΩ, while limiting the device voltage below 100 mV for safe readout. Non-regulated and regulated interfaces were analyzed, generating synaptic currents in the pA–nA range. CMOS–FeCap synapses demonstrated analog weight representation, although bottom-electrode voltage drift remained a key limitation for long-term analog storage. Variability and process corner analyses were used to guide transistor sizing and assess circuit robustness. The neuron circuit implements tunable leaky integrate-and-fire behavior, including adjustable gain, leakage, threshold, output spike width, and refractory period. A current mode comparator reduces the static current of the threshold-detection stage, lowering the energy per spike down to 100 pJ at a firing frequency of 300 Hz. The proposed synaptic and neuronal building blocks provide a foundation for in-memory spiking neural-network chips integrating ferroelectric devices with CMOS technology.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Mazza_Tesi.pdf
accessibile in internet per tutti
Descrizione: Testo tesi
Dimensione
17.99 MB
Formato
Adobe PDF
|
17.99 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Mazza_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
1.89 MB
Formato
Adobe PDF
|
1.89 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/10589/261555