The computing requirements of large language models (LLMs) is doubling every 2 months. This exponential growth raises challenges in energy production, heat dissipation in data centres and portability of artificial intelligence (AI) devices. The development of new inherently parallel computing architectures such as graphics (GPUs) and tensor (TPUs) processing units to perform AI computations have allowed the increase of the energy efficiency of AI workloads. New in memory computing approaches based on emerging non-volatile memories have been proposed to further reduce the area requirements and energy consumption linked with the computation of AI workloads. Among them, the memristor crossbar array holds the promise of a very high compactness and low energy implementation of matrix vector multiplication, the operation at the heart of most AI algorithms. The memristor concept, a two-terminal circuit element with a variable resistance, can be implemented with different phenomena: phase change, conductive filament formation and breaking, magnetic or ferroelectric switching, etc. This thesis focuses on ferroelectric memristors because of their potential for high endurance and high resistance ratio between their high and low resistance states. While most ferroelectric materials are tricky to integrate with the processes of the semiconductor industry, the large-scale, back-end-of-line (BEOL) compatible growth of the ferroelectric phase of GeTe has been demonstrated. This thesis aims to demonstrate the potential of GeTe based memristors for neu- romorphic computing. First, magnetron-sputtered GeTe is used to fabricate devices exploiting the GeTe/metal interface for the realisation of ferroelectric Schottky diodes (FSDs). No resistive switching was observed for the GeTe/Ti, GeTe/Pt or GeTe/Pd contact, but a high resistance in the pristine state and a resistive switching phenomenon was observed for the GeTe/Mg contact. Despite its specificities as a degenerate semi- conductor and its strong Fermi level pinning, being p-type, the formation of a barrier with GeTe is expected for metals with a low work function, like Mg. Actually, GeTe/Mg devices present several resistive switching phenomena which are shown to strongly depend on the presence of oxygen in the system. While the origin of some of those phenomena is linked with the GeTe/metal interface, the role of ferroelectric polarisation in the observed resistive switching can be neither confirmed nor denied. The integration of memristors in crossbar arrays is very compact. However, this structure is inherently parallel, and the application of a voltage difference across a device results in the application of this voltage on all parallel paths. Without a specific scheme to suppress the unwanted current paths, there is a high risk of changing the state of unselected memristors during the writing step. The benefits that ferroelectric memristors bring for the suppression of sneak paths in memristor crossbar arrays are considered and simulated with SPICE simulations. Finally, memristor crossbar arrays have other applications besides neuromorphic computing. Memristors have been shown to realise Boolean logic in a more compact way than traditional CMOS logic. However, they usually suffer from cycle-to-cycle and device- to-device variations that are too high for the exactness requirements of Boolean logic. The impact of variability of the memristor on the good operation of memristive Boolean logic gates is quantified. Qualitatively, as expected, a higher on/off ratio memristor results in a logic gate that is less susceptible to memristor resistance state or switching voltage variations. Despite the challenges associated with the control of the ferroelectric polarisation in the degenerate semiconductor GeTe, the realisation of ferroelectric based memristors like FSDs or ferroelectric tunnel junctions is very promising for AI workloads, a very hot topic, but also for Boolean applications.

Plus d'information :https://www.spintec.fr/phd-defense-by-maxime-culot-memristive-switching-in-gete-based-ferroelectric-schottky-diodes-for-neuromorphic-computing/
​

Pour suivre la soutenance ​​​en visioconférence : https://univ-grenoble-alpes-fr.zoom.us/j/3241920232​

​ ​ ​​