Temporal signal processing with memristor-based in-memory computing systems

Abstract

The execution of increasingly complex artificial intelligence (AI) tasks on edge devices, which are inherently resource-constrained, necessitates exceptional energy efficiency in edge AI hardware. Contemporary AI hardware predominantly expends energy during data transfer between discrete computing and memory units. An emerging technology that facilitates alternative computational paradigms—such as neuromorphic and in-memory computing—is resistive random-access memory (RRAM). RRAM has garnered significant research interest due to its nanoscale dimensions, low power consumption, and intrinsic capabilities that emulate biological synapses and stochastic processes. Furthermore, in-memory computing (IMC) represents a particularly promising architectural approach to address these power consumption challenges in situ. IMC architectures that integrate RRAM with a CMOS logic platform offer a viable solution. These architectures conserve energy by storing the weights of AI models in dense, analogue RRAM cells and performing AI-related computations internally within the RRAM itself. This approach effectively obviates the need for explicit memory access to retrieve weights, thereby improving the system's energy efficiency. In this thesis, I will discuss the fundamental operation and measurement of RRAM devices, as well as the application of resistive memory arrays for time-series tasks such as classification and prediction. First, we introduce an RRAM programming system capable of mapping software-trained weights to RRAM arrays with high programming accuracy, thereby preparing for subsequent simulation calculations. Secondly, we introduce a novel convolutional echo-state network designed for a hybrid analogue-digital architecture that addresses the challenges posed by two specific edge-based spatiotemporal data classification tasks. From a hardware perspective, the system leverages the inherently stochastic programming attributes of RRAM to generate the random weight matrices crucial for the operation of deep learning frameworks. This approach facilitates energy-efficient computation directly within the memory substrate. On the software front, the integration of echo-state networks with randomised convolutional-pooling structures enables the simultaneous capture of both temporal dynamics and channel-wise dependencies within the input data, utilising the random weights provided by the physical configuration of the RRAM array. Thirdly, we propose an energy-efficient, RRAM-based in-memory solver for neural ordinary differential equations (ODEs), specifically designed for a fully analogue system to address two non-dynamic system prediction tasks. This novel system employs a fully analogue, infinite-depth, closed-loop architecture based on ODE principles, enabling continuous-time computation. This fully analogue methodology fosters a dynamic, self-evolving system with inherent memory capabilities, where massively parallel analogue computations are executed within a densely packed crossbar array. The process occurs in continuous time, significantly reducing the need for frequent analogue-to-digital conversions.