COMPUTE-IN-MEMORY ON EMERGING MEMORY TECHNOLOGY: FROM DEVICE TO ALGORITHM

Present-day computing systems largely adhere to the von Neumann architecture, which necessitates data transfer from memory to a processing unit. This architecture presents a significant latency issue when accessing data from memory units, which is a primary performance hindrance for various data-intensive applications, particularly in the intersection of big data and AI. Numerous solutions have been suggested to alleviate and overcome this bottleneck. A notable suggestion has been to physically place memory and logic units closer together. While this approach has seen substantial advancements at both the technology and architectural levels, a revolutionary alternative would be to execute arithmetic kernels directly where data is stored, using memory devices. This concept is referred to as compute-in-memory (CIM).In this dissertation, I will begin by presenting the most recent advancements in the CMOS-compatible ferroelectric memory technologies on aluminum nitride platform. Second, I will present a reconfigurable CIM system on field-programmable ferroelectric diodes in a transistor-free architecture, allowing for multiple essential data operations including storage, parallel search, and neural network. Last, I will discuss the conceptualization and demonstration of a programmable parallel search architecture - analog content-addressable memory (ACAM) on complementary Si-CMOS ferroelectric field-effect-transistor memory. The deployment and acceleration of attentional deep neural network and kernel regression on ACAM will also be presented.