Efficient and Robust Compute-In-Memory for Edge Intelligence

Abstract

With deep neural networks (DNNs) fueling major advancements in artificial intelligence (AI), the demand for efficient computing solutions has never been higher. This dissertation delves into the challenges and innovations in the field of compute-in-memory (CIM) technologies, a key area for advancing sustainable computing to address the growing carbon footprint associated with intensive AI workloads. CIM is an emerging computing paradigm aimed at processing data within memory arrays where data is stored, promising significant gains in energy efficiency and computational speed. This dissertation focuses on building efficient, robust, and heterogeneous CIM solutions for edge intelligence. The dissertation first presents the prototype chip development of CIM primitives as AI inference engine, demonstrating the feasibility of analog computations within resistive random-access memory (RRAM). Additional design challenges are addressed through novel circuit-level design techniques including lightweight on-chip write-verify, in-situ error correction, temperature-tracking references, and embedded model encryption. Exploring beyond traditional scaling methods, the dissertation next proposes vertically stacking silicon dies into a heterogeneous 3-D (H3D) system for the flexibility to combine different process nodes and high-bandwidth data transmission. A H3D integrated accelerator is designed to target vision transformer models through a hybrid analog and digital CIM approach. Thorough thermal and signaling evaluations are conducted to understand the trade-offs of die stacking. Finally, domain-specific CIM architectures for edge computing are investigated, focusing on integrating CIM hardware with sensor frontends for intelligent data volume reduction. An algorithm/hardware co-design approach is proposed to reduce power consumption of portable medical ultrasound imaging and to improve communication bandwidth for autonomous driving, showcasing the potential of CIM to efficiently insert local intelligence in diverse applications.