Interconnect and Memory Design for Intelligent Mobile System

Technology scaling has driven the transistor to a smaller area, higher performance and lower power consuming which leads us into the mobile and edge computing era. However, the benefits of technology scaling are diminishing today, as the wire delay and energy scales far behind that of the logics, which makes communication more expensive than computation. Moreover, emerging data centric algorithms like deep learning have a growing demand on SRAM capacity and bandwidth. High access energy and huge leakage of the large on-chip SRAM have become the main limiter of realizing an energy efficient low power smart sensor platform. This thesis presents several architecture and circuit solutions to enable intelligent mobile systems, including voltage scalable interconnect scheme, Compute-In-Memory (CIM), low power memory system from edge deep learning processor and an ultra-low leakage stacked voltage domain SRAM for low power smart image signal processor (ISP). Four prototypes are implemented for demonstration and verification. The first two seek the solutions to the slow and high energy global on-chip interconnect: the first prototype proposes a reconfigurable self-timed regenerator based global interconnect scheme to achieve higher performance and energy-efficiency in wide voltage range, while the second one presents a non Von Neumann architecture, a hybrid in-/near-memory Compute SRAM (CRAM), to address the locality issue. The next two works focus on low-power low-leakage SRAM design for Intelligent sensors. The third prototype is a low power memory design for a deep learning processor with 270KB custom SRAM and Non-Uniform Memory Access architecture. The fourth prototype is an ultra-low leakage SRAM for motion-triggered low power smart imager sensor system with voltage domain stacking and a novel array swapping mechanism. The work presented in this dissertation exploits various optimizations in both architecture level (exploiting temporal and spatial locality) and circuit customization to overcome the main challenges in making extremely energy-efficient battery-powered intelligent mobile devices. The impact of the work is significant in the era of Internet-of-Things (IoT) and the age of AI when the mobile computing systems get ubiquitous, intelligent and longer battery life, powered by these proposed solutions.PhDElectrical and Computer EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/155232/1/jiwang_1.pd