Date of Award

Summer 7-2015

Embargo Period

8-15-2017

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

Advisor(s)

Larry Pileggi

Abstract

As the fraction of integrated circuit area dedicated to embedded memory continues to increase, the energy spent for transporting data on-chip becomes increasingly larger than the energy needed to perform computation, thereby creating system-level challenges for data-intensive application-domains. One effective approach for this challenge is to use Logic-in-Memory (LiM) blocks, whereby custom application-specific logic is embedded within the memory block to significantly improve the system’s energy, performance, and area efficiency. Recent studies on technology scaling below the 20nm node demonstrate that extremely restrictive patterning enables the automated synthesis of LiM systems by the use of compatible logic and memory patterns. While in-memory processing architectures have been proposed and successfully built for various applications, this dissertation aims to exploit the extremely restrictive patterning to create an end-to-end design methodology for automated synthesis of application-specific LiM blocks. The LiM synthesis methodology eliminates all the full-custom design effort that is inherently needed to build a LiM block, thereby enabling the co-design of algorithms and hardware at an affordable design cost that would be otherwise impractical. Silicon results for two data-intensive applications demonstrate that systems based on synthesized LiM designs can provide dramatic improvements of one to two orders of magnitude of energy and performance efficiency. This methodology further enables rapid design-space exploration for the overall LiM based system by making customization efficient and robust with no extra design-cost. This dissertation attempts to formulate, implement, and validate a novel design methodology that provides automated synthesis of application-specific LiM blocks.

Share

COinS