Computation-in-memory from application-specific to programmable designs based on memristor devices

Abstract

Computation-in-Memory (CIM) is a promising alternative to traditional computing systems where the storage is conceptually separated fromthe computing units. Instead, the CIM paradigm aims to perform the computation where the data resides, alleviating the memory bottleneck and ultimately leading to higher energy efficiency and performance. From the memory technology perspective, memristors, emerging non-volatile memory devices, demonstrate various beneficial characteristics. Although the concept of CIM, in combination with these emerging memory technologies, is in the infancy stage, it shows great potential as a future of computing systems. To further understand and quantify the potential of CIM, more development is required at each abstraction level. In this thesis, we first explore the main potentials for memristor-based computation-in-memory. Then, we study different applications from the CIM perspective to understand different behaviors and patterns of applications and use this knowledge to develop architectural solutions for CIM. Based on that, we study the realization of CIM as a generic and flexible platformat amicro-architecture level.