Analogue In-Memory Computing with Resistive Switching Memories

In the era of pervasive artificial intelligence (AI) and internet of things (IoT), achieving a high energy efficiency is at the top of priority for computing systems. In this scenario, in-memory computing is gaining momentum as a new methodology to overcome the von Neumann architecture and the related memory bottleneck. One of the most promising device for in-memory computation is the resistive switching memory (RRAM), also known as memristor, thanks to controllable conductance, good scaling and relatively low energy consumption. However, to achieve the promised benefits of in-memory computing with RRAM in terms of performance and power consumption, it is necessary to address a number of open challenges at the device, architecture and algorithm levels. This chapter presents the status of in-memory computing with RRAM, including the device concept and characteristics, the computing architectures and the applications. The perspective of analogue computing is analyzed with reference to both matrix vector multiplication (MVM) and inverse MVM to accelerate linear algebra problems that are generally executed with iteration schemes, highlighting the advantages in terms of performance, energy consumption and computational complexity.