Computing with nonvolatile memories for artificial intelligence

In-memory computing shows the key advantages with respect to conventional digital computers, thanks to the non-von Neumann architecture which allows to carry out computation directly within the memory, without any data transfer. While various resistive memories with two or three terminals have been proposed for computation, there are still several concerns in terms of precise programming characteristics, immunity to parasitic resistance/capacitance, and overall architecture to enable the coexistence of analog computation within digital processors. This chapter addresses the status and challenges of in-memory computing in terms of devices, algorithms, and architectures. The device structures will be reviewed and their organization within the array circuit will be discussed. The programming nonideality which limits the accuracy of computation will be described. Finally, the most recent memory array architectures for computing will be reviewed.