Memristive/CMOS Devices for Neuromorphic Applications

Abstract

Neuromorphic computing aims at processing information in a way which is similar to the human brain, therefore featuring the ability to learn, adapt, and generalize with high computing parallelism and energy efficiency. To pursue this ambitious goal, neuromorphic computers must adopt brain-inspired architectures, which radically differ from the von Neumann concept of conventional computing. Neuromorphic computing thus typically relies on neural networks, consisting of neurons and synapses which are ideally compact, scalable, and operated at low voltage and current. Several device technologies have been proposed for neuromorphic computing, including both conventional memory devices based on the complementary metal-oxide semiconductor (CMOS) technology, and alternative memory (or memristor) technologies, where the storage concept is not based on charge, rather it relies on the local modification of the active material. This chapter will review the status of neuromorphic devices including both CMOS-type and memristive devices. First, neuromorphic concepts using floating-gate and static random access memory (SRAM) arrays will be described, followed by memristive synapses and neurons which offer unprecedented density and scalability. The applications and performance of the various neuromorphic computing concepts will be illustrated and discussed.