Design and fabrication of neuromorphic devices towards emulating artificial intelligence

Abstract

Artificial Intelligence (AI) aims to simulate human intelligence in machines so that they can think like humans, and perform human-like cognition [1]. Although the present-day supercomputers can perform excellent calculations, they are very slow and inefficient compared to the human brain! [2,3] For example, a supercomputer takes around 500 s to emulate 5 s of human brain activity by consuming ~ MWs of energy [4]. This is much due to the conventional von Neumann computer architecture, where the memory and processing units are physically separate and are connected by limited interconnects known as bus bars [5,6]. During processing, data shuttles between these units, which makes the process slow and inefficient. Besides, limited transistor density in a processor chip also influences the computational ability. Although the advancing lithography technique is endeavoring to miniaturize transistors to the lowest dimension possible [7] to create a high density on board, the computational performance is still under satisfying. As the transistor scaling down approaches the theoretical limit, their packing density in a chip may not hold the famous Moore’s law any further [8].