A survey of AI accelerators for edge environment

Abstract

The execution of AI models, especially deep learning models, involves large quantity of processor operations and high memory transfer, which is critical to edge environment. To response to this, a number of specific hardware, a.k.a. edge AI accelerators, have been developed to accelerate the AI model execution in edge. Nevertheless, with no dominant design in the domain, existing AI accelerators hold different physical types, hardware features and software capabilities, and users without prior knowledge can barely choose the proper accelerator corresponding to their computation requirements. This paper aims at providing a systematic survey of AI accelerators for edge environment: firstly we briefly introduce the AI computation requirements and AI accelerator history; we then present a list of criteria covering both hardware and software aspects; following the criteria, we survey the emerging and representative AI accelerator products. At last, we summary the current trends and shed light on future directions for AI accelerator design.