INTELLIGENT EDGE COMPUTING ARCHITECTURES FOR REAL-TIME IOT APPLICATIONS

Abstract

The rapid proliferation of the Internet of Things (IoT) devices and also the emergence of latency-sensitive, mission-critical applications (autonomous vehicles, industrial control, augmented reality, remote healthcare) have well driven a huge paradigm shift from the centralized cloud processing to the context of distributed, intelligence-enabled edge form of computing. In this paper, the intelligent edge computing architectures are aimed at analyzing the performance issues of intelligent edge computing networks which should be satisfied by the real-time IoT applications. The recent innovations in the field of multi-access edge computing (MEC), TinyML and on-device inference, federated and split learning at the edge, hardware accelerators, and orchestration architecture that enables a solution to another robust and low-latency decision making. The methodology in terms of its approach to measuring the performance of realistic network heterogeneity and workload patterns is by looking over the literature in a systematic way and provides an experimental framework by which the trade-offs of latency, throughput, energy and accuracy are modelled in realistic edge architecture. The findings show that the hybrid architectures consisting of local on-devil inference and ultra-low latency, edge-level aggregating model and contextual adaptation and cloud updating coordination model with an optimal trade-off between latency and accuracy and resource consumption in the vast majority of the real-tile IoT functions are evidenced. Among the concepts, we also finish the discussion with the principles of the architectural design, open technical challenges (privacy, heterogeneity, resources management, real-time guarantees), directions are also provided on how to approach futuristic research.