Edge AI: Embedded Intelligence, a Review on Hardware and Applications

Abstract

Edge AI brings artificial-intelligence inference from the cloud to resource-constrained devices at the network edge, enabling real-time, low-latency, and privacy-preserving decision-making. This review examines edge AI from a hardware-centric and quantitative perspective. We introduce a unified analytical framework that models inference latency, bandwidth, energy, computational complexity, throughput, and model compression, and use it to compare cloud versus edge execution. We benchmark the principal classes of AI hardware accelerators (VPU, GPU, NPU, and TPU) using public performance and efficiency figures, and present an explicit latency- and energy-aware placement algorithm. We further review representative real-world applications, the main deployment challenges, and future directions including federated learning, neuromorphic computing, and 6G-assisted inference.