Computational intelligence enabled hardware software co design for energy efficient Edge Artificial Intelligence

Abstract Edge Artificial Intelligence (AI) facilitates efficient and high-speed inference by making use of its ability to process data in-device without depending on an external server. Despite its promising possibilities, its deployment still faces many obstacles due to limited computing capacity, memory size, power consumption, and the variety of hardware platforms used at the edge. Thus, this article provides structured narrative review of the latest breakthroughs achieved in the development of energy-efficient Edge AI, focusing specifically on hardware-software co-design practices aimed at improving inference quality. The analysis involves studying state-of-the-art Edge AI hardware infrastructures based on microcontrollers, FPGAs, GPUs, NPUs, ASICs, and some new architectures that utilize in-memory computing, as well as applying various techniques for improving software performance such as model compression, quantization, pruning, neuromorphic computing and techniques applied during compiling, executing, and allocating workloads. Moreover, the article provides comparative evaluation of several hardware platforms, frameworks, scheduling technologies and optimization systems on the basis of their latency, energy-saving capabilities, potential for implementation and complexity. The analysis demonstrates that hardware-aware model optimization and heterogeneous computing significantly improve performance while reducing energy consumption and memory requirements. Finally, the paper discusses current research challenges, including model complexity, real-time execution, security, privacy, and design-space exploration, and outlines future research directions involving AI-driven co-design, automated optimization, and emerging in-memory and neuromorphic computing architectures for sustainable edge intelligence.