Deep neural networks (DNNs) have driven significant breakthroughs but also raised concerns due to their increasing computational and energy demands. With the limitations of Moore’s Law and Dennard scaling, energy-efficient solutions using emerging hardware, approximate computing, and in-memory techniques are crucial for AI systems. This research focuses on hardware-software co-design strategies to improve AI efficiency across platforms like data centers, edge, and embedded devices. Key areas include optimizing training and inference algorithms for large models, model quantization, compression techniques, distributed training, and hardware-aware neural architecture. Additionally, the exploration of analog AI accelerators aims to enhance efficiency, addressing challenges like precision and integration while focusing on energy performance and model accuracy