New survey maps path to green AI: resource-efficient architectures and hardware co-design
Large models' energy crisis meets a comprehensive roadmap for sustainable AI development.
A comprehensive new survey published in the Chinese Journal of Electronics (2026) tackles the growing environmental and computational costs of large AI models. The paper, authored by Linhui Xiao and eight colleagues, systematically reviews resource-efficient architectures and full-stack hardware-software co-design for green AI. On the software side, it covers attention operator optimization (e.g., linear-complexity architectures like Mamba), model sparsification and merging, data-efficient learning, parameter-efficient fine-tuning, and computational compression techniques. These approaches aim to reduce the massive energy consumption of training and inference without sacrificing performance.
On the hardware front, the survey explores energy-efficient AI chips, memory optimization strategies, cross-platform deployment, and sustainable infrastructure. It also examines how large models are being applied to sustainability-critical domains such as DeepSeek, remote sensing interpretation, national-scale infrastructure, and global initiatives. The authors highlight key challenges ahead, including the need for continual learning paradigms, memory-centric hardware design, and standardized evaluation protocols. This survey provides a holistic roadmap for developing large models that are both powerful and environmentally responsible.
- Reviews efficient architectures including attention optimization, linear-complexity models (e.g., Mamba), and model sparsification/merging.
- Covers training efficiencies like data-efficient learning, parameter-efficient fine-tuning, and computational compression.
- Examines energy-efficient hardware (AI chips, memory optimization) and real-world green AI applications in DeepSeek and remote sensing.
Why It Matters
Provides a structured roadmap for reducing AI's environmental footprint while maintaining performance, guiding researchers and industry toward sustainable scaling.