A Survey on the Green Development of Large Models: From Resource-Efficient Architectures to Hardware-Software Co-Design

Jul 16, 2026

Upload by Skyler

The document surveys the green development of large AI models amid concerns about rising computational costs, energy consumption, and environmental sustainability. It emphasizes resource-efficient model architectures and full-stack hardware-software co-design. The review covers advances in efficient construction—attention operator optimization, linear-complexity designs, and model sparsification/merging—plus training and deployment approaches such as data-efficient learning, parameter-efficient fine-tuning, and computational compression. It also evaluates energy-efficient AI hardware including mainstream chips, memory optimization, cross-platform deployment, and sustainable infrastructure, then discusses sustainability-oriented applications and future challenges like continual learning, memory-centric hardware, and standardized evaluation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

Introduction

Resource and energy challenges of large models

Survey scope and gap versus prior work

arXiv :2607 .09084v 1 [ cs .LG] 10 Jul 2026 Chinese Journal of Electronics vol.35, no.5, pp.1– 24, 2026 [https://doi.org/10.23919/cje.2025.00.438](https://doi.org/10.23919/cje.2025.00.438) Review A Survey on the Green Development of Large Models: From Resource-Efficient Architectures to Hardware-Software Co-Design Linhui Xiao 1, Guiping Cao 1, Mingyue Guo 1, Xianchao Guan 1,2, Fan Yang 1, Ming Tao 1 , Xin Li 1,*, Yuxin Peng3, Yaowei Wang2,1,* 1. Pengcheng Laboratory, Shenzhen 518066, China 2. Harbin Institute of Technology (Shenzhen), Shenzhen 518055, China 3. Peking University, Beijing 100080, China Corresponding author: Xin Li, Yaowei Wang; Email: [xinlihitsz@gmail.com](xinlihitsz@gmail.com), [wangyw@pcl.ac.cn](wangyw@pcl.ac.cn). Manuscript Received September 30, 2025; Accepted January 13, 2026; Published Online February 7, 2026 . Abstract— The rapid expansion of large-scale AI models has led to significant performance breakthroughs across diverse domains, yet it has also raised critical concerns regarding computational costs, energy consumption, and environmental sustainability. This survey provides a comprehensive overview of the green development of large models, emphasizing resource-efficient architectures and full-stack hardware-software co-design. We systematically review recent advances in efficient model construction, including attention operator optimization, linear-complexity architectures, and modelsparsification and merging, as well as training and deployment strategies such as data-efficient learning, parameter-efficient fine-tuning, and computational compression. Beyond algorithmic improvements, we explore energy-efficient AI hardware, including mainstream AI chips, memory optimization, cross-platform deployment, and sustainable infrastructure. Furthermore, we examine how large models are being applied to sustainability-critical domains such as DeepSeek, remote sensing interpretation, national-scale infrastructure, and global initiatives. Finally, we discuss key challengesand future directions, highlighting the need for continual learning paradigms, memory-centric hardware, and standardized evaluation protocols. This survey aims to offer a holistic roadmap toward sustainable, scalable, and socially responsible development of large models. Keywords—Green AI, Model Efficiency, Hardware-Software Co-Design, Sustainable Computing, Large Models I. Introduction In recent years, large-scale Artificial Intelligence (AI) models, especially those built upon Transformer architectures [1], have achieved significant breakthroughs in natural language processing, computer vision, and scientific computing. Flagship models such as BERT [2], CLIP [3], LLaVA [4], and the GPT series [5] have continuously expanded the frontier of AI. However, this progress has come with substantial costs: the exponential increase in model parameters, training data, and input length has led to skyrocketing computational demands, resulting in massive energy consumption and environmental concerns, as shown in Table 1 . As model parameters scale from billions to trillions and input sequences extend from 1K to 100K tokens, training and inference costs have increased superlinearly, posing major challenges to the accessibility, scalability, and sustainability of large-scale AI systems.* ∗[https://cje.ejournal.org.cn/article/doi/10.23919/cje.2025.00.438](https://cje.ejournal.org.cn/article/doi/10.23919/cje.2025.00.438) Efficient Figure 1 A triangular framework illustrating the layered relationship among model architecture, training strategies, and green AI hardware. The core of this inefficiency lies in the architectural de- Table 1 Comparison of GPU hours, Energy Usage, and Carbon Footprint for training mainstream large models. † indicates no official data available, and the values are estimated and for reference only. | Model | GPU hours<br>(h) | Eneregy Usage<br>(KWh) | Carbon Emitted<br>(tCO2eq) | | --- | --- | -

FAQs

What main problem does the survey address for large models?

The survey targets the sustainability risks created by rapid scaling of parameters and context length, which drive superlinear training/inference costs, energy use, and environmental impact.

Which core approaches does the survey cover for green large-model development?

It covers resource-efficient architectures (e.g., attention optimization, linear-complexity designs, sparsification/merging) and full-stack hardware-software co-design, including efficient training/deployment techniques and energy-efficient hardware.

How does hardware-software co-design contribute to greener large models?

By aligning model computation strategies with hardware characteristics, the survey highlights improvements such as energy-efficient AI chips, memory optimization, and cross-platform deployment to reduce overall energy demand.

Jul 16, 2026

Upload by Skyler