Energy Efficiency in AI: Powering a Sustainable Future from Edge to Cloud

Latest 50 papers on energy efficiency: Nov. 30, 2025

The relentless march of AI and ML innovation has brought unprecedented capabilities, but it’s also ushered in a growing challenge: energy consumption. From training colossal Large Language Models to deploying intelligent systems on billions of edge devices, the demand for computational power translates directly into significant energy footprints. This isn’t just an environmental concern; it’s an economic and practical one, impacting the scalability and accessibility of AI. Fortunately, a flurry of recent research is tackling this head-on, exploring groundbreaking approaches to make AI more sustainable, efficient, and accessible.

The Big Idea(s) & Core Innovations

At the heart of these advancements lies a common goal: to squeeze more computational value from less energy. Researchers are pushing the boundaries across the entire AI stack, from fundamental algorithms to specialized hardware. For instance, the paper “Quartet: Native FP4 Training Can Be Optimal for Large Language Models” by Roberto L. Castro and colleagues from ISTA and ETH Zürich demonstrates that training LLMs with native FP4 precision can be as accurate as higher-precision methods while being significantly more efficient. This challenges the long-held assumption that high accuracy requires high precision.

Another significant thrust focuses on in-memory computing and neuromorphic architectures. The work on “Compute-in-Memory Implementation of State Space Models for Event Sequence Processing” by Xiaoyu Zhang, Mingtao Hu, and others from the University of Michigan pioneers an energy-efficient hardware implementation of State Space Models (SSMs) using memristors. Their approach leverages the inherent physics of memristors for real-time, event-driven processing, dramatically reducing FLOPs (62x to 131x) compared to traditional CNNs. Similarly, “NL-DPE: An Analog In-memory Non-Linear Dot Product Engine for Efficient CNN and LLM Inference” by Authors A, B, and C (affiliations Institution X, Y, Z) introduces an analog in-memory dot product engine that promises substantial energy and latency reductions for CNN and LLM inference.

On the software and algorithmic front, “FAST: Topology-Aware Frequency-Domain Distribution Matching for Coreset Selection” by Jin Cui, Boran Zhao, and others from Xi’an Jiaotong University presents a DNN-free coreset selection framework that reduces power consumption by an astounding 96.57% and offers a 2.2× speedup on CPU, all while preserving fine-grained semantic information. This is particularly impactful for applications like LLM instruction tuning. In a similar vein, “Temporal-adaptive Weight Quantization for Spiking Neural Networks” by Zhang Han, Meng Qingyan, and Ma Zhengyu from Pengcheng Laboratory, introduces TaWQ, a method that dynamically adapts weight quantization in Spiking Neural Networks (SNNs), leading to improved performance and energy efficiency for neuromorphic computing.

Edge computing is another crucial battleground for energy efficiency. The paper “Characterizing and Understanding Energy Footprint and Efficiency of Small Language Model on Edges” by Choi and colleagues from NVIDIA and the University of Tokyo emphasizes the need to optimize inference pipelines for small language models on edge devices. Meanwhile, “TT-Edge: A Hardware-Software Co-Design for Energy-Efficient Tensor-Train Decomposition on Edge AI” by P. Narayanan et al. from NCSU and Synopsys introduces a co-designed framework that optimizes both algorithms and hardware for significant energy savings in tensor operations on edge devices. For IoT, “Edge-Based Predictive Data Reduction for Smart Agriculture: A Lightweight Approach to Efficient IoT Communication” by Fathalla, Li, Salah, and Mohamed demonstrates how lightweight LSTM models at the edge can achieve over 90% data reduction, saving energy and bandwidth in smart agriculture.

Under the Hood: Models, Datasets, & Benchmarks

This wave of research relies on innovative models, rigorous benchmarks, and sometimes, entirely new hardware paradigms:

• Quantized LLM Training: Quartet’s method is implemented with highly optimized CUDA kernels for NVIDIA’s Blackwell architecture, demonstrating that FP4 can be competitive with FP16/FP8. The code is available at https://github.com/IST-DASLab/Quartet.
• State Space Models (SSMs) in CIM: The University of Michigan team re-parameterizes SSMs for real-valued coefficients and leverages WOₓ-based memristors in RRAM CIM chips. They benchmark against Spiking Heidelberg Digits (SHD), Spiking Speech Commands (SSC), DVS128 Gesture, and DVS128 Lips datasets.
• DNN-free Coreset Selection: FAST, developed by Xi’an Jiaotong University researchers, is a novel framework that uses frequency-domain distribution matching and characteristic function distance (CFD) for coreset selection. Its power consumption reduction makes it ideal for LLM instruction tuning.
• SNN Quantization: TaWQ, by Pengcheng Laboratory, introduces a temporal-adaptive weight quantization for SNNs, outperforming static methods. Its code is open-sourced at https://github.com/ZhangHanN1/TaWQ.
• Edge AI Hardware: NX-CGRA (“NX-CGRA: A Programmable Hardware Accelerator for Core Transformer Algorithms on Edge Devices”) by J. Devlin et al. from EPFL, and FERMI-ML (“FERMI-ML: A Flexible and Resource-Efficient Memory-In-Situ SRAM Macro for TinyML acceleration”) by Sandeep Kumar and colleagues, represent specialized hardware designs for energy-efficient Transformer and TinyML workloads on edge devices, leveraging SRAM macros.
• Sustainable Scheduling Algorithms: Papers like “Developing an Algorithm Selector for Green Configuration in Scheduling Problems” and “Instance Configuration for Sustainable Job Shop Scheduling” by Carlos March, Christian Pérez, and Miguel Salido from Universitat Politècnica de València, contribute machine learning-based algorithm selectors (e.g., XGBoost) and a comprehensive set of 500 public test instances to benchmark energy-efficient solutions for Job Shop Scheduling Problems (JSP). Code for the algorithm selector is at https://github.com/carlosmarch/AlgorithmSelectorForGreenConfiguration.
• Federated Learning for Green AI: FairEnergy (“FairEnergy: Contribution-Based Fairness meets Energy Efficiency in Federated Learning”) by Xiaowei Chen et al. from the University of Toronto, provides a framework and code (https://github.com/FairEnergy-FL/FairEnergy) to balance fairness and energy efficiency in federated learning.

Impact & The Road Ahead

The implications of this research are profound. We’re seeing a paradigm shift where energy efficiency is no longer an afterthought but a core design principle across AI development. From optimizing training of massive LLMs to making tiny ML models viable on resource-constrained devices, these advancements pave the way for a greener, more accessible AI future.

• Sustainable AI: The explicit focus on carbon-aware and energy-efficient systems, whether in smart ports (“Assessing the Technical and Environmental Impacts of Energy Management Systems in Smart Ports”) or microservices (“On the Effectiveness of Microservices Tactics and Patterns to Reduce Energy Consumption”), directly addresses AI’s environmental impact.
• Democratized AI: Efficient edge and federated learning make powerful AI accessible on mobile devices, IoT sensors, and distributed networks, reducing reliance on centralized, energy-hungry data centers.
• Robust & Resilient Systems: Innovations like opportunistic DTN for disaster relief (“Improving Resiliency of Vital Services in Flood-Affected Regions of Bangladesh Using Next-Generation Opportunistic DTN Edge Ad Hoc Networks”) and AI-enhanced microgrids (“AI-Enhanced IoT Systems for Predictive Maintenance and Affordability Optimization in Smart Microgrids”) highlight AI’s role in building more robust societal infrastructure.
• Future Networks (6G): Papers like “Toward hyper-adaptive AI-enabled 6G networks for energy efficiency” and “Agentic AI-Empowered Conversational Embodied Intelligence Networks in 6G” show AI as a foundational element for sustainable and intelligent next-generation communication systems, enabling dynamic resource allocation and real-time decision-making.

The road ahead involves continued exploration of algorithm-architecture co-design, novel materials (like memristors), and biologically inspired computing (SNNs). As “Analog Physical Systems Can Exhibit Double Descent” from the University of Pennsylvania demonstrates, even the fundamental physics of computation holds untapped potential. By embracing these interdisciplinary approaches, we can ensure AI’s powerful capabilities are harnessed responsibly, paving the way for a truly sustainable and intelligent world.