Energy Efficiency in AI/ML: Powering a Sustainable Future from Edge to HPC

Latest 50 papers on energy efficiency: Dec. 7, 2025

The relentless march of AI and Machine Learning is transforming industries, but this progress comes with a significant and growing carbon footprint. From massive data centers powering Large Language Models (LLMs) to ubiquitous IoT devices at the edge, the energy demands of AI are escalating. Addressing this challenge is not just an environmental imperative but a technological one, driving innovation in hardware, software, and algorithmic design. Recent research, as explored in a fascinating collection of papers, reveals exciting breakthroughs aimed at making AI truly sustainable.

The Big Ideas & Core Innovations: Smarter, Leaner AI for All

The central theme across these papers is a concerted effort to decouple AI performance from its energy cost, often by embedding energy-awareness directly into the system’s design. This manifests in several key areas:

• Hardware-Software Co-design for Edge AI: A novel approach from the University of Example and EdgeTech Inc. in their paper, “Hardware-aware Neural Architecture Search of Early Exiting Networks on Edge Accelerators”, integrates hardware constraints directly into Neural Architecture Search (NAS). By leveraging early-exiting mechanisms, they achieve faster, more energy-efficient inference on edge devices. Similarly, the Nara Institute of Science and Technology (NAIST) demonstrates that non-AI-specialized Coarse-Grained Linear Arrays (CGLAs) can offer substantial energy efficiency gains for LLM inference, up to 44.4x over high-performance GPUs, as detailed in “Efficient Kernel Mapping and Comprehensive System Evaluation of LLM Acceleration on a CGLA”. This is a significant shift, suggesting that data transfer, not just computation, is a major bottleneck.

• Sustainability-Aware LLMs: Large Language Models are notoriously power-hungry. Research from the University of Innsbruck, Austria, and the University of Klagenfurt, Austria tackles this in “Toward Sustainability-Aware LLM Inference on Edge Clusters”. They propose prompt routing strategies on heterogeneous edge clusters, finding an optimal batch size of four for balancing latency, energy, and carbon footprint. Complementing this, William Blaskowicz’s “SLO-aware GPU Frequency Scaling for Energy Efficient LLM Inference Serving” introduces throttLL’eM, a framework dynamically scaling GPU frequencies at the iteration level to cut energy consumption by up to 43.8% while meeting service-level objectives (SLOs). The theme of intelligent resource allocation extends to “Energy-Aware Data-Driven Model Selection in LLM-Orchestrated AI Systems” from the University of Example, advocating for data-driven model selection to reduce computational costs.

• Neuromorphic and Analog Computing: The future of ultra-low-power AI might lie in mimicking the brain. “Efficient Eye-based Emotion Recognition via Neural Architecture Search of Time-to-First-Spike-Coded Spiking Neural Networks” by Shandong University introduces TNAS-ER, an NAS framework for Spiking Neural Networks (SNNs), achieving impressive accuracy with dramatically reduced parameters and operations (1.6x and 5.6x respectively) and ultra-low energy (0.05 J per inference). Furthermore, groundbreaking work from the University of Liège, Belgium, in “A Neuromodulable Current-Mode Silicon Neuron for Robust and Adaptive Neuromorphic Systems” presents a current-mode silicon neuron with robust neuromodulation, operating at a mere 40-200 pJ/spike. Even more fundamental, the University of Pennsylvania demonstrates in “Analog Physical Systems Can Exhibit Double Descent” that analog physical systems can exhibit complex machine learning phenomena like double descent without digital computation, hinting at a truly energy-minimal path forward.

• Sustainable Networking and HPC: Energy efficiency isn’t just about the chips; it’s about the entire ecosystem. Papers like “Energy Efficient Sleep Mode Optimization in 5G mmWave Networks via Multi Agent Deep Reinforcement Learning” by North Carolina State University show how Multi-Agent Deep Reinforcement Learning (MARL) can optimize 5G mmWave networks, boosting energy efficiency to 0.60 Mbit/Joule. Universidade de São Paulo (USP), Brazil, offers a comprehensive “Energy Efficiency in Network Slicing: Survey and Taxonomy”, emphasizing AI’s role in future sustainable networks. For High-Performance Computing (HPC), the University of Chicago in “Core Hours and Carbon Credits: Incentivizing Sustainability in HPC” proposes Energy-Based and Carbon-Based Accounting to incentivize sustainable user behavior, reducing energy use by up to 40%.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often built upon or necessitate new tools and resources:

• throttLL’eM Framework: Proposed by William Blaskowicz in “SLO-aware GPU Frequency Scaling for Energy Efficient LLM Inference Serving”, this framework optimizes GPU frequency scaling for LLMs. Code available at https://github.com/WilliamBlaskowicz/throttLL-eM.
• TNAS-ER Framework: Developed by Shandong University for “Efficient Eye-based Emotion Recognition via Neural Architecture Search of Time-to-First-Spike-Coded Spiking Neural Networks”, this is the first NAS framework for TTFS SNNs.
• ESACT Sparse Accelerator: Introduced by University of Example and Research Institute for AI in “ESACT: An End-to-End Sparse Accelerator for Compute-Intensive Transformers via Local Similarity”, for efficient transformer models. Code at https://github.com/ESACT-Project/esact.
• Energy Profiling Model for Data Pipelines: Proposed by Technische Universität Berlin in “Energy Profiling of Data-Sharing Pipelines: Modeling, Estimation, and Reuse Strategies” to optimize energy usage in cross-organizational data sharing. Code is available: https://github.com/Sepide-Masoudi/Energy-profiling-in-federated-data-sharing-pipelines.
• HiFiNet for WSN Fault Identification: A hierarchical framework by Phenikaa University, Vietnam, and Hanoi University of Science and Technology, Vietnam, described in “HiFiNet: Hierarchical Fault Identification in Wireless Sensor Networks via Edge-Based Classification and Graph Aggregation”, using Intel Lab Data and NASA’s MERRA-2 data for simulations.
• WebAssembly Runtimes for IoT: “WebAssembly on Resource-Constrained IoT Devices: Performance, Efficiency, and Portability” from the Faculty of Electrical Engineering and Computing, Zagreb, Croatia, evaluates WAMR and wasm3 on microcontrollers like Raspberry Pi Pico and ESP32. Code links are provided: https://github.com/appcypher/awesome-wasm-runtimes, https://github.com/wasm3/wasm3, https://github.com/bytecodealliance/wasm-micro-runtime.
• Carbon-Aware Intrusion Detection: “Carbon-Aware Intrusion Detection: A Comparative Study of Supervised and Unsupervised DRL for Sustainable IoT Edge Gateways” integrates Deep Reinforcement Learning (DRL) with IoT security to reduce carbon footprint.
• Visual-Word Tokenizer (VWT): A training-free method from the University of Sussex, UK, in “Visual-Word Tokenizer: Beyond Fixed Sets of Tokens in Vision Transformers” to reduce energy costs in vision transformers. Code at https://github.com/wearepal/visual-word-tokenizer.
• FAST Coreset Selection Framework: From Xi’an Jiaotong University, in “FAST: Topology-Aware Frequency-Domain Distribution Matching for Coreset Selection” a DNN-free framework achieving 96.57% power reduction and 2.2x speedup on CPU.
• TaWQ for SNNs: Pengcheng Laboratory, China, introduces “Temporal-adaptive Weight Quantization for Spiking Neural Networks” to improve efficiency. Code at https://github.com/ZhangHanN1/TaWQ.
• HERMES Middleware: Proposed by INESC INOV, Instituto Superior Técnico, Universidade de Lisboa, Portugal, in “HERMES: Heterogeneous Application-Enabled Routing Middleware for Edge-IoT Systems” for flexible routing in IoT networks. Code available at https://github.com/jequinhatavares/HERMES.
• Microbenchmarking NVIDIA’s Blackwell: “Microbenchmarking NVIDIA’s Blackwell Architecture: An in-depth Architectural Analysis” by NVIDIA Corporation and The Ohio State University offers detailed insights into GPU performance and optimizations. Code: https://github.com/NVIDIA/Blackwell-Microbenchmarking.

Impact & The Road Ahead: Green AI for a Smarter World

The implications of this research are profound. We are moving towards an era where AI systems are not only intelligent but also inherently sustainable. The advancements in hardware-aware NAS, dynamic frequency scaling for LLMs, and innovative neuromorphic designs promise significant reductions in computational energy demands, making powerful AI more accessible and environmentally responsible. For instance, the improvements in HPC sustainability through incentive models, as shown in “Core Hours and Carbon Credits: Incentivizing Sustainability in HPC”, are critical for managing the carbon footprint of large-scale research and enterprise AI.

Looking ahead, the convergence of AI with other fields like wireless communications (e.g., “A Spatial Array for Spectrally Agile Wireless Processing” by Nokia for 5G-Advanced and 6G, or “Low-Power Double RIS-Assisted Mobile LEO Satellite Communications”) and smart infrastructure (e.g., “Assessing the Technical and Environmental Impacts of Energy Management Systems in Smart Ports”) will multiply these gains. The development of biologically inspired AI, as seen in neuromorphic systems, holds the potential for breakthroughs in energy efficiency that could redefine what’s possible for AI at the edge. The next frontier involves refining these hybrid approaches, bridging the gap between theoretical models and real-world deployment, and fostering a collaborative ecosystem where sustainability is a core design principle for all AI/ML endeavors. The journey toward truly green AI is exhilarating, and these papers mark crucial steps on that path.