Energy Efficiency in AI/ML: From Silicon to Sustainable Systems

Latest 50 papers on energy efficiency: Nov. 16, 2025

The relentless march of AI/ML, while delivering unprecedented capabilities, comes with an increasingly significant environmental footprint. The sheer computational power required to train and run complex models is pushing the boundaries of energy consumption, making energy efficiency a critical frontier. Fortunately, a flurry of recent research, as highlighted in the papers summarized here, is tackling this challenge head-on, from novel hardware designs to intelligent software orchestration and sustainable deployments.

The Big Idea(s) & Core Innovations

At the heart of these advancements lies a common theme: a fundamental shift towards hardware-software co-design and domain-specific optimizations. The paper “The Role of Advanced Computer Architectures in Accelerating Artificial Intelligence Workloads” emphasizes that AI is no longer merely an application running on hardware; it’s an active partner in the design process. This symbiotic relationship is yielding groundbreaking results, as seen in the development of Processing-in-Memory (PIM) and neuromorphic computing architectures. A prime example is the event-driven spiking compute-in-memory macro based on SOT-MRAM, detailed in “An Event-Driven Spiking Compute-In-Memory Macro based on SOT-MRAM”, which significantly enhances energy efficiency for neuromorphic applications by processing asynchronous and sparse data more effectively. Similarly, the work on “Self-correcting High-speed Opto-electronic Probabilistic Computer” by Quantum Dice Limited, Oxford, UK, introduces a self-correcting optoelectronic probabilistic computer using quantum photonic p-bits, achieving remarkable speed and energy consumption figures by harnessing robust electronic control.

Beyond specialized hardware, intelligent software-level optimizations are proving equally vital. “AIM: Software and Hardware Co-design for Architecture-level IR-drop Mitigation in High-performance PIM” from Peking University and Houmo AI, introduces an innovative co-design approach that leverages workload characteristics to dynamically adjust power delivery, leading to substantial IR-drop mitigation and energy savings in PIM systems. In the realm of large language models (LLMs), “Leveraging LLMs to Automate Energy-Aware Refactoring of Parallel Scientific Codes” by Matthew T. Dearing et al. (University of Illinois Chicago and Argonne National Laboratory) presents LASSI-EE, an LLM-based framework that automates the generation of energy-efficient parallel scientific code, achieving up to 48% energy reduction across diverse hardware. This demonstrates how AI can become self-aware of its own energy consumption, fostering greener development.

Under the Hood: Models, Datasets, & Benchmarks

To drive these innovations, researchers are developing and utilizing a range of specialized tools and benchmarks:

• Hardware Architectures & Accelerators: Numerous papers focus on optimizing for specific architectures. “Low-cost yet High-Performant Sparse Matrix-Matrix Multiplication on Arm SME Architectures” introduces LOOPS, a hybrid framework that leverages Arm SME’s vector and matrix units for efficient sparse matrix multiplication, outperforming GPUs in energy efficiency. For edge devices, “FPGA-Accelerated RISC-V ISA Extensions for Efficient Neural Network Inference on Edge Devices” and “LL-ViT: Edge Deployable Vision Transformers with Look Up Table Neurons” demonstrate the power of FPGAs and custom RISC-V extensions for low-latency, energy-efficient inference. The latter, by EdgeAI Research Lab, MIT, uses novel lookup table neurons to make Vision Transformers viable on resource-constrained hardware.
• Benchmarking Frameworks: Critical for evaluating and comparing energy-efficient solutions are new benchmarks. “Smart but Costly? Benchmarking LLMs on Functional Accuracy and Energy Efficiency” introduces BRACE, a framework with novel CIRC and OTER rating methods to benchmark LLMs on both accuracy and energy. For HPC data centers, Hewlett Packard Enterprise and Oak Ridge National Laboratory’s “LC-Opt: Benchmarking Reinforcement Learning and Agentic AI for End-to-End Liquid Cooling Optimization in Data Centers” provides a comprehensive benchmark for liquid cooling optimization, featuring a high-fidelity digital twin of the Frontier supercomputer.
• Software Frameworks & Protocols: For distributed and memory-intensive applications, “BIPPO: Budget-Aware Independent PPO for Energy-Efficient Federated Learning Services” introduces a budget-aware PPO framework to enhance energy efficiency in federated learning. In memory systems, “WIRE: Write Energy Reduction via Encoding in Phase Change Main Memories (PCM)” by Zhang, Li, and Wang (University of California, Stanford, and Tsinghua University) presents a novel coding mechanism, WIRE, to reduce bit flips and extend PCM lifetime, achieving significant write energy improvements.
• Public Code Repositories: Many of these papers are accompanied by open-source code, fostering reproducibility and further research. Notable examples include LOOPS-SpMM, BRACE for LLM benchmarking, sustain-lc for liquid cooling, and AEOSBench for Earth observation satellite scheduling.

Impact & The Road Ahead

These advancements have profound implications across the AI/ML landscape. From improving the sustainability of large-scale cloud infrastructure with energy-aware container orchestration as seen in “Experimenting with Energy-Awareness in Edge-Cloud Containerized Application Orchestration” by D. Ali and R. C. Sofia from fortiss research institute, to making AI accessible on tiny edge devices, the research is pushing the boundaries of what’s possible. Conversational agents like SPARA, presented in “Conversational Agents for Building Energy Efficiency – Advising Housing Cooperatives in Stockholm on Reducing Energy Consumption” by KTH Royal Institute of Technology, demonstrate AI’s potential to directly guide energy-saving efforts in real-world scenarios, achieving expert-level precision in advising on building retrofitting.

Looking ahead, the integration of insights from biological systems, as explored in “The Demon Hidden Behind Life’s Ultra-Energy-Efficient Information Processing – Demonstrated by Biological Molecular Motors” by Toshio Yanagida et al., promises to revolutionize computing by leveraging probabilistic fluctuations and noise as computational resources, vastly outperforming current digital computation in energy efficiency. This could inspire entirely new AI architectures. Moreover, “Geospatial Foundation Models to Enable Progress on Sustainable Development Goals” introduces SustainFM, a framework to evaluate FMs against SDGs, highlighting the importance of not just accuracy, but also energy efficiency and ethical considerations for responsible AI deployment. The path forward involves continued interdisciplinary collaboration, pushing the boundaries of hardware-software co-design, and drawing inspiration from nature to build a truly sustainable and intelligent future for AI.