Jul 2026· International journal of computer information systems and industrial management applications· Vol 18, pp. 407-418· 0 citations
TL;DR
These findings demonstrate that meaningful reductions in AI-related carbon emissions are achievable through coordinated system-level and model-level interventions without requiring fundamental changes to underlying hardware infrastructure.
Abstract
The rapid growth of artificial intelligence training and inference workloads has produced a corresponding rise in data center energy consumption and associated carbon emissions, raising concerns about the environmental sustainability of continued AI scaling. This paper proposes a Sustainable AI Computing framework that combines carbon-aware dynamic workload scheduling with neural network optimization to reduce the operational carbon footprint of AI systems without proportionally sacrificing model performance. The proposed scheduler performs both spatial carbon shifting, routing deferrable workloads to data center regions with lower real-time grid carbon intensity, and temporal carbon shifting, delaying flexible jobs to lower-carbon time windows within deadline constraints, while a complementary neural network optimization module applies adaptive pruning, quantization, mixed-precision computation, and early-exit inference to reduce per-job energy consumption. The framework was evaluated across a simulated multi-region data center testbed spanning five regions with heterogeneous grid carbon intensity profiles, using representative training and inference workloads including convolutional and transformer-based models. Experimental results show that the proposed hybrid carbon-aware scheduler and optimizer combination reduces relative carbon emissions to 44% of a static round-robin baseline, compared to 71% for spatial shifting alone and 68% for temporal shifting alone, while cluster GPU utilization improves from an average of 52% under baseline scheduling to 78% under the proposed scheduler. Accuracy-energy trade-off analysis identifies a Pareto-efficient operating region in which up to 55% energy reduction per inference is achievable with less than one percentage point of accuracy degradation, beyond which further compression yields diminishing accuracy returns. Ablation results confirm that scheduling and model optimization contribute complementary and largely additive carbon reductions, with the combined framework outperforming either component in isolation. These findings demonstrate that meaningful reductions in AI-related carbon emissions are achievable through coordinated system-level and model-level interventions without requiring fundamental changes to underlying hardware infrastructure.
The rapid growth in artificial intelligence (AI) demand has significantly increased the electricity consumption and carbon emissions of computing centers. How to schedule AI requests across computing centers to reduce carbon emissions and electricity costs while maintaining low latency is an essential research problem. Existing schedulers reduce emissions by shifting workloads or balancing resources but usually simplify power system modeling, ignore transmission-side costs and carbon emissions, or make local decisions without batch-level coordination. To better address these problems, we first develop an ILP-based scheduler to get optimized results, but it faces scalability limitations. Then, we propose RAPID, a region-aware and power-informed scheduling framework that integrates static and online heuristic schedulers for large-scale AI request scheduling. Experiments based on real-world GenAI traces and Chinese regional power profiles show that RAPID significantly reduces carbon emissions, electricity costs, and total energy consumption compared to methods from previous works while maintaining zero Service Level Agreement (SLA) violations.
B. Ding, Cai-Ning Wang, Ka-Fei Tang et al.· Sustainability· 0 citations
The fast-growing Artificial Intelligence (AI) has led to the enormous processing requirements previously unknown, which cause high energy usage and related carbon emissions. Although the previous studies have concentrated on enhancing the efficiency of the algorithms, less attention has been given on the timing and placement of AI workloads. This paper introduces a Carbon-Aware AI Scheduler a full-stack application that optimizes the carbon footprint of AI workloads by utilizing smart time-shifting and geo-shifting operations. The system combines actual carbon intensity data, a hybrid model of forecasting which uses historical baselines and artificial diurnal curves, and daily job urgency and flexibility optimization algorithms. Experimental testing in over 50 regions throughout the U.S. shows that carbon emissions can be reduced by as much as 87 percent in case of geo-shifted workloads and by 2530 percent in case of time-shifted workloads in one region. The scheduler has a latency of sub-200 ms, which is acceptable in real-time application. Findings suggest that carbon-conscious scheduling, alongside the clear user feedback, may have a considerable effect on the behavior of developers and have a positive impact on the environmental impact of AI systems.
Harshraj N. Gadbail, Rajendra M. Rewatkar, Sujal Zade et al.· International Conference on...· 0 citations
Deep learning workloads have experienced rapid growth which has resulted in higher energy consumption and increased carbon emissions for contemporary data centres. The existing solutions of carbon tracking and static scheduling systems provide insufficient capacity to implement carbon awareness in actual machine learning operational processes. In this paper, we present EcoSchedAI (Eco-Aware Scheduling using Artificial Intelligence), a carbon-aware scheduling framework that integrates both real-time and predicted carbon intensity into the AI training process. The system introduces the concept of carbon opportunity windows, enabling proactive scheduling of training jobs during low-carbon periods. It also incorporates job-level constraints, allowing flexible and practical deployment in real-world scenarios. The framework uses a convolutional neural network (CNN) to evaluate three different scheduling methods which include baseline and rule-based and reinforcement learning-based approaches together with multi-region carbon assessment. The results demonstrate that EcoSchedAI can reach a maximum carbon reduction of 9.72% while sustaining model performance through its small operational impact. This research demonstrates how predictive scheduling and constraint-aware scheduling methods can create environmentally friendly AI systems which operate at large scales.
Haarika Alla, K. Madhura, Shweta S. Aladakatti· Discover Artificial Intellig...· 0 citations
This work introduces a novel LLM-based predictive scheduling system designed to enhance operational efficiency while reducing the environmental impact of data centers, using an LLM to predict key metrics such as execution time and energy consumption from source code.
Hanzhao Wang, Jingxuan Wu, Yumeng Li et al.· 0 citations
Modern smart agriculture increasingly relies on edge computing for real-time, high-concurrency tasks such as wide-area drone-based crop monitoring. However, highly volatile workloads and severe environmental noise in agricultural Internet of Things (IoT) networks often lead to resource congestion and high latency when relying on traditional reactive scheduling. To address these challenges, this paper proposes a hybrid prediction-driven proactive resource scheduling method for edge computing. We construct a Variational Mode Decomposition-Convolutional Neural Network-Attention-Bidirectional Long Short-Term Memory (VMD-CNN-Attention-BiLSTM) model to filter environmental noise and accurately capture the spatio-temporal features of bursty traffic. Furthermore, a deep reinforcement learning scheduling algorithm based on Proximal Policy Optimization (PPO) incorporates future workload trends into its state space, dynamically optimizing task offloading. To evaluate the proposed Predictive Computational Scheduling Framework (PCSF), we developed a custom edge computing simulation environment and synthesized a hybrid dataset combining real-world server logs from the Alibaba Cluster Trace with deep learning inference workloads derived from a Wheat Plant Diseases image repository. Simulations demonstrate that the prediction model achieves a Root Mean Square Error of 0.030 and a Mean Absolute Error of 0.0215. Compared to static and reactive baselines, the PCSF reduces average task timeout violations to 2.2 and total system energy consumption by nearly 40%. This proactive mechanism effectively overcomes decision-making lags, enabling efficient, low-latency computing resource allocation for modern agricultural facilities.
Shizhen Bai, Rong-Hua Chen, Yongbo Tan et al.· Applied Sciences· 0 citations
The electricity demand of artificial intelligence (AI) data centers is growing faster than the grids hosting them can decarbonize, yet carbon-aware computing is rarely audited at the system level: workloads follow scalar-price or carbon signals, whose claimed savings need not survive network-wide dispatch. We develop a carbon-aware co-optimization framework that schedules checkpoint intervals, workload migration, dynamic voltage and frequency scaling, and behind-the-meter storage jointly with unit commitment, pricing system CO2 emissions beyond the market-internalized level, with split-conformal scenario bands and an out-of-sample cost certificate. On IEEE 39/118-bus systems mapped to 2025 French and German measurements, carbon-priced dispatch reduces emissions by 29.5% at a generation-cost premium below 0.01% (the carbon-signal headroom of this mapped system); system-level decomposition attributes more than 99% of this reduction at reference penetration to generation redispatch, an operator-side instrument. AI-load flexibility adds 0.23% of system emissions today, equivalent to about one quarter of the data-center fleet’s own carbon footprint, and 3.3% at gigawatt scale, where it cuts worst-day load shedding by up to 49%. The endogenous checkpoint policy recovers the reliability-optimal interval, whereas following the average carbon-intensity signal can increase system-level emissions. Conformal calibration maintains at-or-above-nominal coverage (97.5%); empirical bands under-cover (86.0%).
Jun-Qing Zhang, Yi Wang· Sustainability· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
