Search Results (34)

In enclosed cold-aisle data-center rooms, hotspot temperatures within the cold aisles are directly related to server inlet air safety and air-conditioning operation performance. Therefore, high-accuracy temperature prediction is of great significance for thermal-risk warning and operational optimization. This study focuses on an enclosed cold-aisle data-center room in a large data center located in a hot-summer and warm-winter region. The maximum temperature of one selected cold aisle is taken as the prediction target, and a hybrid prediction model integrating TSception, Transformer, and TCN is proposed to characterize multi-scale local disturbances and cross-period temporal dependencies. The model inputs and historical time window are determined through cold–hot separation analysis, cold-aisle temperature-stability comparison, lag analysis between supply air temperature and cold-aisle temperature, and correlation analysis of candidate variables. The experimental results show that the proposed model achieves an MAE of 0.285 °C, an RMSE of 0.438 °C, an NRMSE of 4.76%, and a MAPE of 1.13% in predicting the cold-aisle temperature of the AB aisle, outperforming Persistence, BP, LSTM, XGBoost, and several ablation models. Further multi-aisle experiments demonstrate that the proposed model consistently maintained low prediction errors in independent prediction tasks across different cold aisles, indicating good applicability of the model architecture for characterizing temperature time series in multiple cold aisles within the same data-center room. The proposed method can provide a reference for hotspot-risk identification in enclosed cold aisles and operational optimization of precision air-conditioning systems. Full article

Small modular reactors (SMRs) offer firm low-carbon heat and power, and data centers concentrate large, continuous electrical and cooling loads. Transparent tools for screening their thermal integration are scarce. To the authors’ knowledge, this paper develops the first steady-state thermodynamic framework that couples SMR steam extraction, solar thermal input, and recovered data-center liquid-cooling heat through a single mixing-tank thermal bus serving both an absorption chiller and an organic Rankine cycle (ORC). The framework’s novelty is in its focus on the structural level rather than the component level. Two consistency requirements are built into its equations. First, the data-center control volume closes exactly, so that recovered heat reduces the residual cooling demand and heat removal equals IT dissipation. Second, delivered cooling is credited identically in every configuration compared, so that apparent gains cannot arise from asymmetric accounting. The framework identifies the governing mechanism of the architecture: a small 120 °C extraction stream (1.01% of core thermal power at the activation bound) unlocks the larger 70 °C recovered stream, which cannot drive the chiller alone. At the margin-constrained design point (2.93% extraction), direct liquid recovery removes 20.9 MW_th, absorption cooling serves the remaining 16.2 MW_th, and net electricity is 5.8 MW_e above the all-electric reference. An itemized estimate places the integration-specific parasitic loads at 0.6–1.5 MW_e (central value 1.0 MW_e), which reduces the increment over the liquid-cooled non-integrated reference from $+0.5$ MW_e (gross) to approximately $-0.5$ MW_e (net). A 20,000-sample Monte Carlo analysis across seven uncertain parameters shows the net-of-parasitics gain over the all-electric reference is positive with 92% probability (median $+4.1$ MW_e), while the increment over the non-integrated reference is positive with only 29% probability. A compact exergy inventory attributes 7.1 MW of destruction to the recovery train (process heat exchanger 2.0, mixing 1.2, ORC 2.0, chiller 2.0). The architecture’s robust value therefore lies in thermally driven cooling and the productive use of recovered heat, not in net energy. The framework is a screening tool rather than a validated plant model; a companion study populates it with published plant, climate, and equipment data. Full article

Large-scale artificial intelligence systems increasingly disclose energy and carbon metrics, but their freshwater costs remain less consistently measured. This paper introduces the Water Cost of Intelligence (WCI), a per-query metric combining direct water consumed for on-site data-center cooling with indirect water consumed during electricity generation, weighted by local scarcity using Aqueduct 4.0 Baseline Water Stress (BWS) scores. We first reconstruct Google’s disclosed Gemini direct-water figure from Google’s own reported parameters, an internal consistency check on the implementation rather than an independent validation. Expanding the accounting boundary to include electricity-generation water raises the estimate for a median large language model (LLM) prompt by 179% under a uniform national water-intensity value. Parameterizing that intensity by the regional generation mix instead changes the estimate substantially and reverses the regional ordering, depending on whether hydroelectric reservoir evaporation is allocated to generation: the same grid is the least water-intensive of those studied under one convention and the most water-intensive under the other. A region cannot be characterized as water-efficient in terms of electricity without first establishing that convention. Direct-only reporting can be internally accurate yet boundary-incomplete, and regional scarcity can change the interpretation of identical physical water use by an order of magnitude. Full article

Data centers are large and rapidly growing electricity consumers, and cooling systems account for a substantial share of their energy demand. A key contribution of this study is a climate-sensitive, hourly techno-economic comparison of three data-center cooling configurations under consistent operating assumptions: a conventional water-cooled centrifugal chiller baseline, a magnetic bearing chiller (MBC) system, and an MBC system integrated with aquifer thermal energy storage (ATES). The comparison is performed for Phoenix, Arizona, and Fairbanks, Alaska, which represent substantially different cooling climates in the U.S. Hourly simulations use identical information technology (IT) load profiles, identical aggregate installed chiller capacity represented by two 4058 kW chiller units, common water-side economizer controls, and site-specific weather and electricity tariffs. Results show that the MBC system reduces annual cooling-system electricity consumption from 1169.4 to 957.4 MWh in Phoenix (18.1%) and from 361.6 to 319.4 MWh in Fairbanks (11.7%). Peak cooling-system electrical demand decreases by 119.4 kW in Phoenix and 71.6 kW in Fairbanks. Relative to the centrifugal baseline, the MBC case gives a 5.8-year simple payback in Phoenix but is not economically attractive in Fairbanks under the assumed tariff. The MBC-only case gives the lowest annual cooling electricity use in both climates. The MBC + ATES case is treated only as a screening-level, discharge-assisted cold-storage scenario rather than a full techno-economic assessment of seasonal ATES, and no site-specific hydrogeological feasibility assessment is performed. Under the assumed O&M cost structure, MBC + ATES gives a higher discounted value of savings than MBC-only, but this economic result is not caused by additional cooling-electricity savings relative to MBC-only. The MBC + ATES case also has a longer payback period because of its higher capital cost. These results show that the value of advanced cooling configurations depends on climate, free-cooling availability, electricity pricing, storage assumptions, and economic assumptions within the modeling framework considered in this study. Full article

Modern data centers are consuming more energy than ever before due to the rapid growth of cloud services, artificial intelligence (AI), and large-scale digital applications. As energy demand continues to rise, accurate load forecasting has become an important tool for improving energy management and operational planning. However, predicting data-center power demand remains challenging because computing workloads, cooling systems, and facility operations are closely connected and constantly changing. This review examines the use of artificial intelligence (AI) and machine learning (ML) techniques for data-center load forecasting. The surveyed literature covers traditional statistical methods, supervised learning algorithms, deep learning models, probabilistic forecasting techniques, and physics-informed hybrid approaches. Important topics such as forecasting horizons, feature selection, performance evaluation, and practical deployment challenges are also discussed. The reviewed studies indicate that advanced approaches, including long short-term memory (LSTM), gated recurrent unit (GRU), transformer-based models, and digital twin (DT) frameworks, can improve forecasting performance and support more energy-efficient operation. These methods can also assist carbon-aware and grid-interactive data-center management. Several challenges remain. These include limited data availability, poor generalization, interpretability issues, and real-time implementation constraints. In this paper, the reviewed studies are classified and analyzed to provide a clear reference for researchers working in AI-based data-center energy forecasting. Full article

AI-driven growth is pushing data-center electricity demand from 415 TWh (2024) toward 945 TWh by 2030. Demand-response research has matured on operational levers but treats the physical envelope as largely exogenous, leaving construction-phase decisions that bound achievable flexibility unmodeled. This article integrates three previously separate bodies of literature (data-center demand response, grid-interactive efficient buildings, and stochastic optimal control under irreversibility) into a single framework that prices construction-phase flexibility as a portfolio of real options, pairing elicitation-derived (FAHP) weights with simulation-derived (Sobol) variance indices. None of the individual techniques is new; the contribution is their synthesis and the finding that architectural and site decisions carry the dominant financial leverage in the model, whereas a literature-grounded synthetic-persona prior (twenty-five LLM-simulated personas) prioritizes mechanical-electrical systems; a divergence we frame as a screening diagnostic between an LLM prior and the model. The Flex-by-Design Readiness Index (FDRI) is a nineteen-dimension taxonomy across architectural, MEP, and site-urban layers, weighted by Fuzzy AHP. The FDRI–ROV model formalizes the decision as stochastic optimal control under irreversibility. Calibrated to PJM, ERCOT, and CAISO (2024–2026) for a 100 MW plant at N = 10,000 paths over thirty years, it yields +$79 M net option value at Full FDRI for PJM (additive upper bound; substitution-corrected ≈ +$41 M, 1.7× CapEx; 2.4× CapEx PJM, 2.1× ERCOT, 2.8× CAISO); on both the additive (2.4–2.8×) and corrected (1.7×) bases the pre-registered H2 threshold of Vtotal/Ctotal ≥ 3× is not met. Sobol decomposition places architectural and site layers at ST ≈ 0.56 each versus MEP at 0.15, exposing waste-heat-export and regulatory-avoided-cost dimensions as under-recognized leverage. Out-of-sample validation against four hyperscale projects yields 11% MAPE, reported as an n = 4, single-period proof of concept.A pro-rata extrapolation across all ~43 GW of incremental U.S. capacity gives a nominal ~$34 billion through 2035, but this applies the single most optimistic scenario uniformly; applying the substitution-corrected per-plant value with competition, policy, and adoption decay multipliers, the defensible 2035 opportunity is ≈$3–$18 billion (central ≈$7 billion), with $34 billion retained only as an undecayed ceiling. Full article

The rapid scaling of large language model (LLM) training has intensified demand for Graphics Processing Unit (GPU) clusters balancing throughput with energy efficiency. While NVIDIA’s H100 and B200 architectures are increasingly deployed in production datacenters, their comparative behavior under distributed training remains insufficiently characterized beyond vendor specifications, leaving datacenter operators without empirical guidance on metrics such as TFLOPs/kW and tokens-per-kilojoule. This work presents a system-level evaluation of single-node 8× H100 and 8× B200 configurations using Distributed Data Parallel (DDP) training across LLMs and vision–language models (VLMs) ranging from 7B to 32B parameters, spanning various real AI workload scenarios. We benchmark end-to-end throughput, utilization, power, energy, TFLOPs/kW, and tokens-per-kilojoule, complemented by architectural analysis explaining observed behavioral differences. Across LLM workloads, B200 achieves higher utilization (1–6%), faster training (up to 15%), and greater compute efficiency (up to 32% higher TFLOPs/GPU), attributable to higher memory bandwidth and large streaming multiprocessor (SM) count. However, B200 exhibits lower TFLOPs/kW and tokens-per-kilojoule, revealing a fundamental trade-off: throughput gains come at a measurable energy cost per useful token. VLM results further expose model-dependent asymmetries, with B200 consuming disproportionately more energy for lighter compute kernels due to elevated baseline power draw. These findings provide an empirical framework distinguishing compute efficiency from energy efficiency across next-generation GPU nodes, offering practical guidance for energy-aware AI datacenter design. Full article

The rapid growth of artificial intelligence (AI) workloads is reshaping semiconductor design across architecture, interconnect, memory hierarchy, packaging, timing, and design automation. Rather than converging on a single hardware solution, the field is expanding into a heterogeneous ecosystem that includes data-center graphics processing units (GPUs), edge neural processing units (NPUs), and application-specific integrated circuits (ASICs), field-programmable gate array (FPGA)-based and hybrid AI system-on-chip (SoC) platforms, chiplet-enabled systems, and emerging beyond-conventional-silicon approaches such as photonic, neuromorphic, and analog in-memory processors. This paper presents a comprehensive review of AI-on-chip systems from a cross-layer perspective. It examines AI chip architectures and hardware platforms, network-on-chip (NoC) designs for AI communication patterns, and algorithm–hardware co-design methods for model acceleration, including compression, quantization, and sparsity-aware optimization. It also reviews clocking, synchronization, and clock-domain-crossing (CDC) challenges in large heterogeneous systems and chiplets, as well as manufacturing, advanced packaging, and reliability issues, including two-and-a-half-dimensional (2.5D) and three-dimensional (3D) integration, thermal and mechanical constraints, assembly quality, and long-term yield considerations. In parallel, the paper surveys the growing role of AI in chip design itself, covering machine-learning-assisted analysis, Bayesian and reinforcement-learning-based optimization, and the emerging use of large language models (LLMs) and AI agents for register-transfer level (RTL) generation, design-space exploration, and autonomous electronic design automation (EDA) workflows. Finally, it discusses beyond-silicon AI chip directions and the broader economic and industry context shaping cloud, on-premises, and edge deployment. By integrating these topics into a unified framework, this review highlights the key technological drivers, system-level tradeoffs, and future research directions that will define next-generation scalable, reliable, and energy-efficient AI-on-chip systems. Full article

The rapid proliferation of Large Language Models (LLMs) is increasing electricity demand from data centers, creating new challenges for power-demand forecasting and grid planning. A key difficulty is that architecture- and deployment-related information that affects inference load is often private to LLM providers. This paper proposes a two-stage, mechanism-assisted forecasting framework under information asymmetry. In the first stage, a stylized incentive mechanism elicits verifiable reduced-form demand parameters from LLM providers at a chosen reporting precision. In the second stage, the elicited parameters are incorporated into forecasting models as architecture- and deployment-informed features. Using calibrated synthetic scenarios constructed from public data-center energy reports, open LLM-inference energy benchmarks, and secondary public estimates, we find that incorporating elicited parameters reduces the mean squared error (MSE) of the ResNet forecasting backbone by 65.1% relative to an architecture-agnostic ResNet baseline. Similar improvements are observed for a gradient-boosting model, indicating that the main empirical value comes from procuring informative provider-side demand features rather than from a specific neural architecture. The results should be interpreted as a proof-of-concept demonstration rather than a full operational model of LLM serving or power-system dispatch. Full article

AI infrastructure is entering a constraint-dominated regime in which power access, cooling, water conditions, reliability, and financing jointly shape cost, sustainability, and operational risk. Yet the metrics used to evaluate these systems remain fragmented across facility engineering, compute/workload performance, and economic or risk analysis, with definitions that often sit at different layers and under different boundaries. This fragmentation weakens cross-layer reasoning and makes decision-traceable trade-off analysis difficult. This paper proposes a structured, decision-oriented measurement architecture for AI infrastructure metrics. The framework combines a 6 × 3 taxonomy, which organizes metrics across six layers and three semantic domains, with a procedural workflow built around a problem card, variable registry, minimality gate record, activated-cell map, boundary log, metric ledger, and a results sheet with case-pack manifest. Within this protocol, the Metric Propagation Graph is used as a case-specific dependency representation for tracing decision-facing metrics back to minimal boundary-consistent inputs. It is introduced as a traceability layer within the framework rather than as a stand-alone graph-theoretic method. The paper is illustrated through one fully worked case and one scoped portability illustration. The first is a fully worked large-load planning case for the Northern Virginia data-center corridor within PJM’s Dominion zone, showing that a boundary-consistent integrated metric can reverse the ranking obtained under a simpler screening view. The second is a scoped portability illustration for hourly matching under dual Scope 2 boundaries. Its purpose is not to provide a second full empirical validation, but to show how the same dossier logic, boundary discipline, and traceable metric construction transfer to a distinct decision setting. Full article

Multi-campus AI data centers procure identical hardware and service SKUs across geographically heterogeneous locations, yet finance and operations require a single system-level benchmark (“world price”) per SKU for budgeting, chargeback, and capacity planning. Naive deployment-weighted aggregation preserves total cost but can induce Simpson-type aggregation bias, where heterogeneous location mixes reverse global SKU rankings and weaken managerial decision signals. This study formalizes the problem of location-robust, cost-preserving aggregation and develops two mathematically structured operators for production cost pipelines. The first operator applies a two-way fixed-effects decomposition to separate global SKU effects from campus-specific premia, followed by normalization to guarantee exact cost preservation. This yields an interpretable benchmark that performs well when campus coverage is sufficiently broad and location effects remain approximately additive. The second operator solves a constrained convex common-weight optimization, producing a unified set of non-negative campus weights that preserves total cost while providing the strongest protection against dominance reversals in the ordered setting. Simulation experiments and a semi-real calibrated AI datacenter OPEX illustration show that both operators substantially improve ranking stability relative to naive blending, while the convex operator serves as the more conservative safeguard under adverse heterogeneity. The resulting detect–correct–validate workflow provides a scalable decision-support framework for robust cost aggregation in distributed AI infrastructure and illustrates how symmetry-preserving aggregation operators can stabilize benchmarking in large heterogeneous systems. Full article

Data-center power lines are nearing their thermal and operational limits, creating a need for higher transfer capability, lower voltage regulation, and improved transmission efficiency. Although series capacitor compensation is a well-established transmission technique, its application to large data-center interconnections requires a clearer understanding of how compensation level affects controllable power delivery under practical voltage regulation requirements. This paper develops analytical transmission-line models without and with series compensation and applies them to the grid-to-data-center transmission interface. The study quantifies how series compensation affects voltage regulation, reactive power requirement, transferable power, and transmission efficiency under two operating regimes: an unconstrained receiving-end voltage case and a constrained terminal-voltage case. The results show that, when the receiving-end voltage is not strictly regulated, increasing the degree of series compensation significantly reduces voltage regulation and reactive power demand while enhancing power transfer capability. However, when the sending-end and receiving-end voltages are constrained to remain at or near nominal values, the maximum transferable power increases only up to an optimal compensation level, beyond which it declines as compensation approaches 100%. The analysis further shows that coordinated regulation of voltage magnitude and angle becomes necessary at high compensation levels to maintain controllable and efficient power transfer. Overall, the paper provides a data-center-oriented framework for identifying when series compensation improves power delivery and when additional transmission control becomes necessary. Full article

Carbon-neutral industrial parks require large consumers, such as data centers, to balance low-carbon operation and service reliability. This paper proposes a resilience-constrained stochastic dispatch framework for a data-center virtual power plant (VPP) with renewable generation, short-duration batteries, and long-duration storage units. The dispatch is formulated as a two-stage stochastic program with normal and outage scenarios. To solve the resulting large mixed-integer problem, we develop a hybrid quantum–classical L-shaped method: the integer master is solved heuristically by quantum annealing, while scenario subproblems are solved exactly by classical optimization. In a case study based on real-world industrial-park data, the proposed storage strategy eliminates critical load shedding for the tested 6 h outage scenarios with a 3.7% increase in expected daily cost. The QA-driven method reaches the same best-known objective as the classical baseline with an empirical 1.36× runtime speedup. Full article

As the exponential growth in advanced compute workloads drives intra-datacenter interconnects to ever increasing bitrates, optical networking equipment has risen to the challenge by shifting from NRZ signaling to bandwidth efficient modulation methods such as PAM4. As these modulation schemes introduce an inherent SNR penalty, maintaining low bit error rates (BER) forces optical links to operate at significantly higher optical powers. However, increasing the optical power leads to photodetectors reaching one of their fundamental bottlenecks caused by the space-charge effect, limiting their ability to provide a high-speed response under high-power illumination. This work presents the design, fabrication, and characterization of a waveguide-integrated photodiode with dual optical inputs (DIPD) designed to overcome this limitation. Specifically, we demonstrate that combining a dual-fed architecture with targeted cross-sectional geometric optimizations effectively distributes the photocurrent density to delay the onset of space-charge saturation. Experimental validation demonstrates a high responsivity of ≈0.91 [A/W] (for O-band wavelengths) and a large electro-optic bandwidth (EOBW) of ≈58 [GHz], all under high-power illumination and CMOS driving voltages. Full article

The increasing energy consumption of large datacenters, with cooling systems constituting a significant portion, calls for efficient thermal management strategies. Conventional computational fluid dynamics (CFD) methods, although accurate, are time-consuming for supporting real-time tasks in dynamic datacenter environments. Machine learning (ML)-based methods, particularly artificial neural network (ANN)-based surrogate models, have emerged as potential alternatives, but they struggle with generalization across diverse working conditions. Meanwhile, ML models’ performance in large datacenters still remains unclear. This research introduces a hybrid surrogate model combining ANNs and CNNs for the precise and rapid prediction of 3D temperature distributions in large datacenters. The proposed method incorporates an ANN for feature processing and a CNN for decoding spatial features, leveraging both to capture complex airflow patterns and temperature distributions under varying conditions. A dataset of 500 CFD-simulated temperature fields based on a real datacenter is established for model training and validation. The CFD method is evaluated by comparing the simulation results with experimental data. Results of the ML models’ performance indicate that the proposed hybrid surrogate model outperforms the conventional ANN model, reducing mean absolute error (MAE) by 87.44%. Additionally, the model is 300,000 times faster than CFD simulations, offering an efficient solution for further supporting real-time thermal management. Full article