Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Data centers”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 289 records · Page 16
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Exploring DAOS as a Burst Buffer for a 100 Gbps DAQ Real-Time Streaming System

We present an experimental evaluation of a burst buffer for a real-time DAQ streaming system designed to transmit instrument data to remote data centers. The system is based on EJ-FAT, a load balancing system capable of Nx 100Gbps streams, distributing data from event sources to processing nodes. We explore applying the DAOS system as a burst buffer to serve a number of purposes: improve resiliency, elasticity and add new functions into the processing pipeline. In the evaluation a sender transmits events over a 100Gbps network to a receiver integrated with DAOS to store the reassembled events using DAOS APIs. We evaluate the system for possible bottlenecks and provide end-to-end evaluation with a burst buffer using DAOS storage abstractions. We show that a receiver node can support 38.1 Gbps. This proves the viability of our approach and allows us to extend this work to investigate scale-out properties and new streaming optimizations.

Mei, Xinxin↗
DOE OSTI2026

Public Reference Data for Megawatt-Scale Hydrogen Electrolysis - NLR Historical Wind

The U.S. Department of Energy and the National Laboratory of the Rockies (NLR) demonstrate hydrogen electrolysis from variable sources, hydrogen compression and storage, and hydrogen fuel cell power production using megawatt-scale equipment at NLR’s Flatirons Campus as part of the Advanced Research on Integrated Energy Systems (ARIES) initiative. This dataset represents part of that effort and is intended for academic, national laboratory, industrial, and other stakeholders to plan, design, and validate models of megawatt-scale hydrogen technologies and diverse energy infrastructure nationwide. These data provide a baseline for how existing hydrogen electrolysis technologies perform when coupled with various energy technologies. Future datasets will demonstrate how existing hydrogen fuel cell technologies can provide controllable, dispatchable, and variable power output for artificial intelligence (AI) data centers and other variable loads. This dataset entry describes hydrogen production by conducting a statistical analysis of historical wind data over a five-year period (2020-2025) from a single 1.5MW turbine manufactured by General Electric (GE) located at NLR’s Flatirons Campus, to generate an experimental test profile that was deployed on a 1.25-MW proton exchange membrane type MC250 electrolyzer system manufactured by Nel Hydrogen . [1] While the electrolyzer balance-of-plant supports up to 2.5 MW of electrolysis, NLR only has a single 1.25-MW electrolysis stack. The historical wind data provided several metrics, however, the analysis particularly focused on the measured power output by the wind turbine. The power output time series of data for each day was categorized by total energy generation and standard deviation, and the day that represented the highest combination of these two metrics was chosen – December 25th, 2022. This process was then repeated for a moving four-hour window within this day to identify the most statistically variable period. Finally, this four-hour period was scaled by 65% to match the 1.25 MW electrolyzer. The electrolysis system controls hydrogen production by varying DC current applied to the stack, from a maximum of 3000 A to a minimum safe operation of 300 A, or 10%. Because the current – voltage characteristic changes as the stack ages and efficiency degrades, the actual minimum safe operating power changes over time. The historical wind profiles were translated from power (kilowatts) to current (amperes) using a curve fit with calibration data and sent to the electrolyzer power supply at 1 Hz frequency. For more details on the statistical analysis process, see the presentation labeled “ Public Reference Data for Megawatt-Scale Hydrogen Electrolysis” provided with each data entry. These datasets report relevant hydrogen balance-of-plant and system data, all captured at 1 Hz, including hydrogen mass production measured with an Emerson Coriolis flow meter. Each .zip file represents a single wind turbine electrolysis experiment and is formatted as follows: {technology}_{scaling factor}-{electrolyzer ramp rate in amperes/second} For instance, “wind-GE1.5MW_0.65-400.zip” represents the hour-long experiment using historical data from the wind-GE1.5MW turbine, scaled to 65%, with the electrolyzer power supply set to a maximum ramp rate (gain and slew) of 400 A/s. Each .zip folder contains the following files: A .csv file containing raw data An .xlsx file explaining all the fields in the raw data. A .png plot showing the time series of hydrogen production, electrolysis power consumption, and wind power input. A PDF file detailing the historical wind data statistical analysis used to generate the wind profile. An experiment labeled “characterization_200.zip” demonstrates the MC250 electrolyzer steady-state response with 30-minute load steps for a total duration of 5 hours. Finally, a .csv file is provided with all simulated wind experiments combined into one dataset labeled "combined_historical_wind_experiments.csv". NLR also built an AI/machine-learning predictive model based on these datasets. The model ingests the electrolyzer current command in amperes, as well as various pressures and temperatures across the system, and predicts hydrogen output in kilograms per hour. The complete model can be found at https://huggingface.co/NatLabRockies/ptmelt-hydrogen-electrolysis [1] nelhydrogen.com/product/mc-series-electrolyser .

08 HYDROGEN↗
DOE OSTI2022

RCSB Protein Data Bank (RCSB.org): delivery of experimentally-determined PDB structures alongside one million computed structure models of proteins from artificial intelligence/machine learning

Abstract The Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB), founding member of the Worldwide Protein Data Bank (wwPDB), is the US data center for the open-access PDB archive. As wwPDB-designated Archive Keeper, RCSB PDB is also responsible for PDB data security. Annually, RCSB PDB serves >10 000 depositors of three-dimensional (3D) biostructures working on all permanently inhabited continents. RCSB PDB delivers data from its research-focused RCSB.org web portal to many millions of PDB data consumers based in virtually every United Nations-recognized country, territory, etc. This Database Issue contribution describes upgrades to the research-focused RCSB.org web portal that created a one-stop-shop for open access to ∼200 000 experimentally-determined PDB structures of biological macromolecules alongside >1 000 000 incorporated Computed Structure Models (CSMs) predicted using artificial intelligence/machine learning methods. RCSB.org is a ‘living data resource.’ Every PDB structure and CSM is integrated weekly with related functional annotations from external biodata resources, providing up-to-date information for the entire corpus of 3D biostructure data freely available from RCSB.org with no usage limitations. Within RCSB.org, PDB structures and the CSMs are clearly identified as to their provenance and reliability. Both are fully searchable, and can be analyzed and visualized using the full complement of RCSB.org web portal capabilities.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2023

E85 Fueling Infrastructure Trends: A Decade in Review

This report provides information on ethanol fueling infrastructure and industry trends over the last 10 years. It is informed primarily by ethanol fueling station location data collected through the U.S. Department of Energy's Alternative Fuels Data Center (AFDC) Station Locator from 2011 through 2021 (AFDC 2021a). Industry stakeholders, including the Renewable Fuels Association and Growth Energy, also provided data and additional context around trends seen in the data.

33 ADVANCED PROPULSION SYSTEMS↗
DOE OSTI2026

Municipality of Anchorage Household Travel Survey

The 2002 Household Travel Survey for the municipality of Anchorage, Alaska, collected demographic, socioeconomic, and travel information about households and persons (age five and older) on an assigned, 24-hour travel period. The main objective of the study, conducted by NuStats, was to improve the transportation system. 2,035 households were recruited to participate in the study. Of these, 1,293 completed travel logs depicting detailed data on driving habits, such as purpose of the trips, time of the data, and mode of transportation—spanning from April 1, 2002, to May 17, 2002. This dataset is part of the Metropolitan Travel Survey Archive, which includes travel surveys from numerous public agencies across the United States and is archived by the Transportation Secure Data Center to ensure their continued public availability.

1Hz data↗
DOE OSTI2026

2001 Atlanta Household Travel Survey

The 2001 Atlanta Household Travel Survey collected demographic, socioeconomic, and travel information on work and non-work travel behavior for a 48-hour travel period. The study was conducted on behalf of the Atlanta Regional Commission, and it is an essential element in the transportation planning and modeling efforts for the 13-county Atlanta region. The main objective of the study was to produce data that could be used to develop and calibrate travel demand models for use in travel forecasting, land use planning, and air quality planning to improve the transportation system. The second component in the survey was the deployment of an electronic travel diary with person-based GPS and an accelerometer to collect health and activity information. Travel data includes trip generation, trip distribution, and modal choice. The survey recruited a total of 12,184 households to participate in the study. Of these, 8,069 households (66%) completed travel. The 8,069 households, when weighted, represent 21,323 persons, 14,449 vehicles, and 126,127 places visited from April 2001 through April 2002. This dataset is part of the Metropolitan Travel Survey Archive, which includes travel surveys from numerous public agencies across the United States and is archived by the Transportation Secure Data Center to ensure their continued public availability.

1Hz data↗
DOE OSTI2021

RouteE: A Vehicle Energy Consumption Prediction Engine

The emergence of connected and automated vehicles and smart cities technologies create the opportunity for new mobility modes and routing decision tools, among many others. To achieve maximum mobility and minimum energy consumption, it is critical to understand the energy cost of decisions and optimize accordingly. The Route Energy prediction model (RouteE) enables accurate estimation of energy consumption for a variety of vehicle types over trips or sub-trips where detailed drive cycle data are unavailable. Applications include vehicle route selection, energy accounting and optimization in transportation simulation, and corridor energy analyses, among others. The software is a Python package that includes a variety of pre-trained models from the National Renewable Energy Laboratory (NREL). However, RouteE also enables users to train custom models using their own data sets, making it a robust and valuable tool for both fast calculations and rigorous, data-rich research efforts. The pre-trained RouteE models are established using NREL's Future Automotive Systems Technology Simulator paired with approximately 1 million miles of drive cycle data from the Transportation Secure Data Center, resulting in energy consumption behavior estimates over a representative sample of driving conditions for the United States. Validations have been performed using on-road fuel consumption data for conventional and electrified vehicle powertrains. Transferring the results of the on-road validation to a larger set of real-world origin-destination pairs, it is estimated that implementing the present methodology in a green-routing application would accurately select the route that consumes the least fuel 90% of the time. The novel machine learning techniques used in RouteE make it a flexible and robust tool for a variety of transportation applications.

47 OTHER INSTRUMENTATION↗
DOE OSTI2022

Prospects of federated machine learning in fluid dynamics

Physics-based models have been mainstream in fluid dynamics for developing predictive models. In recent years, machine learning has offered a renaissance to the fluid community due to the rapid developments in data science, processing units, neural network based technologies, and sensor adaptations. So far in many applications in fluid dynamics, machine learning approaches have been mostly focused on a standard process that requires centralizing the training data on a designated machine or in a data center. In this article, we present a federated machine learning approach that enables localized clients to collaboratively learn an aggregated and shared predictive model while keeping all the training data on each edge device. We demonstrate the feasibility and prospects of such a decentralized learning approach with an effort to forge a deep learning surrogate model for reconstructing spatiotemporal fields. Our results indicate that federated machine learning might be a viable tool for designing highly accurate predictive decentralized digital twins relevant to fluid dynamics.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Public Reference Data for Megawatt-Scale Hydrogen Electrolysis - NLR Historical Solar PV

The U.S. Department of Energy and National Laboratory of the Rockies (NLR) demonstrate hydrogen electrolysis from variable sources, hydrogen compression and storage, and hydrogen fuel cell power production using megawatt-scale equipment at NLR’s Flatirons Campus as part of the Advanced Research on Integrated Energy Systems (ARIES) research platform. This dataset represents part of that effort and is intended for academic, national laboratory, industrial, and other stakeholders to plan, design, and validate models of megawatt-scale hydrogen technologies and diverse energy infrastructure nationwide. These data provide a baseline for how existing hydrogen electrolysis technologies perform when coupled with various energy technologies. Future datasets will demonstrate how existing hydrogen fuel cell technologies can provide controllable, dispatchable, and variable power output for artificial intelligence data centers and other variable loads. This dataset entry describes the behavior of a 1.25-MW proton exchange membrane MC250 electrolyzer system, manufactured by Nel Hydrogen , [1] when fed historical data generated by the 430-kW, fixed-axis solar photovoltaic (PV) array located at NLR’s Flatirons Campus. (While the electrolyzer balance of plant supports up to 2.5 MW of electrolysis, NLR only has a single 1.25-MW electrolysis stack.) Solar PV power output data for the 2020 calendar year were categorized on a daily basis by total energy generation and standard deviation. Each day was then ranked by these metrics, and the 25th, 50th, and 100th percentiles were selected. The 75th percentile day did not exhibit sufficient variability to make for a valuable experiment. A similar process was used for the related historical wind dataset . [2] The historical days in 2020 that represented these percentiles are Dec. 19, March 29, and May 4, respectively. The entire solar day’s power profile was then fed through the MC250 electrolyzer. Due to its length, the 100th percentile day experiment was split into two parts, and the final 3 hours of the solar day were not captured. These final 3 hours contained no spikes or dips of interest and simply represented a slow decay of input solar power. Also, a single timestamp (13:13:47 on Jan. 14, 2026) was lost in the hydrogen system supervisory control and data acquisition. Finally, during the 25th percentile experiment (solar day Dec. 19, 2020) data recording was lost from 11:00:13 to 11:14:45. The roughly 15 minutes of the solar profile were rerun at the end of the experiment and spliced into this time slot during post-processing. The electrolysis system controls hydrogen production by varying direct current applied to the stack, from a maximum of 3,000 A to a minimum safe operation of 300 A, or 10%. Because the current–voltage characteristic changes as the stack ages and efficiency degrades, the actual minimum safe operating power changes over time. The historical solar profiles were translated from power (kilowatts) to current (amperes) using a curve fit with calibration data and sent to the electrolyzer power supply at 1-Hz frequency. For more details on the statistical analysis process, see the slide deck “Public Reference Data for Megawatt-Scale Hydrogen Electrolysis: NLR Historical Solar PV Analysis and Profile Generation” accessible with this data entry. These datasets report relevant hydrogen balance-of-plant and system data, all captured at 1 Hz, including hydrogen mass production measured with an Emerson Coriolis flow meter. Each .zip file represents a single solar PV electrolysis experiment and is formatted as: {technology}_{percentile}_{scaling factor} For instance, “solarPV-430kW_25_2x.zip” reports the experiment using the 25th percentile solar data from the historical 2020 solar PV dataset, scaled to 200%. Scaling factors were applied to the generated solar PV power output files to more closely match the 1.25-MW capacity of the electrolyzer. Each .zip folder contains the following files: A .csv file containing raw data. An .xlsx file explaining all the fields in the raw data. A .png plot showing the time series of hydrogen production, electrolysis power consumption, and solar power input. A PDF file detailing the historical solar data statistical analysis used to generate the solar profile. An experiment labeled “characterization_200.zip” demonstrates the MC250 electrolyzer steady-state response with 30-minute load steps for a total duration of 5 hours. Finally, a .csv file is provided with all experiments combined into one dataset labeled "combined_solarPV_experiments.csv". [1] nelhydrogen.com/product/mc-series-electrolyser . [2] data.nlr.gov/submissions/316 .

08 HYDROGEN↗
DOE OSTI2022

Experimental Unevaluated Nuclear Data List (XUNDL)

Contains nuclear structure and decay data from published articles. Data include nuclear level properties, such as energy, half-life, decay modes, spin and parity, as well as energy and intensity for radiation types such as gammas, alphas, betas and neutrons. For articles published in Physical Review C and the European Physical Journal, compilation takes place during the article submission process. Data from approximately 250 articles are added per year. All the compilation effort is supported by the US Nuclear Data Program and coordinated by the National Nuclear Data Center at Brookhaven National Laboratory, which is also responsible for its web dissemination.

Basic Nuclear Research, nuclear structure and deca↗
DOE OSTI2020

Virtual Metering for Monitoring Building Energy Consumption

The United States Department of Energy (DOE) has standard metering requirements of commercial buildings for optimizing energy performance. The guiding principles are to continuously track and optimize energy performance and install building-level meters for electricity, natural gas, and steam. Some buildings at Los Alamos National Laboratory (LANL) have physical submeters monitoring their energy consumption, but these meters have proven to be unreliable. And, in most cases, replacing them has proven to be a slow process. Installing new submeters also requires a temporary lockout of the circuit on which they are being installed. Many buildings at LANL contain laboratories with ongoing experiments or data centers, which makes an equipment power outage nearly impossible to plan. This inability to plan power outages results in long-term submeter failures. Although most submeters are eventually replaced, failures lead to missing consumption data for some unpredictable, extended time. A building automation system (BAS) is a system that provides control and monitoring on a building to maintain the operational performance of the building and occupancy comfort. Many buildings at LANL currently have a BAS, and all new renovations and installs will include installing a BAS if one does not already exist. The intended purpose for a BAS is primarily to monitor the health and efficiency of a building; however, it is also possible to calculate equipment power and energy consumption using BAS information. This project aims to use virtual meters to monitor building energy consumption as a cost-effective and minimally labor-intensive alternative to installing physical submeters. The fault detection and diagnostics tool, SkySpark, provides a centralized database for all the data from the various BAS that are active at LANL. This data includes the information that is needed to create virtual meters for heating, ventilating, and air conditioning (HVAC) systems in most buildings, including heating and cooling loads. 9 This report begins with a detailed summary of the project, including the reasoning, procedure, and results. The specific processes of creating the various virtual meters are then identified. Then the limitations are discussed. And, lastly, the results and future potential are presented.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2023

CommAwareNet: Towards Communication‐Aware Smart Facilities: Designing an Energy‐Efficient High‐Data‐Rate and Reliable Hybrid THz/VLC Comm. Arch. Reinforced with Intelligent Surfaces for Future Network

This project will develop foundations for energy-efficient, high-data-rate, and reliable communication links for next-generation advanced wireless technologies. The Internet of Everything is emerging as a major technology that can establish a massively connected network among both human-type users and machine-type devices. Particularly, massive machine-type communication is becoming the dominant communication paradigm for numerous emerging areas, including data center networking and wireless backhaul, healthcare (e-health), manufacturing (industry 4.0), utilities, transportation (connected cars and public safety), and virtual or augmented reality for human-machine interaction applications. The rising demand for high data rates, especially in indoor scenarios and densely populated outdoor environments, will eventually overload conventional RF-based technologies. This project aims to conduct fundamental research to design, optimize, and validate a communication-aware, energy-efficient, and reliable network (CommAwareNet) architecture for future smart facilities and environments that require unprecedented high throughput needs for massive numbers of the machine- and human-type users. The scientific and technical merit of this work lies in the research required to bring the proposed transformative CommAwareNet solution into fully functional and practical architecture.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2026

Resonance Self-Shielding: Why it is so Important

This paper is one of a series that I am writing to document my 58 years of experience with ENDF and Neutron Transport calculations, beginning when I worked at the National Nuclear Data Center (NNDC), Brookhaven National Laboratory (BNL), from 1967 to 1972. During those years I was the head of the computer unit of NNDC, assigned to develop computer codes to pre-process, view and test ENDF/B data. Since then, I have continued to support the ENDF effort without any official position or monetary compensation, because I realized how important accurate nuclear data is for use in use in our Engineering applications. It is so important to realize that regardless of how accurate or even perfect our application codes may be to transport particles, without accurate nuclear data we are in a “Garbage In = Garbage Out” situation.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2023

Precursor Analysis Report: Ryuk Ransomware Attack on Universal Health Services 2020

The Ryuk Ransomware Attack on Universal Health Services (UHS) 2020 Precursor Analysis Report leverages publicly available information about the 2020 UHS cyber attack and catalogs anomalous observables for each technique employed in the attack. This analysis is based upon the methodology of the Cybersecurity for the Operational Technology Environment (CyOTE) program. UHS manages over 400 hospitals and is one of the largest healthcare providers in the United States with 3.5 million patients each year. On 27 September 2020, UHS suffered a widespread ransomware attack that resulted in a denial of service to critical internet-dependent healthcare systems including workstations, phones, and data centers. Employees resorted to filing patient details with pen and paper, while other facilities had to redirect ambulances and urgent patients to other facilities for adequate care. Adversaries carried out the attack with Ryuk, a ransomware that encrypts data and generates a RyukReadMe.txt ransom note with the ransom fee to decrypt the data, varying from 15 Bitcoin (BTC) to 50 BTC, equivalent to roughly $\$$353,892 to $\$$964,617. UHS did not pay the ransom and was able to recover data through backups, but still reported an impact of $\$$67 million dollars in recovery costs. On 29 October, one month after the attack, UHS made an official statement that their systems had been restored and they were resuming normal operations. Researchers and analysts identified 18 unique techniques (used in a sequence of 19 steps) utilized during the attack with a total of 185 observables using MITRE ATT&CK® for Industrial Control Systems. The CyOTE program assesses observables accompanying techniques used prior to the triggering event to identify opportunities to detect malicious activity. If observables accompanying the attack techniques are perceived and investigated prior to the triggering event, earlier comprehension of malicious activity can take place. Fourteen of the identified techniques used during the UHS cyber attack were precursors to the triggering event. Analysis identified 106 observables associated with these precursor techniques, 82 of which were assessed to have an increased likelihood of being perceived in the 30 days to two hours preceding the triggering event. The response and comprehension time could have been reduced if the observables had been identified earlier. The information gathered in this report contributes to a library of observables tied to a repository of artifacts, data sources, and technique detection references for practitioners and developers to support the comprehension of indicators of attack. Asset owners and operators can use these products if they experience similar observables or to prepare for comparable scenarios.

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEF↗
DOE OSTI2024

Compilation of a Comprehensive Earthquake Catalog and Relocations in the Caucasus Region

Instrumental seismic monitoring has a long history in the Caucasus and started in 1899 when the first seismograph was installed in Tbilisi, Georgia. Much of the analog paper records from this time period are preserved in the Tbilisi archives because Georgia served as the regional data center. In the 1990s, due to the collapse of the Soviet Union and the political turmoil in the region, the analog networks and the communication between the newly formed national networks deteriorated. In Georgia, for the next 13 yr, the seismic network coverage was poor until the 2002 Tbilisi earthquake. Following this earthquake, the first permanent digital seismic station in Georgia was established in Tbilisi in 2003. The digital era progressively improved the ability to collect and archive data and today more than a hundred broadband seismic stations (including temporary arrays) are operating in the southern Caucasus. Until recently, the region lacked a coordinated effort to catalog all analog and digital era data collected by different countries into a single repository. As a result of collaboration between Lawrence Livermore National Laboratory, the Ilia State University, and the Republican Seismic Survey Center of Azerbaijan, a comprehensive earthquake catalog was compiled for the Caucasus and neighboring areas as part of a broader probabilistic seismic hazard assessment project. Here this project digitized Soviet-era paper bulletins, compiled a unified earthquake catalog from regional bulletins, developed 1D reference velocity model, and used it to relocate the events. The final catalog contains 16,963 events with magnitudes 3.7 and above, bringing together all the available data sets in the Caucasus region from 1900 to 2015, significantly improving locations, and generating the most complete earthquake catalog in the region, temporally and geographically.

58 GEOSCIENCES↗
DOE OSTI2022

Psychrometric Chart v1.0.0

Psychart is a graphical program developed for NERSC for accurately monitoring air conditions for the air cooled HPC equipment. Psychart will live directly in the OMNI system as a Grafana plugin. It employs the use of Psychrolib and the Grafana Starter Panel to create a front-end program that plots the state of the environmental air on a psychrometric chart. It is highly customizable, allowing the user to modify graph bounds, optionally show ASHRAE data center comfort regions, change the series point color, and more. What sets Psychart apart from other existing psychrometric charts is its ability to plot in real-time. Other psychrometric charts that can be found online require manual input of data and do not seamlessly integrate with Grafana.

Ventura, Nicolas↗
DOE OSTI2025

Empirically-calibrated H100 node power models for accurate AI training energy estimation

Accurately quantifying the energy use of artificial intelligence (AI) training is critical for infrastructure planning, carbon accounting, and sustainable data center operation, but few studies have directly measured the power consumption of production workloads on contemporary hardware. By combining empirical measurements from Brookhaven National Laboratory during AI training on 8-graphics-processing-unit H100 systems with open-source benchmarking data, we develop statistical models relating computational intensity to node-level power consumption. We measure the gap between manufacturer-rated thermal design power (TDP) and actual power demand during AI training. Our analysis reveals that even computationally intensive workloads operate at only 76% of the 10.2 kW TDP rating. Our architecture-specific model, calibrated to floating-point operations, predicts energy consumption with 11.4% mean absolute percentage error, significantly outperforming TDP-based approaches (27%–37% error). We identified distinct power signatures between transformer and convolutional neural network architectures, with transformers showing characteristic fluctuations that may impact grid stability. These results provide a measurement-grounded basis for improving AI training energy estimates, enabling more reliable infrastructure sizing, cost projections, and environmental impact assessments.

Newkirk, Alex C↗