Alex C. Newkirk
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Alex C. Newkirk has a total of 28 co-authors. Here are the top co-authors by number of shared publications:
- Arslan Munir
Florida Atlantic University
4 shared publications - Hayat Ullah
Department of Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL 33431, USA
3 shared publications - Xi Yu
Computing and Data Sciences Directorate, Brookhaven National Laboratory, Upton, NY, USA
3 shared publications - Arman Shehabi
Energy Analysis Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, United States of America
2 shared publications - Constantine Samaras
Carnegie Mellon University
2 shared publications Erica Fuchs
Carnegie Mellon University
2 shared publications
Latest Publications
- Open Access
Scalable and Energy-Efficient AI: System-Level Profiling of NVIDIA GPU Clusters for Distributed LLM Training
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.
Open AccessScalable and Energy-Efficient AI: System-Level Profiling of NVIDIA GPU Clusters for Distributed LLM Training
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.
- Top Cited Papers
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (3) , 1-12
Cooling Matters: Benchmarking Large Language Models and Vision-Language Models on Liquid-Cooled Versus Air-Cooled H100 GPU Systems
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (3) , 1-12
Top Cited PapersCooling Matters: Benchmarking Large Language Models and Vision-Language Models on Liquid-Cooled Versus Air-Cooled H100 GPU Systems
- Open Access
- 22 December 2025
- Published by Springer Nature in Research Square
Technoeconomic Analysis of Microgrids for AI Data Centers in the Continental United States
Rapid growth of artificial intelligence demands data centers of unprecedented scale and deployment speed. To power these facilities, firms typically connect to the commercial electric grid or purchase power directly from existing or proposed generators. However, these plans overlook likely delays from grid interconnection and natural gas turbine availability. We compare grid-connected and off-grid data center energy infrastructure (AC-coupled PV+storage, DC-coupled PV+storage, islanded natural gas, and grid baseline) across the continental U.S. We combine models in the literature for location- specific cooling energy use, solar production, and battery degradation and develop data center component-level power and microgrid performance models. To estimate oppor- tunity costs of construction delays, we quantify deployment speed value using GPU spot market prices. We find that DC-coupled microgrids reduce costs 17% (7¢/kWh) on average versus AC counterparts by eliminating inverters and improving efficiency. When valuing deployment speed, all-DC microgrids are lowest-cost in 70%-89% of con- tinental U.S. locations depending on assumptions about natural gas plant construction timelines. These findings suggest that islanded microgrids can serve large data center loads without ratepayer-funded grid expansion, meaning expedited permitting for self- supplied data center power could accelerate AI deployment while preventing residential ratepayers from cross-subsidizing private infrastructure.
- 22 December 2025
- Published by Springer Nature in Research Square
Open AccessTechnoeconomic Analysis of Microgrids for AI Data Centers in the Continental United States
Rapid growth of artificial intelligence demands data centers of unprecedented scale and deployment speed. To power these facilities, firms typically connect to the commercial electric grid or purchase power directly from existing or proposed generators. However, these plans overlook likely delays from grid interconnection and natural gas turbine availability. We compare grid-connected and off-grid data center energy infrastructure (AC-coupled PV+storage, DC-coupled PV+storage, islanded natural gas, and grid baseline) across the continental U.S. We combine models in the literature for location- specific cooling energy use, solar production, and battery degradation and develop data center component-level power and microgrid performance models. To estimate oppor- tunity costs of construction delays, we quantify deployment speed value using GPU spot market prices. We find that DC-coupled microgrids reduce costs 17% (7¢/kWh) on average versus AC counterparts by eliminating inverters and improving efficiency. When valuing deployment speed, all-DC microgrids are lowest-cost in 70%-89% of con- tinental U.S. locations depending on assumptions about natural gas plant construction timelines. These findings suggest that islanded microgrids can serve large data center loads without ratepayer-funded grid expansion, meaning expedited permitting for self- supplied data center power could accelerate AI deployment while preventing residential ratepayers from cross-subsidizing private infrastructure.
- Open Access
- 8 December 2025
- Published by IOP Publishing in Environmental Research: Energy
- Vol. 2 (4) , 045016
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.
- 8 December 2025
- Published by IOP Publishing in Environmental Research: Energy
- Vol. 2 (4) , 045016
Open AccessEmpirically-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.
- Open Access
- 11 August 2025
- Published by Springer Nature in Research Square
Assessment of Vulnerability to Disruptions to the Supply of Semiconductors
While semiconductor supply chain vulnerabilities draw significant attention, the consequences of losing access to specific semiconductor devices remain poorly understood. We combine structured interviews with 21 semiconductor experts across six sectors with component origin analysis of an FPV drone and smartphone to assess for which devices and process nodes loss of access would be most catastrophic. Experts indicated that existing infrastructure—grid, data centers, defense, transportation—would maintain operations for years due to semiconductor durability and stockpiling. Instead, the primary impact would be the inability to produce new products and expand infrastructure. All expert-proposed substitution strategies required 18-36 months to execute, suggesting that reducing these constraints may be important to improve crisis response and longer-term resilience. These findings suggest industry and policymakers should evaluate the costs and benefits of strategies to accelerate substitution, such as design precertification, component standardization, or modular architectures.
- 11 August 2025
- Published by Springer Nature in Research Square
Open AccessAssessment of Vulnerability to Disruptions to the Supply of Semiconductors
While semiconductor supply chain vulnerabilities draw significant attention, the consequences of losing access to specific semiconductor devices remain poorly understood. We combine structured interviews with 21 semiconductor experts across six sectors with component origin analysis of an FPV drone and smartphone to assess for which devices and process nodes loss of access would be most catastrophic. Experts indicated that existing infrastructure—grid, data centers, defense, transportation—would maintain operations for years due to semiconductor durability and stockpiling. Instead, the primary impact would be the inability to produce new products and expand infrastructure. All expert-proposed substitution strategies required 18-36 months to execute, suggesting that reducing these constraints may be important to improve crisis response and longer-term resilience. These findings suggest industry and policymakers should evaluate the costs and benefits of strategies to accelerate substitution, such as design precertification, component standardization, or modular architectures.
- Open Access
Cooling Matters: Benchmarking Large Language Models and Vision-Language Models on Liquid-Cooled Versus Air-Cooled H100 GPU Systems
The unprecedented growth in artificial intelligence (AI) workloads, recently dominated by large language models (LLMs) and vision-language models (VLMs), has intensified power and cooling demands in data centers. This study benchmarks LLMs and VLMs on two HGX nodes, each with 8x NVIDIA H100 graphics processing units (GPUs), using liquid and air cooling. Leveraging GPU Burn, Weights and Biases, and IPMItool, we collect detailed thermal, power, and computation data. Results show that the liquid-cooled systems maintain GPU temperatures between 41-50 degrees Celsius, while the air-cooled counterparts fluctuate between 54-72 degrees Celsius under load. This thermal stability of liquid-cooled systems yields 17 percent higher performance (54 TFLOPs per GPU vs. 46 TFLOPs per GPU), improved performance per watt, reduced energy overhead, and greater system efficiency than the air-cooled counterparts. These findings underscore the energy and sustainability benefits of liquid cooling, offering a compelling path forward for hyperscale data centers s
Open AccessCooling Matters: Benchmarking Large Language Models and Vision-Language Models on Liquid-Cooled Versus Air-Cooled H100 GPU Systems
The unprecedented growth in artificial intelligence (AI) workloads, recently dominated by large language models (LLMs) and vision-language models (VLMs), has intensified power and cooling demands in data centers. This study benchmarks LLMs and VLMs on two HGX nodes, each with 8x NVIDIA H100 graphics processing units (GPUs), using liquid and air cooling. Leveraging GPU Burn, Weights and Biases, and IPMItool, we collect detailed thermal, power, and computation data. Results show that the liquid-cooled systems maintain GPU temperatures between 41-50 degrees Celsius, while the air-cooled counterparts fluctuate between 54-72 degrees Celsius under load. This thermal stability of liquid-cooled systems yields 17 percent higher performance (54 TFLOPs per GPU vs. 46 TFLOPs per GPU), improved performance per watt, reduced energy overhead, and greater system efficiency than the air-cooled counterparts. These findings underscore the energy and sustainability benefits of liquid cooling, offering a compelling path forward for hyperscale data centers s
- Top Cited PapersOpen Access
- 25 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 61740-61747
Single-Node Power Demand During AI Training: Measurements on an 8-GPU NVIDIA H100 System
The expansion of artificial intelligence (AI) applications has driven substantial investment in computational infrastructure, especially by cloud computing providers. Quantifying the energy footprint of this infrastructure requires models parameterized by the power demand of AI hardware during training. In this work, we measured the instantaneous power draw of an 8-GPU NVIDIA H100 HGX node during the training of open-source image classifier (ResNet) and large-language models (Llama2-13b).We characterize power demand for a single node configuration, providing foundational data for future multi-node studies. The maximum observed power draw was approximately 8.4 kW, 18% lower than the manufacturer-rated 10.2 kW, even with GPUs near full utilization. Holding model architecture constant, increasing batch size from 512 to 4096 images for ResNet reduced total training energy consumption by a factor of 4. These findings can inform capacity planning for data center operators and energy use estimates by researchers. Future work will investigate the impact of cooling technology and carbon-aware scheduling on AI workload energy consumption.
- 25 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 61740-61747
Top Cited PapersOpen AccessSingle-Node Power Demand During AI Training: Measurements on an 8-GPU NVIDIA H100 System
The expansion of artificial intelligence (AI) applications has driven substantial investment in computational infrastructure, especially by cloud computing providers. Quantifying the energy footprint of this infrastructure requires models parameterized by the power demand of AI hardware during training. In this work, we measured the instantaneous power draw of an 8-GPU NVIDIA H100 HGX node during the training of open-source image classifier (ResNet) and large-language models (Llama2-13b).We characterize power demand for a single node configuration, providing foundational data for future multi-node studies. The maximum observed power draw was approximately 8.4 kW, 18% lower than the manufacturer-rated 10.2 kW, even with GPUs near full utilization. Holding model architecture constant, increasing batch size from 512 to 4096 images for ResNet reduced total training energy consumption by a factor of 4. These findings can inform capacity planning for data center operators and energy use estimates by researchers. Future work will investigate the impact of cooling technology and carbon-aware scheduling on AI workload energy consumption.
- Open Access
- 17 July 2024
- Published by Springer Nature in Energy Efficiency
- Vol. 17 (6) , 1-20
Expert and operator perspectives on barriers to energy efficiency in data centers
It was last estimated in 2016 that data centers (DCs) comprise approximately 2% of total US electricity consumption. However, this estimate is currently being updated to account for the massive increase in computing needs due to streaming, cryptocurrency, and artificial intelligence (AI). To prevent energy consumption that tracks with increasing computing needs, it is imperative we identify energy efficiency strategies and investments beyond the low-hanging fruit solutions. In a two-phased research approach, we ask: What non-technical barriers still impede energy efficiency (EE) practices and investments in the data center sector, and what can be done to overcome these barriers? In particular, we are focused on social and organizational barriers to EE. In Phase I, we performed a literature review and found that technical solutions are abundant in the literature, but fail to address the top-down cultural shifts that need to take place in order to adapt new energy efficiency strategies. In Phase II, reported here, we interviewed 16 data center operators/experts to ground-truth our literature findings. Our interview protocols focus on three aspects of DC decision-making: procurement practices, metrics and monitoring, and perceived barriers to energy efficiency. We find that vendors are the key drivers of procurement decisions, advanced efficiency metrics are facility-specific, and there is convergence in the design of advanced facilities due to the heat density of parallelized infrastructure. Our ultimate goals for our research are to design DC decarbonization policies that target organizational structure, empower individual staff, and foster a supportive external market.
- 17 July 2024
- Published by Springer Nature in Energy Efficiency
- Vol. 17 (6) , 1-20
Open AccessExpert and operator perspectives on barriers to energy efficiency in data centers
It was last estimated in 2016 that data centers (DCs) comprise approximately 2% of total US electricity consumption. However, this estimate is currently being updated to account for the massive increase in computing needs due to streaming, cryptocurrency, and artificial intelligence (AI). To prevent energy consumption that tracks with increasing computing needs, it is imperative we identify energy efficiency strategies and investments beyond the low-hanging fruit solutions. In a two-phased research approach, we ask: What non-technical barriers still impede energy efficiency (EE) practices and investments in the data center sector, and what can be done to overcome these barriers? In particular, we are focused on social and organizational barriers to EE. In Phase I, we performed a literature review and found that technical solutions are abundant in the literature, but fail to address the top-down cultural shifts that need to take place in order to adapt new energy efficiency strategies. In Phase II, reported here, we interviewed 16 data center operators/experts to ground-truth our literature findings. Our interview protocols focus on three aspects of DC decision-making: procurement practices, metrics and monitoring, and perceived barriers to energy efficiency. We find that vendors are the key drivers of procurement decisions, advanced efficiency metrics are facility-specific, and there is convergence in the design of advanced facilities due to the heat density of parallelized infrastructure. Our ultimate goals for our research are to design DC decarbonization policies that target organizational structure, empower individual staff, and foster a supportive external market.
- Open Access
- 10 January 2023
- Published by Springer Nature in Energy Efficiency
- Vol. 16 (1) , 1-18
Organizational and psychological measures for data center energy efficiency: barriers and mitigation strategies
It was last estimated that in 2020, data centers comprised approximately 2% of total US electricity consumption, with an estimated annual growth rate of 4%. As our country increasingly relies on information technology (IT), our data centers (DCs) will need to increase their energy efficiency (EE) to stabilize their energy consumption. The task of studying EE in DCs is complicated by the interconnected nature of humans and mission-critical technical systems. Moreover, the literature tends to focus on technology solutions such as improvements to IT equipment, cooling infrastructure, and software, without addressing organizational and psychological drivers. Our research demystifies the complex interactions between humans and DCs, by asking What non-technical barriers impede EE investment decision-making and/or implementing energy management strategies? To begin to answer this question, we perform a literature review of 86 resources, ranging from peer-reviewed journal publications to handbooks. We also consider related fields such as organizational behavioral management and energy intensive buildings. We develop a public Zotero library, perform content coding, and complete a rudimentary network analysis. Our findings from the literature review suggest that (1) technological solutions are abundant in the literature but fall short of providing practical guidance on the pitfalls of implementation, (2) making energy efficiency a priority at the executive level of organizations will be largely ineffective if the IT and facilities staff are not directly incentivized to increase EE, and (3) there is minimal current understanding of how the individual psychologies of IT and facilities staff affect EE implementation in DCs. In the next phase of our research, we plan to interview data center operators/experts to ground-truth our literature findings and collaboratively design decarbonization policy solutions that target organizational structure, empower individual staff, and foster a supportive external market.
- 10 January 2023
- Published by Springer Nature in Energy Efficiency
- Vol. 16 (1) , 1-18
Open AccessOrganizational and psychological measures for data center energy efficiency: barriers and mitigation strategies
It was last estimated that in 2020, data centers comprised approximately 2% of total US electricity consumption, with an estimated annual growth rate of 4%. As our country increasingly relies on information technology (IT), our data centers (DCs) will need to increase their energy efficiency (EE) to stabilize their energy consumption. The task of studying EE in DCs is complicated by the interconnected nature of humans and mission-critical technical systems. Moreover, the literature tends to focus on technology solutions such as improvements to IT equipment, cooling infrastructure, and software, without addressing organizational and psychological drivers. Our research demystifies the complex interactions between humans and DCs, by asking What non-technical barriers impede EE investment decision-making and/or implementing energy management strategies? To begin to answer this question, we perform a literature review of 86 resources, ranging from peer-reviewed journal publications to handbooks. We also consider related fields such as organizational behavioral management and energy intensive buildings. We develop a public Zotero library, perform content coding, and complete a rudimentary network analysis. Our findings from the literature review suggest that (1) technological solutions are abundant in the literature but fall short of providing practical guidance on the pitfalls of implementation, (2) making energy efficiency a priority at the executive level of organizations will be largely ineffective if the IT and facilities staff are not directly incentivized to increase EE, and (3) there is minimal current understanding of how the individual psychologies of IT and facilities staff affect EE implementation in DCs. In the next phase of our research, we plan to interview data center operators/experts to ground-truth our literature findings and collaboratively design decarbonization policy solutions that target organizational structure, empower individual staff, and foster a supportive external market.
- Open Access
- 1 June 2021
- Published by Office of Scientific and Technical Information (OSTI)
U.S. Industrial and Commercial Motor System Market Assessment Report Volume 2: Advanced Motors and Drives Supply Chain Review
Electric motors in commercial and industrial applications account for almost a third of total electric grid load in the United States. With these applications representing such a significant level of energy use, increasing efficiency in these applications can help substantially increase resilience and lower total energy costs. This report examines stationary high-efficiency electric motor technologies and their supply chains.
- 1 June 2021
- Published by Office of Scientific and Technical Information (OSTI)
Open AccessU.S. Industrial and Commercial Motor System Market Assessment Report Volume 2: Advanced Motors and Drives Supply Chain Review
Electric motors in commercial and industrial applications account for almost a third of total electric grid load in the United States. With these applications representing such a significant level of energy use, increasing efficiency in these applications can help substantially increase resilience and lower total energy costs. This report examines stationary high-efficiency electric motor technologies and their supply chains.