OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION

OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION Aug 2025 AUTHOR: Sana Babayan Vanestan Sharif University of Technology (SUT) SUPERVISORS: Natalia Diana Szczepanek (IT-TC-LCG) Robin Hofsaess (IT-TC-LCG) Ewoud Ketele (IT-TC-LCG) CERN openlab Report 08/2025 PROJECT SPECIFICATION This project aims to investigate the relationship between performance, power consumption, and stability of GPUs under varying power limit and frequency configurations, using High-Energy Physics (HEP) workloads. While CPUs have been extensively studied for performance and energy efficiency across different clock frequencies, similar systematic investigations for GPUs remain limited. Standard power limit settings guarantee stable operation across all units of the same GPU model; however, many GPUs can reliably operate at lower power limits, reducing power consumption without significantly compromising performance. Leveraging access to a range of HEP-related GPU workloads, we will conduct a controlled study on multiple GPUs. The study will quantify the effects of different power and frequency limit configurations on performance, energy usage, and stability. These results will contribute to the understanding of possible optimization strategies for the WLCG. Given the increasing computational demands of the upcoming High-Luminosity LHC, investigating GPU configurations with the best performance-to-energy balance is key for environmentally sustainable scientific computing. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 1 CERN openlab Report 08/2025 ABSTRACT The increasing computational demands of the High-Luminosity Large Hadron Collider (HL-LHC) pose significant challenges for the energy efficiency of the WLCG’s compute infrastructure. Enhancing power efficiency is essential for achieving more environmentally sustainable scientific computing. Since GPUs can play a central role in meeting these demands, their default power and frequency settings are not necessarily optimal for maximizing the ratio of computational performance to energy consumed. This study explores how tuning GPU power limits and operating frequencies can improve the performance-to-energy ratio for High-Energy Physics (HEP) workloads. We ran representative HEP workloads on NVIDIA A100, V100, and V100S GPUs under various power and frequency configurations measuring their effects on performance, energy consumption, and efficiency. Our results show that running GPUs at maximum power or frequency does not necessarily deliver the best performance-to-energy ratio. In some cases, lowering these settings improves the performanceto-energy ratio, showing that a reduction in power draw does not always translate into a proportional loss in performance. These findings illustrate the potential for systematic tuning of GPU operating parameters to reduce the energy consumption and carbon footprint of large-scale scientific computing. When applied at the scale of the WLCG, even modest improvements in the performance-to-energy ratio could lead to significant reductions in energy usage and environmental impact, particularly as the number of GPUs utilized in High-Energy Physics computing is expected to grow in the future. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 2 CERN openlab Report 08/2025 ACKNOWLEDGMENTS I would like to express my sincere gratitude to the HEPiX Benchmarking Working Group for their valuable support and contributions to this work. I am especially thankful to my head supervisor, Domenico Giordano, for his continuous guidance and encouragement throughout the project. I also extend my appreciation to my supervisors, Natalia Diana Szczepanek, Robin Hofsaess, and Ewoud Ketele, for their invaluable advice and support, and to Markus Schulz for his insightful suggestions and for helping to steer the research in the right direction. Finally, I would like to thank the CERN Summer Student Openlab Programme [10] for providing me with the opportunity to carry out this research in such an inspiring environment. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 3 CERN openlab Report 08/2025 TABLE OF CONTENTS 1 Introduction 5 2 Background 5 2.1 GPU Performance and Energy Efficiency Concepts . . . . . . . . . . . . . . . . . . 5 2.2 Benchmarking and Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . 6 3 Related Work 7 4 Methodology 8 5 Implementation / Experimental Setup 9 5.1 Hardware Platforms Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 5.1.1 Power and Frequency Limits . . . . . . . . . . . . . . . . . . . . . . . . . . 9 5.2 Benchmark Configuration and Execution . . . . . . . . . . . . . . . . . . . . . . . . 10 5.3 Automation and Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 5.4 Post-Processing..................................... 11 6 Results and Analysis 11 6.1 PowerLimit....................................... 11 6.2 FrequencyLimit..................................... 17 7 Discussion and Conclusion 18 8 Future Work 18 9 Appendix 18 9.1 BenchmarkExecutionSteps .............................. 18 10 References 19 OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 4 CERN openlab Report 08/2025 1 Introduction The Large Hadron Collider (LHC) at CERN generates enormous volumes of data, requiring substantial computing resources for processing and analysis. The Worldwide LHC Computing Grid (WLCG) is a global distributed computing infrastructure that brings together around 160 computing centers across more than 40 countries. Its mission is to store, distribute, and analyze the roughly 200 Petabytes of data expected from LHC operations each year [8]. CERN itself provides about 20% of the total WLCG resources [7]. With the upcoming High-Luminosity LHC (HL-LHC) upgrade, the average number of simultaneous proton–proton interactions per bunch crossing (pile-up) is expected to climb to up to 200, significantly raising detector occupancy and hit multiplicities. These conditions increase the volume of raw data and also exacerbate combinatorial ambiguity in pattern recognition, making track and event reconstruction far more computationally intensive [28]. This surge in computational demand intensifies the need for sustainable computing practices within the WLCG, where even small efficiency gains can have a substantial cumulative effect due to the system’s scale. Graphics Processing Units (GPUs) are strong candidates to meet these growing computational requirements thanks to their massive parallelism and high throughput. However, they are typically operated at maximum frequency and power limits, which does not always correspond to optimal energy efficiency. In high-performance computing (HPC) and AI workloads, the traditional focus on minimizing time-to-solution [4] often overlooks the trade-off between runtime and power consumption. Since total energy usage depends on both, achieving true efficiency requires a balanced consideration of performance, power, and energy. Optimizing GPU energy efficiency is non-trivial: lowering clock frequencies may reduce power draw but increase execution time, potentially negating the savings. Moreover, efficient operation must consider not only GPU behavior but also system-level and application-level factors, from cooling and server infrastructure to algorithmic and scheduling choices. Addressing these challenges is crucial to reduce operational costs, limit environmental impact, and enable sustainable HPC and AI computing. This study investigates how tuning GPU operating parameters can reduce power consumption, improve sustainability, and lower the CO 2 footprint of large-scale scientific computing, aligning with CERN’s environmental goals (Figure 1). We focus on three GPU models, NVIDIA A100, V100, and V100S, and evaluate multiple configurations to understand their impact on both performance and energy efficiency. The primary objective is to identify configurations that maximize performance per unit energy while maintaining acceptable throughput levels. The remainder of this report is structured as follows: Section 2provides the necessary background on GPU architecture and energy-efficiency concepts. Section 3reviews related work in energy-efficient GPU and HPC research. Section 4outlines the methodology adopted in this study, while Section 5 details the implementation and experimental setup used to evaluate different GPU configurations. Section 6presents and analyzes the results, and Section 7discusses the findings and their broader implications for sustainable high-performance computing. Section 8highlights potential directions for future work. The Appendix (Section 9) contains additional data, measurement details, and configuration tables that complement the main text. Section 10 lists the references used in this study. 2 Background 2.1 GPU Performance and Energy Efficiency Concepts High-Energy Physics (HEP) experiments, such as those conducted at CERN, traditionally mainly rely for their computations on Central Processing Units (CPUs), which have been optimized for low-latency, OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 5 CERN openlab Report 08/2025 Figure 1: CERN’s commitment to environmentally responsible research, integrating sustainability, energy efficiency, and innovative technologies to minimize its environmental footprint [5]. serialized task execution. While CPUs excel at quickly completing individual tasks, their throughput is limited when handling large-scale parallel workloads. In contrast, GPUs are designed for throughput-oriented computing and massive parallelism. Equipped with thousands of cores capable of executing many operations simultaneously, GPUs are well-suited for HEP workloads that involve numerous independent calculations running concurrently. The clock speed of a GPU, measured in MHz or GHz, determines the rate at which instructions are executed. Modern GPUs employ dynamic frequency scaling and power-saving features to maintain energy efficiency. Techniques such as adaptive clock boosting allow GPUs to adjust their operating frequency in response to workload intensity and thermal conditions, optimizing the balance between performance and energy consumption [19]. Of course, GPU energy efficiency depends not only on raw computational power but also on workload-specific configurations. However, adjusting GPU core and memory clocks and setting static power limits via management tools like NVIDIA’s nvidia-smi [23] or NVML [22] can influence energy consumption. Evaluating throughput (events per second) alongside throughput per watt provides a robust metric to assess both performance and energy efficiency of an HEP application. 2.2 Benchmarking and Performance Evaluation Benchmarking plays a vital role in quantifying and comparing compute performance in a consistent, reproducible manner. For this purpose, the HEPiX Benchmarking Working Group [16] defines and maintains standardized benchmarks for performance assessments within the WLCG. Notably, HEPScore23 (HS23), which replaced the older HS06 metric in 2023 [11], is tailored to the WLCG, as it provides representative benchmark results by invoking real-world HEP application as executed payloads. The HEP Benchmark Suite [15], shown in Figure 2, is a modular toolkit developed at CERN that orchestrates multiple benchmarks across diverse hardware platforms and collects comprehensive system metadata and performance and utilization indicators enabling verbose performance studies OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 6 CERN openlab Report 08/2025 and a direct comparisons between data centers. The suite supports the execution of a variety of benchmarks including HEPScore23, HS06, and SPEC CPU2017 [26]. Benchmark results are integrated into CERN’s monitoring systems. This information supports informed decisions on procurement of resources and furthermore enables decisions on the optimization of GPU power and frequency settings to improve energy efficiency. Figure 2: High-level architecture of the HEP Benchmark Suite [15]. The suite is organized into three main components: (i) Plugins, which enable services such as hardware metadata collection and connectivity to monitoring tools; (ii) Run Logic, which configures and executes benchmarks; and (iii) Data Processing, which validates results, collects logs, generates reports, and publishes them. The suite supports multiple benchmarks including HS06, SPEC CPU2017, and HEPscore for CPUs and GPUs, providing a modular and extensible framework for performance evaluation. 3 Related Work GPU power efficiency is of course relevant in many fields and also has been evaluated in different scientific domains. Training deep neural networks (DNNs) is increasingly energy-intensive, yet most prior work focuses on reducing training time without optimizing energy use. You et al. [32] demonstrate a tradeoff between performance and energy consumption, proposing an online framework that automatically tunes batch size and GPU power limits to improve energy efficiency by up to 75.8%. Their approach adapts dynamically via just-in-time energy profiling, unlike prior offline methods. They also highlight that limiting GPU power can reduce energy use significantly without sacrificing performance. You et al. [32] complements existing research on energy optimization in DNN training by focusing on end-to-end energy-to-accuracy metrics with fully online optimization. Schoonhoven et al. [25] introduce energy monitoring and optimization features within Kernel Tuner, a generic GPU autotuning tool, to explore tuning for energy efficiency versus execution time. Their model for GPU power consumption effectively narrows the search space by estimating optimal clock frequencies. Experiments show that tuning clock frequency alongside kernel parameters can improve energy efficiency by over 40%, often with moderate performance trade-offs. Compared to power capping, frequency tuning offers finer control over power consumption and enables larger energy OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 7 CERN openlab Report 08/2025 savings. This model-guided approach reduces the autotuning search space by up to 82%, facilitating more efficient energy-aware GPU optimizations. Wang et al. [31] propose power consumption models for General-Purpose GPU (GPGPU) applications based on program slicing, which decomposes programs into smaller units to analyze energy use at a fine granularity. Their models consider factors like computation intensity and the number of active streaming multiprocessors (SMs), distinguishing between sparse and dense branching programs. With an average prediction error below 6%, their approach provides application programmers with a practical tool to estimate and profile power consumption, enabling targeted energy optimizations. This method offers valuable insights into program-level power behavior and aims to be extended for different GPU architectures in future work. Tang et al. [27] empirically study the effect of GPU Dynamic Voltage and Frequency Scaling (DVFS) on the energy consumption and performance of deep neural network (DNN) training and inference across various GPU architectures. The results show that optimizing core frequency can reduce energy use by up to 23.1% during training and 26.4% during inference, while improving performance by as much as 33%. These findings highlight the potential of GPU DVFS as an effective approach for enhancing energy efficiency in DNN workloads without significant performance loss. The study also suggests future work on layer-wise DVFS schemes and system-level scheduling to further optimize energy consumption in multi-task GPU environments. Bharadwaj et al. [3] propose DUB, a technique combining dynamic underclocking and bypassing in the GPU Network-on-Chip (NoC) to optimize power savings for heterogeneous workloads. They identify that GPU kernel performance is highly sensitive to NoC clock frequencies, which vary during execution phases, a nuance poorly addressed by traditional DVFS. DUB dynamically reduces NoC frequency and bypasses retimer flops and routers during low-load, latency-sensitive phases, achieving 26% power savings with only a 3% performance loss. This outperforms conventional DVFS methods, which typically incur higher performance degradation for less power savings. Although most of the introduced approaches are not feasible for most of the LHC use cases, they underline the relevance of the topic and are representative for the ongoing research in the field. 4 Methodology The present study was conducted using a preliminary HEPScore GPU benchmark, which provides a rating of the GPU performance for a given hardware. For the experiments, the preliminary CMS-FlowSim-GPU workload [29] was selected. FlowSim is an experimental GPU workload designed for the CMS experiment at CERN. It is based on deep learning methods and mirrors the inference step of simulating the event-driven process from particle generation to detector signal reconstruction (CMS FlashSim) [13,30]. This workload is a candidate for potential HEP applications that can utilize GPUs on larger scales. The results of the benchmark are collected in JSON files produced by the Suite. These files contain detailed runtime measurements, including system load, clock frequency, and power consumption for both CPU and GPU components. From these measurements, the throughput reported by the benchmark was extracted, and the throughput-per-watt metric was computed from the measured (GPU) power draw for further analysis: • Throughput (events/s): The number of simulated events per second, serving as the primary indicator of performance. • Throughput per watt: The number of events processed per unit of GPU core energy consumed, reflecting the energy efficiency of execution. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 8 CERN openlab Report 08/2025 Figure 6: Throughput per watt of V100S, A100, and V100 GPUs across power limits. All GPUs show peak efficiency at intermediate caps, with improvements of up to 31% compared to maximum power settings. This shows that efficiency optimizations based on a reduced power limit are possible. A further observation is that the V100 and V100S GPUs display smooth, gradual throughput loss with decreasing power limits, whereas the A100 shows slightly steeper losses at lower caps (around 150–175 W). Despite these differences, all GPUs follow a qualitatively similar trajectory: more strict power limits lead to disproportionately smaller reductions in throughput, thereby improving overall energy efficiency. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 15 CERN openlab Report 08/2025 Figure 7: Normalized throughput loss vs. normalized imposed power limit. Throughput decreases as the power cap is lowered, but the reduction is consistently smaller than the relative decrease in power budget, resulting in a potential for efficiency optimizations. Figure 8: Normalized throughput loss vs. normalized consumed power. Across all GPUs, reductions in consumed power significantly outweigh the associated performance penalties, highlighting improved energy efficiency at lower caps. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 16 CERN openlab Report 08/2025 6.2 Frequency Limit Figure 9shows the throughput-per-watt ratio for the V100S and A100 GPUs under different enforced frequency limits, measured both on bare metal and within a VM. The results consistently demonstrate that the highest energy efficiency is not achieved at maximum frequency but rather near the 1 GHz range. For the A100 on BM, efficiency increases steadily until peaking at around 1035MHz, after which it begins to decline. Operating at this peak yields roughly a 21.92% improvement in throughput per watt compared to running at its maximum frequency of 1410MHz. The V100S on VM follows a similar trend, with peak efficiency occurring near 975MHz, corresponding to an improvement of about 30.52% relative to maximum-frequency operation at 1597MHz. The V100 on BM also achieves its best efficiency around the 975MHz range with approximately 21.07% improvement in throughput-per-watt compared to running at its maximum frequency of 1530MHz, though at a lower absolute level compared to the A100 and V100S, highlighting differences in architectural efficiency. These findings emphasize that modest reductions from maximum clock speeds can deliver substantial energy-efficiency gains. While higher frequencies continue to improve raw throughput, the power cost rises disproportionately beyond the efficiency peak. Conversely, at very low frequencies, throughput losses outweigh power savings. Overall, operating GPUs near the 1 GHz range provides the most sustainable balance between performance and energy consumption, underlining that maximum performance does not coincide with maximum efficiency. Figure 9: Throughput per watt as a function of frequency limit for V100S (VM), V100 (BM), and A100 (BM). All three configurations achieve peak efficiency near 1 GHz, with improvements of nearly 30% compared to maximum-frequency operation. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 17 CERN openlab Report 08/2025 7 Discussion and Conclusion The experiments demonstrate that GPU performance and energy consumption are closely linked and can be effectively influenced through power and frequency limits. Reducing the power budget lowers throughput, but the normalized analysis shows that performance losses are consistently smaller than the corresponding reductions in power draw. For example, operating the GPUs at 60 to 70% of the maximum power budget reduced throughput by only 10 to 15%, while lowering power consumption by roughly 30 to 40%. This asymmetry indicates that significant energy savings can be achieved with relatively modest performance penalties. A similar trend was observed with frequency scaling. Although the reduced clock speeds resulted in lower throughput, they also led to a disproportionately large reduction in power consumption compared to higher frequency limits. Across all tested GPUs, the most energy-efficient operating points occurred near 1 GHz improving the throughput per watt by 20 to 30% compared to maximum-frequency operation. This shows that an adaption of the frequency limits can lead to an increased power efficiency for the studied workload. Taken together, these results show that GPU utilization can be optimized for a more sustainable computing when operating them under moderate limits. Although the exact efficiency peaks vary across architectures and environments, careful tuning of these parameters can yield meaningful energy savings with moderate performance trade-offs. In cases where short latencies, and therefore the maximum achievable performance, are less important, this approach therefore offers good potential for more environmentally sustainable computing. However, two further aspects have to be taken into account: First, the investigated throughput per watt only focuses on the power draw of the GPU. Here, it has to be considered as well that a reduced peak performance leads to a longer runtime and the power savings can easily be absorbed by the machine overhead when calculating in the extended runtime. And second, GPU applications are very heterogeneous. This means that the presented results must not be seen as general applicable rules for an increased energy efficiency, but for each and every (HEP) GPU application, an individual optimum has to be found. In conclusion, the results are very promising as they show that under given circumstances a remarkable energy efficiency with simple optimizations can be achieved, enabling a more environmentally sustainable computing. 8 Future Work Future work may extend these findings in several directions. Different workloads can be explored to test to what extent the observed trade-offs generalize across different HEP applications. Running longer workloads and performing additional repeated tests can improve the precision and reliability of the measurements. Additional GPU models and system configurations can be evaluated to capture architectural variations. Furthermore, multi-dimensional optimization can be investigated by combining power, frequency, and other tunable parameters, with the aim of identifying operating regions that maximize efficiency while keeping throughput at a sufficient level. 9 Appendix 9.1 Benchmark Execution Steps The main steps performed by run_HEPscore.sh include: OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 18 CERN openlab Report 08/2025 1. Parsing of command-line arguments (site name, executor type, GPU count, module selection, etc.). 2. Validation of parameters (e.g., supported versions). 3. Creation of working directories and configuration files (bmkrun_config.yml). 4. Installation of the suite via a Python virtual environment. 5. Execution of workloads through the bmkrun command. 6. Storage of results and logs (with optional publication to AMQ). An example command for executing a GPU workload is: ./run_HEPscore.sh -s CERN -x v2.0 -b all -g 1 where: •-s specifies the site name, •-x selects the HEPSCORE version, •-b selects the plugin set (“all” for all available plugins), •-g specifies the number of GPUs to be tracked. 10 References [1] Atumation bash script for testing different frequency limits.url: https://gitlab.cern. ch / hep - benchmarks / hep - benchmark - studies/ - /blob / BMK - 1718 / analysis / hepscore23/projects/summer_student_2025/Automation/A100/auto_freq.sh (visited on 09/18/2025). [2] Atumation bash script for testing different power limits.url: https://gitlab.cern.ch/hepbenchmarks/hep-benchmarkstudies/-/blob/BMK1718/analysis/hepscore23/ projects/summer_student_2025/Automation/A100/power_benchmark_runner.sh (visited on 09/18/2025). [3] Srikant Bharadwaj et al. “DUB: Dynamic underclocking and bypassing in nocs for heterogeneous GPU workloads”. In: Proceedings of the 15th IEEE/ACM International Symposium on Networkson-Chip. 2021, pp. 49–54. [4] CERN. Energy and Power Efficiency for Applications on the Latest NVIDIA Technology. https://www.nvidia.com/en-us/on-demand/session/gtc24-s62419/ . Accessed: 2025-08-15. [5] CERN. Environmentally responsible research. https://home.cern/about/whatwedo/environmentally-responsible-research. Accessed: 2025-08-14. [6] CERN. Results uploaded to cern gitlab. https://gitlab.cern.ch/hep-benchmarks/hepbenchmark-studies/-/tree/BMK-1718/analysis/hepscore23/projects/summer_ student_2025/Analyses. Accessed: 2025-08-15. [7] CERN. Worldwide LHC Computing Grid (WLCG). https : / / home . cern / science / computing/grid. Accessed: 2025-08-14. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 19 CERN openlab Report 08/2025 [8] CERN Data Centre passes the 200-petabyte milestone.url: https://home.cern/news/ news/computing/cern-data-centre-passes-200-petabyte-milestone (visited on 10/11/2025). [9] CERN openlab.url:https://openlab.cern/ (visited on 09/18/2025). [10] CERN openlab summer student programme.url: https://openlab.cern/cern-openlabsummer-student-programme/ (visited on 09/18/2025). [11] D. Giordano CERN/IT). HEPScore benchmark status.url: https://indico.cern.ch/ event/1225116/contributions/5519006/attachments/2713539/4712490/GDB-1309-2023-giordano.pdf (visited on 09/18/2025). [12] CMS FlowSim GPU workload gitlab.url: https://gitlab.cern.ch/hep-benchmarks/ hep-workloads/container_registry/24351 (visited on 09/18/2025). [13] FlowSim.url: https://github.com/francesco-vaselli/FlowSim/tree/benchmark (visited on 11/10/2025). [14] Google Cloud. GPU models in Compute Engine. https://cloud.google.com/compute/ docs/gpus#performance_comparison_chart. Accessed: 2025-08-19. 2025. [15] HEP Benchmark Suite. https://gitlab.cern.ch/hep-benchmarks/hep-benchmarksuite. Accessed: 2025-08-14. [16] HEPiX Benchmarking Working Group. https://w3.hepix.org/benchmarking.html . Accessed: 2025-08-14. [17] Link to the repository providing hepscore-gpu.yaml configuration file. https://gitlab. cern.ch/hep-benchmarks/hep-score/-/blob/BMK-1646/hepscore/etc/hepscoregpu.yaml. Accessed: 2025-08-19. 2025. [18] Link to the repository providing run 𝐻𝐸 𝑃𝑠𝑐𝑜𝑟𝑒.𝑠ℎ. . https : / / gitlab . cern . ch / hep - benchmarks/hep-benchmark-suite/-/tree/GPU-additions/examples/hepscore . Accessed: 2025-08-19. 2025. [19] Maximizing Energy and Power Efficiency in Applications with NVIDIA GPUs. https:// developer.nvidia.com/blog/maximizingenergyandpowerefficiencyinapplications-with-nvidia-gpus/. Accessed: 2025-08-14. [20] NVIDIA Ampere Architecture.url: https://www.nvidia.com/en-us/data-center/ ampere-architecture/ (visited on 09/18/2025). [21] NVIDIA Volta Architecture.url: https://www.nvidia.com/en-us/data-center/voltagpu-architecture/ (visited on 09/18/2025). [22] Nvidia-nvml. https://developer.nvidia.com/management-library-nvml . Accessed: 2025-08-14. [23] Nvidia-smi. https://docs.nvidia.com/deploy/nvidia-smi/index.html . Accessed: 2025-08-14. [24] OpenStack: Open Source Cloud Computing Infrastructure.url: https://www.openstack. org/ (visited on 09/18/2025). [25] Richard Schoonhoven et al. “Going green: optimizing GPUs for energy efficiency through model-steered auto-tuning”. In: 2022 IEEE/ACM International Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS). IEEE. 2022, pp. 48–59. [26] SPEC CPU 2017 benchmark.url: https : / / www . spec . org / cpu2017/ (visited on 10/11/2025). OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 20 CERN openlab Report 08/2025 [27] Zhenheng Tang et al. “The impact of GPU DVFS on the energy and performance of deep learning: An empirical study”. In: Proceedings of the Tenth ACM International Conference on Future Energy Systems. 2019, pp. 315–325. [28] The High-Luminosity upgrade of the LHC.url: https://cds.cern.ch/record/2263093/ files/10.1088_1742-6596_706_2_022002.pdf (visited on 10/11/2025). [29] Francesco Vaselli et al. “End-to-end simulation of particle physics events with Flow Matching and generator Oversampling (2024)”. In: arXiv preprint arXiv:2402.13684 (). [30] Francesco Vaselli et al. “FlashSim: Accelerating HEP simulation with an end-to-end Machine Learning framework”. In: EPJ Web of Conferences (2024). url: https : / / api . semanticscholar.org/CorpusID:269703146. [31] Haifeng Wang and Qingkui Chen. “Power Estimating Model and Analysis of General Programming on GPU.” In: J. Softw. 7.5 (2012), pp. 1164–1170. [32] Jie You, Jae-Won Chung, and Mosharaf Chowdhury. “Zeus: Understanding and optimizing { GPU } energy consumption of { DNN } training”. In: 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). 2023, pp. 119–139. OPTIMIZING THE OPERATION OF GPUS TO REDUCE POWER CONSUMPTION 21