2025 · 2025 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON)
Energy Consumption Analysis of Large Language Models Across CPU and GPU Using Diverse Metric Types
Evidence basis: full-text-reviewed · Review status: catalog-reviewed; paper-author approval pending
energy-efficiency llm-efficiency performance-modeling machine-learning performance-analysis
LLM energy consumption CPU energy GPU energy green AI static metrics dynamic metrics NLP workloads energy prediction CASCON 2025
Core contribution: The paper models LLM inference energy separately on CPU and GPU using hardware counters, device metrics, and task/model features, then compares classical and neural regressors across language tasks.
Catalog abstract summary
The paper presents a multi-level analysis of energy consumption for large language model workloads across CPU and GPU settings, combining diverse static and dynamic metrics to build and validate energy-prediction models across NLP tasks.
Source: Public abstract/metadata mirror and IEEE bibliographic record, paraphrased; full text not obtained.
Problem and motivation
Energy demand depends on the model, task, context, hardware, and execution device. Aggregate power alone does not explain these differences or provide a practical predictor for choosing a lower-energy inference configuration.
Method and contribution
The experiments use an Intel Core i7-12700K, an NVIDIA RTX 3060 Ti, 32 GB system memory, and 8 GB GPU memory. perf, psutil, and pynvml collect 11 CPU/system metrics and GPU utilization, memory, temperature, and fan signals at the stated one-second sampling interval; clocks are locked. More than 50 Hugging Face models are tested across sentence similarity, summarization, question answering, and generation groups, with 20 random configurations per model and 10 repetitions. Linear/Ridge/LASSO/SGD, MLP, and random-forest models are evaluated with five- and ten-fold cross-validation and MAPE.
Findings and evidence
The reported examples show higher total GPU energy than CPU energy over a 10-second window in some comparisons, but higher CPU energy per iteration; MiniLM is given as approximately 0.4 J on GPU versus 1.2 ± 0.1 J on CPU. GPT2-XL is among the high-demand cases. CPU instruction/cycle counters and GPU utilization/temperature are useful predictors, while GPU memory becomes more important for generation and long context. The reported cross-validation MAPE reaches 2.58% for a five-fold CPU Ridge model and 6.28% for a five-fold GPU linear model; ten-fold results include 2.16% CPU SGD and 4.59% GPU MLP.
Limitations and future directions
Limitations: The study is inference-only on one CPU/GPU platform with fixed clocks, one-second sampling, and no systematic DVFS, accelerator, batching, concurrency, or data-transfer analysis. NVML metrics are relatively high-level, and the transfer of the learned models to other hardware, runtimes, and serving policies is not established.
Future work: Evaluate heterogeneous accelerators and serving stacks, include batching/concurrency and host-device transfer, model DVFS and thermal effects, measure training and end-to-end service energy, and report leakage-safe cross-hardware generalization with uncertainty intervals.
Sources and identifiers
- Published version published
- IEEE record and public abstract metadata public_abstract
When to cite this paper
Cite this paper when modeling or comparing LLM inference energy on CPU and GPU using counters and device telemetry.
- perf/psutil/pynvml features and task/model-aware regressors.
- The multi-task, multi-model evaluation with cross-validation and MAPE.
- The distinction between total-window energy and per-iteration energy.
- The fixed-hardware, inference-only boundary of the measurements.
Citation
@inproceedings{ezzatiJivan2025energyconsumption,
author = {Tong Zhang and Leila Tahmooresnejad and Naser Ezzati-Jivan},
title = {Energy Consumption Analysis of Large Language Models Across CPU and GPU Using Diverse Metric Types},
year = {2025},
booktitle = {2025 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON)},
pages = {301-310},
publisher = {IEEE},
doi = {10.1109/CASCON66301.2025.00056},
url = {https://doi.org/10.1109/CASCON66301.2025.00056}
}Other citation formats for Word and reference managers
Zhang, T., Tahmooresnejad, L., & Ezzati-Jivan, N. (2025). Energy Consumption Analysis of Large Language Models Across CPU and GPU Using Diverse Metric Types. In 2025 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON) (pp. 301-310). https://doi.org/10.1109/CASCON66301.2025.00056T. Zhang, L. Tahmooresnejad, and N. Ezzati-Jivan, "Energy Consumption Analysis of Large Language Models Across CPU and GPU Using Diverse Metric Types," in 2025 IEEE International Conference on Collaborative Advances in Software and COmputiNg (CASCON), pp. 301-310, 2025, doi: 10.1109/CASCON66301.2025.00056