Cargando...
Fecha
2025-09
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editorial
Citas
0 citas en 
0 citas en 
Resumen
Este Trabajo Fin de Máster se sitúa en el contexto del creciente consumo energético asociado al entrenamiento y uso de modelos de inteligencia artificial, especialmente los basados en redes neuronales profundas (Deep Learning). Como respuesta a esta problemática, se ha de- sarrollado una herramienta de monitorización denominada Energy Report System (ERS), orientada al análisis del consumo energético y las emisiones de carbono asociadas a procesos de entrenamiento e inferencia de modelos de aprendizaje profundo. El objetivo principal es proporcionar una solución que permita evaluar el impacto computacional y ambiental de estos modelos, integrando métricas de uso de recursos y estimaciones de huella de carbono en función de la ubicación geográfica.
ERS ha sido implementado para sistemas Linux y es capaz de registrar métricas relacionadas con la CPU, GPU, memoria y temperatura, a nivel de sistema, proceso e incluso bloques de código específicos. Se ha desarrollado además una métrica compuesta que combina precisión del modelo, consumo energético, emisiones de carbono y tiempo de ejecución, con el fin de facilitar comparaciones entre distintos modelos.
Para validar la herramienta, se han entrenado y evaluado cuatro arquitecturas de redes neuronales (ResNet-18, VGG-19, GoogLeNet v1 y MobileNetV2) utilizando el conjunto de datos CIFAR-10. Durante estos experimentos, se ha recogido información detallada sobre el comportamiento energético de cada modelo en los procesos de entrenamiento e inferencia en dos sistemas de pruebas.
El sistema desarrollado constituye una herramienta de apoyo al análisis de eficiencia en entornos de machine learning y puede ser útil en futuras investigaciones orientadas a mejorar la sostenibilidad de los modelos de IA.
This Master’s Thesis addresses the increasing energy consumption associated with the training and deployment of artificial intelligence models, particularly those based on Deep Neural Networks. In response to this issue, a monitoring tool called Energy Report System (ERS) has been developed. It is designed to analyze the energy consumption and carbon emissions generated during the training and inference process of Deep Learning models. The main objective is to provide a solution capable of assessing both the computational and environmental impact of these models by integrating resource usage metrics and carbon footprint estimations based on geographic location. ERS has been implemented for Linux systems and is capable of recording metrics related to CPU, GPU, memory, and temperature, at system level, process level, and even within specific code blocks. A composite metric has also been proposed, combining model accuracy, energy consumption, carbon emissions, and execution time, in order to facilitate comparisons between different models. To validate the tool, four neural network architectures (ResNet-18, VGG-19, GoogLeNetv1 and MobileNetV2) were trained and evaluated using the CIFAR-10 dataset. During these experiments, detailed information was collected on the energy performance of each model during both training and inference stages in two test systems. The developed system serves as a supporting tool for efficiency analysis in machine learning environments and may prove useful in future research aimed at improving the sustainability of AI models.
This Master’s Thesis addresses the increasing energy consumption associated with the training and deployment of artificial intelligence models, particularly those based on Deep Neural Networks. In response to this issue, a monitoring tool called Energy Report System (ERS) has been developed. It is designed to analyze the energy consumption and carbon emissions generated during the training and inference process of Deep Learning models. The main objective is to provide a solution capable of assessing both the computational and environmental impact of these models by integrating resource usage metrics and carbon footprint estimations based on geographic location. ERS has been implemented for Linux systems and is capable of recording metrics related to CPU, GPU, memory, and temperature, at system level, process level, and even within specific code blocks. A composite metric has also been proposed, combining model accuracy, energy consumption, carbon emissions, and execution time, in order to facilitate comparisons between different models. To validate the tool, four neural network architectures (ResNet-18, VGG-19, GoogLeNetv1 and MobileNetV2) were trained and evaluated using the CIFAR-10 dataset. During these experiments, detailed information was collected on the energy performance of each model during both training and inference stages in two test systems. The developed system serves as a supporting tool for efficiency analysis in machine learning environments and may prove useful in future research aimed at improving the sustainability of AI models.
Descripción
Categorías UNESCO
Palabras clave
deep learning, eficiencia energética, monitorización, Green AI, sostenibilidad, deep learning, energy efficiency, monitoring, Green AI, sustainability
Citación
Quesada Pablos, Alberto. Trabajo Fin de Máster: "Análisis del consumo energético de redes neuronales profundas". Universidad Nacional de Educación a Distancia (UNED), 2025
Centro
E.T.S. de Ingeniería Informática