Post-Doctoral position TSAI

Descrizione

Post-Doctoral position TSAI Sintesi in italiano (traduzione automatica): Il Laboratorio di Informatica, Sistemi, Elaborazione delle Informazioni e Conoscenza (LISTIC) presso l'Università della Savoia Mont Blanc offre una posizione di Ricercatore post-dottorato. Il ruolo si concentra sulla ricerca di soluzioni per ridurre l'impatto ambientale e il consumo energetico dell'addestramento distribuito in ambienti di calcolo eterogenei. Le principali mansioni includono la valutazione e l'implementazione di algoritmi innovativi, nonché lo sviluppo di uno strumento open-source chiamato DAHL. È richiesta una laurea in Informatica o un campo correlato, con esperienza in machine learning e sistemi distribuiti. La posizione è basata in Francia e mira a contribuire a una maggiore efficienza energetica nel settore dell'Intelligenza Artificiale. The Laboratory of Computer Science, Systems, Information Processing, and Knowledge (LISTIC) is a Research Unit (UR) at the University of Savoie Mont Blanc (USMB), recognized by the Ministry since 2003 as a research team (EA3703). Its scientific research focuses on machine learning and information fusion, as well as networks and systems. Its work involves the modeling, specification, and development of theories, algorithms, and systems for extracting and managing knowledge, particularly in the fields of Earth observation and human behavior. Title : Energy-Aware Distributed AI on Heterogeneous Clusters In 2020, Information and Communication Technologies (ICT) already accounted for 4.4% of global electricity consumption (Malmodin et al, 2024), a share expected to increase especially with the emergence of Artificial Intelligence (AI) tools, such as large language models. Indeed, ADEME (the French Agency for Ecological Transition) anticipates a tripling of greenhouse gas emissions and an 80% increase in electricity consumption for the sector by 2050. Training these models requires huge computational resources, and therefore energy resources, resulting in substantial greenhouse gas emissions (Patterson et al., 2021) and significant water consumption for server cooling (Li et al., 2021) This increase in energy demand is driven by the rapid expansion of computational requirements: since 2012, the computing power needed to train state-of-the-art AI models has doubled approximately every three to four months (Amodei et al., 2018). This trend has been made possible by the using of parallel computing units, such as GPUs, which significantly accelerate machine-learning processes. GPU-equipped clusters coupled with traditional computing unit allow computations to be parallelized across multiple machines, but they introduce distributed systems issues, including task scheduling, fault tolerance, and workload allocation. Despite their importance, the intersection between energy efficiency, environmental footprint, AI, and distributed systems remains largely underexplored (Orgerie et al., 2014), as most research continues to prioritize performance and scalability over environmental considerations (Guo et al, 2024; Joost Verbraeken et al., 2021). Recent findings from the ADEME–Arcep report (ADEME, 2025) highlight the huge carbon footprint of ICT sector, with 50% of emissions attributable to terminals, and 46% to data centers. Most device-related emissions come from manufacturing, distribution, and end-of-life processes rather than from operational use. These insights challenge the current paradigm of homogeneous clusters, which rely on identical machines connected via uniform high-speed links. For example, GPT-3 was trained using 10,000 V100 GPUs (Patterson, 2021), consuming energy comparable to that of a European nuclear reactor operating for one hour. We believe that current distributed AI frameworks are not adapted to heterogeneous computing environments. Algorithms designed for homogeneous platforms treat slower machines as bottlenecks, generating idle time, inefficiencies, and unnecessary energy consumption. This design leads to premature replacement of slower hardware, necessitates investment in large sets of identical machines, and limits the reuse of existing resources. Consequently, the conventional approach to AI training worsens the environmental impact of large-scale computational infrastructures. The aim of this postdoc is to evaluate, compare, and implement new solutions to reduce the environmental impact and the energy consumption of distributed training in a heterogeneous computing environment. An open-source tool, DAHL, is currently under development. DAHL is a machine learning framework designed to minimize the energy footprint of AI training on a single machine and to operate efficiently in heterogeneous computing environments. It targets to train a Convolutional Neural Network (CNN) on several datasets. By leveraging the dive Annuncio in inglese. Fonte: Euraxess (Commissione europea).