Year: 2025
Type: conference-paper
Abstract: The recent advancement in artificial intelligence (AI) boosts the demand for AI-related workloads, significantly increasing GPU deployment and associated power consumption in data centers.This trend accelerates hardware upgrades, introducing the need for GPU upgrades planning, reallocation, and retirement.Existing lifecycle planning methods typically rely on static provisioning or heuristicbased strategies, which make it challenging to address uncertainties in workload demand, hardware performance degradation, power limitations, etc.As a result, poor resource utilization and increased operating costs are common in data centers.We propose an AI-assisted stochastic optimization framework for GPU lifecycle planning in data centers.This framework utilizes large language models to generate predictive scenarios for future workload demands, hardware performance decay, etc., in different cases.These AI-generated insights are integrated into a stochastic optimization model, which helps stakeholders make decisions about the timing of upgrades and retirements of GPU and cooling systems in data centers.
Authors: Chengyi Nie, Anna Xing, Imran Latif, Zhenhua Liu
Institutions: Stony Brook University, Brookhaven National Laboratory
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funder: National Science Foundation
Awards: Collaborative Research: CNS Core: Medium: Dynamic Data-driven Systems - Theory and Applications, Collaborative Research: CNS Core: Small: Optimizing Large-Scale Heterogeneous ML Platforms, Collaborative Research: CNS Core: Large: Systems and Verifiable Metrics for Sustainable Data Centers, CAREER: An adaptive framework to accelerate real-time workloads in heterogeneous and reconfigurable environments
OpenAlex ID: W4411346140