Year: 2025
Type: article
Abstract: This article examines the significant challenges of GPU inefficiency in machine learning model training workflows, addressing how suboptimal resource utilization leads to computational waste and increased costs. It explores the various factors contributing to this inefficiency, including improper batch size, inadequate memory management, data loading bottlenecks, and hardware configuration mismatches. The article presents a comprehensive framework for identifying, measuring, and optimizing GPU performance through advanced techniques such as dynamic batching, mixed-precision training, and efficient data pipeline engineering. By implementing these strategies, organizations can achieve more sustainable and cost-effective model training practices while maintaining computational performance, ultimately supporting broader accessibility of AI research and reducing the environmental impact of large-scale machine learning operations.
Author: Ramesh Mohana Murugan
Institution: Anna University, Chennai
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4409045689