Lay Summary

This study explores how to speed up artificial intelligence training by efficiently distributing workload across multi-GPU computing systems. Using high-performance NVIDIA H100 hardware, the researchers evaluated strategies for training image recognition models and fine-tuning Large Language Models (LLMs). By benchmarking different parallelization techniques, data distribution strategies, and memory management settings, they analyzed how data transfers and memory limits create processing bottlenecks. Their findings offer actionable guidance on allocating tasks effectively and optimizing GPU usage, helping developers train complex AI models faster and with greater efficiency. (86 words)

Full Abstract

This work is concerned with the evaluation of the performance of parallelization of learning and tuning processes for image classification and large language models. For machine learning model in image recognition, various parallelization methods are developed based on different hardware and software scenarios: simple data parallelism, distributed data parallelism, and distributed processing. A detailed description of presented strategies is given, highlighting the challenges and benefits of their application. Furthermore, the impact of different dataset types on the tuning process of large language models is investigated. Experiments show to what extent the task type affects the iteration time in a multi-GPU environment, offering valuable insights into the optimal data utilization strategies to improve model performance. Furthermore, this study leverages the built-in parallelization mechanisms of PyTorch that can facilitate these tasks. Performance profiling is incorporated into the study to thoroughly evaluate the impact of memory and communication operations during the training/tuning procedure. Test scenarios are developed and tested with numerous benchmarks on the NVIDIA H100 architecture showing efficiency through selected metrics.

Metadata

• Publication Date: March 19, 2025
• Authors: Marcin Lawenda, Krzesimir Samborski, Kyrylo Khloponin, Łukasz Szustak
• DOI: 10.48550/arXiv.2503.15252