On Benchmarking Multiple GPU Computing Resources for Faster Training of Deep Neural Networks

Abstract

In this technical note, we explore a method to benchmark the performance of a Lambda Quad Deep Learning Workstation on training seven leading deep neural network (DNN) models using multiple GTX-1080 Ti graphics processing units (GPUs). We compute the average number of images processed per second for each DNN and quantify the consistent improvements in speed/performance when additional GPUs are assigned to the task. Our results show that the multi-GPU performance statistics calculated at the US Army Combat Capabilities Development Command Army Research Laboratory (CCDC ARL) are in close agreement to those calculated at Lambda Labs. The benchmarking of an accessible distributed computing resource provides an important step toward mitigating the excessive computational costs often associated with modern machine-learning applications, to include the training of DNNs on large data sets.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Feb 06, 2020
Accession Number
AD1091936

Entities

People

  • Arnold D. Tunick

Organizations

  • United States Army Research Laboratory

Tags

Communities of Interest

  • Autonomy
  • Engineered Resilient Systems

DTIC Thesaurus Topics

  • Agreements
  • Artificial Intelligence Computing
  • Artificial Intelligence Software
  • Data Sets
  • Deep Learning
  • Distributed Computing
  • Graphics Processing Unit
  • Image Processing
  • Information Science
  • Information Systems
  • Learning
  • Machine Learning
  • Military Research
  • Networks
  • Neural Networks
  • Statistics
  • Training

Fields of Study

  • Computer science

Readers

  • Military Science and Technology Research and Modernization.
  • Neural Network Machine Learning.
  • Parallel and Distributed Computing.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks