Performance Benchmarking of GPU and TPU on Google Colaboratory for Convolutional Neural Network

Abstract

Convolutional Neural Network (CNN) is a sub-class of artificial intelligence deep learning technique. In this sub-class, algorithms are inspired by the structure and function of the brain called artificial neural network. CNN outperforms in image processing tasks such as image recognition, image classification, face recognition, object detection, etc. CNN consists of very large neural networks and is capable to process an enormous amount of unstructured data. To process these types of data, huge computational power is required, which should be both data intensive as well as compute intensive. For performance benchmarking of CNN, an experiment on hand written digit recognition dataset— MNIST has been performed. The main objective of this paper is a comparative study of CNN performance’s on accelerated computational power, i.e., Graphics Processing Unit (GPU) and Tensor Processing Unit (TPU). To access the GPU and TPU resources, Google colaboratory cloud platform called google Colab which has been utilized. After running CNN model for MNIST, it is observed that TPU takes comparatively less epoch time for different batch sizes as compared to GPU. This research paper is also a good source for AI researchers whose research area intersect with High-performance computing (HPC).