Performance Analysis of Convolutional Neural Network Models

• Published in: 2019 9th International Conference on Advances in Computing and Communication (ICACC) pp. 22 - 26
• Main Authors: Kala, S, Paul, Debdeep, Jose, Babita R, Mathew, Jimson, Nalesh, S
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2019
• Subjects: Convolutional neural networks; execution time; FPGA; graphics processing unit; training
• DOI: 10.1109/ICACC48162.2019.8986201

Deep Neural Networks (DNNs) have become a promising solution for numerous machine learning tasks, which include object detection, classification, weather forecasting, video surveillance and safety management. Convolutional Neural Network (CNN) is a class of DNN, which can achieve state-of-art accuracy and high performance for most of the computer vision tasks. Training of CNNs is computationally intensive and may take multiple days for completion. In this work we have performed the feed forward path implementation and training of AlexNet, VGG-16 and ResNet models on Nvidia GeForce GTX 1080 Ti GPU with 3584 CUDA Cores, 1582 MHz, 11GB GDDR5X, using Caffe framework. Inference and training of these models are also performed on Nvidia Jetson TX2 hardware platform. We also present a detailed analysis of graphics processing unit (GPU) and general purpose processor (GPP) implementations of AlexNet, VGG-16 and ResNet-50 network models. A hardware architecture for performing convolution operations in feed forward path of CNN has also been implemented in Virtex-7 XC7VX690T FPGA, with throughput of 184 GFLOPS for AlexNet, 172 GFLOPS for VGG-16 and 197 GFLOPS for ResNet-50 model.