Benchmarking Analysis of CNN Architectures for Artificial Intelligence Platforms

The prompt innovations in Digital Technologies with the availability of credible data have led to the emergence of an era of Artificial Intelligence and Deep Learning. This demonstrates their effectiveness in solving complex problems, particularly in image classification and object recognition applications, with the assistance of Convolutional Neural Networks (CNNs). However, such algorithms need to be executed within a certain time frame specifically applications like High Performance Computing (HPC), Autonomous vehicles, Gaming etc. The requirement of time certainty brings hardware accelerators into the picture. These hardware accelerators not only accelerate time-critical tasks but have also been proven effective in enhancing the throughput of CNNs. In this research, we have tried to tune in performances of different CNN models i.e. Alexnet, SqueezeNet1.1, GoogleNet-v1, and VGG-16 of the Caffe framework which have been pre-trained and converged, using the Bench-Marking and Cross Check Tools of OpenVINO Toolkit. The tool define the performance of CNN Models based on latency, throughput, absolute and relative differences in each layer of these models. The CNN models have been simulated on platforms like Intel i5-8265 CPU, 1.60 Ghz and Integrated GPU UHD graphics 620 using OpenVINO Toolkit, which helped in running the simulations on Windows 10. Furthermore, this study is expected to direct the future development of an efficient accelerator on specialized hardware accelerators and also be useful for deep learning researchers.