Performance Analysis of Deep Neural Networks for Object Classification with Edge TPU
Ahmad Ammar Asyraaf Jainuddin, Yew Cheong Hou, Mohd Zafri Baharuddin, Salman Yussof
Abstract
Deep learning becomes a more popular, widespread, and common tool in almost any task that requires information extraction from a large dataset. Hence, the data transmission speed between the data-gathering devices and processing units can be crucial in hardware selection depending on the machine learning application. Generally, the processing unit is usually centralized, and the data transferring time will increase when the data-gathering devices were installed further away from the processing unit. The work aims to provide the performance analysis on Google's new machine learning hardware called Edge TPU that was created specifically for edge devices. Furthermore, the work also reviewed the different types of deep neural network models as current benchmarks in deep learning were tested with different hardware used in edge applications. The review also discussed the comparison of the performance of the edge device using the deep neural networks in Tensorflow. From the results obtained, the performance of the edge device with the Edge TPU is faster than the device without it.