Benchmarking Jetson Platform for 3D Point-Cloud and Hyper-Spectral Image Classification

Modern innovations of embedded system platforms (hardware accelerations) play a vital role in revolutionizing deep learning into practical scenarios, transforming human efforts into an automated intelligent system such as autonomous driving, robotics, IoT (Internet-of-Things) and many other useful applications. NVIDIA Jetson platform provides promising performance in terms of energy efficiency, favorable accuracy, and throughput for running deep learning algorithms. In this paper, we present benchmarking of Jetson platforms (Nano, TX1, and Xavier) by evaluating its performance based on computationally expensive deep learning algorithms. Previously, most of the benchmark results were based on 2-D images with conventional deep learning models for image processing. However, the implementation of many other complex data types at Jetson platform has remained a challenge. We also showed the practical impact of optimizing the algorithm vs improving the hardware accelerations by deploying a diverse range of dense and intensive deep learning architectures at all three aforementioned Jetson platforms, to make a better comparison of performance. In this regard, we have used two entirely different data-types, namely (i) ModelNet-40(Princelon-3D point-cloud) data-set along with PointNet deep learning architecture for classification of 3D point-cloud, and (ii) hyperspectral images (HSI) datasets (KSC and Pavia) alongside stacked auto-encoders(SAE) to classify HSI correspondingly. This will broaden the scope of edge-devices to handle 3-D and HSI data whilst real-time classification will be processed at edge-server under the umbrella of edge-computing. The selection of (i) was made to exploit GPU heavily as the code uses TensorFlowgpu whereas (ii) was chosen to challenge the CPU cores of each platform as the code is based on Theano and may suffer from under-utilizing the GPU cores. We have presented the detailed evaluation exclusively in term of performance indices as inference time, the maximum number of concurrent processes, resource utilization per process and efficiency.