出版社：Vilnius University, University of Latvia, Latvia University of Agriculture, Institute of Mathematics and Informatics of University of Latvia
摘要：The design space of a computer vision engineer is very large when it comes to the selection of hardware for high performance and energy efficient computing. By comparing a few potential (Parallella with Epiphany co-processor, Radxa Rock2 with Mali T764 GPU and Airvision Core X1 with Nvidia Tegra X1) platforms we have narrowed down this design space. This paper analyses three recent heterogeneous platforms for a typical image processing application ¨C convolution based Sobel filter. By measuring platforms energy consumption while computing processing intensive task of image filtering we are trying to identify best fit embedded heterogeneous computing platform for energy and sized constrained environment. Platforms were selected by the ability to incorporate parallel computing on a co-processor or a GPU, but also should use less than 10 Watts of electrical power and should be no larger than a credit card so it would be suitable for such constrained environments as small UAVs. The results shown that GPU platforms are more efficient compared with single core CPU application and co-processor technology, though Nvidia Tegra X1 processor was best performing ( 142 x faster than a single core application and 29x faster than nearest GPU opponent) and the most energy efficient (used 84x less energy than a CP U and 12x than GPU opponent).
