Studying execution time and memory transfer time of image processing using GPU cards

No Thumbnail Available
Issue Date
Asaduzzaman, Abu
Jojigiri, Srinivas
Sabu, Thushar
Tailam, Sanath

Asaduzzaman, A., Jojigiri, S., Sabu, T., & Tailam, S. (2021). Studying execution time and memory transfer time of image processing using GPU cards. Paper presented at the 2021 IEEE 11th Annual Computing and Communication Workshop and Conference, CCWC 2021, 689-695. doi:10.1109/CCWC51732.2021.9376170


With advancements in technology, the size and quantity of captured images to be analyzed are rapidly increasing. As the number of pixels in an image increases (more than 20 megapixels), additional processing power is required to process the images in a time-efficient manner. Modern computer systems with multicore Central Processing Unit (CPU) and General Purpose Graphics Processing Unit (GPU, short for GPGPU) show promise in processing images. However, different GPU cards provide different kinds of performance on different imaging techniques. In this work, we study execution time and memory overhead of three different GPU cards on two image processing algorithms, namely Sobel filter along with Bilateral filter and Canny edge detector. GPU cards used are Nvidia Tesla/Fermi C2075, GeForce/Maxwell GTX-965, and Tesla/Kepler K20. Experimental results with more than 120 megapixels show that K20 provides the best execution time speedup (more than 300x) for the Sobel filter. Whereas, GTX-965 provides the best speedup (almost 120x) for the Canny edge detector. It is also observed that the memory transfer time due to GTX-965 is minimum (about 0.21 seconds for the Sobel filter and 0.23 seconds for the Canny detector). For both algorithms, GTX-965 performs better or about the same as C2075. This study is helpful to select the right GPU card for a given problem.

Table of Content
Click on the DOI link to access this conference paper at the publishers website (may not be free).