Browsing by Author "Kuzu, Ersin"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Hesaplama ağırlıklı algoritmaların programlanmasında grafik işlemci (GPU) kullanımının incelenmesi
    (Bahçeşehir Üniversitesi Fen Bilimleri Enstitüsü, 2014-08) Kuzu, Ersin; Tunga, Alper
    It is obvious that the expectations will continue to increase along with the technology. Therefore, a single processing unit will never adeqaute in application areas in which high processing features are needed. The problems are shared with multiple processing units in both past and future to increase the total processing capacity. Because central processing units are cheap, it is preferred to use them to construct processor farms. In the recent past, the rapid growth of the game industry was the driving force for the development of the graphics processors. On the other hand, graphics processors are completely specialized hardware on computational calculations. Programming these graphics processors for general purposes (GPGPU) lets this hardware to be used in any kind of calculations. The standard programming language (OpenCL) independent of vendors lets these processors to be used as multi-core structure, calculation-oriented architecture and instruction set. Low cost is another feature to choose them. In this study, the performance analysis of some compute-intensive algorithms was carried out on CPU and GPU with OpenCL threads. Experimental results were obtained using computers equipped with 2-core CPU (Intel Core 2 Duo T9550), 448x2-core NVIDIA GPU (2xNvidia Tesla M2050), 1600x2-core ATI GPU (Ati Radeon HD 5970 X2) and which support OpenCL general programming platform technology. For this purpose, with the reference of 2-core CPU calculation results, the acceleration, performance, power consumption and cost graphs were obtained. According to the results; at least 13 times, up to 1001 times acceleration is obtained in approximately compute-intensive algorithms. In advanced password recovery algorithms, at least 152 times, up to 905 times acceleration, up to 1190 times performance increase, up to 119 times energy saving, and up to 618 times lower cost were obtained.