Abstract
The number of applications that use GPU accelerated calculations always grows. The software conversion to such calculation type is although complex but gives enormous energy efficiency and performance. In the publication is presented a method of the data sets, in which a temporary storage is occurred in the shared memory of GPU. The loading of values to shared memory via massive data access makes it possible to exploit the full power of the High Bandwidth Memory. It is demonstrated by CUDA codes for Cellular Automaton application and corresponding indexing of the data in global and shared memory. The real and the theoretical Speed-Up are described and shown in this publication.