Journals - MOST Wiedzy

Abstract

A computer simulation of an atomic fluid on a GPU was implemented using the CUDA architecture. It was shown that the programming model for efficient numerical computing applications was changing with the development of the CUDA architecture. The introduction of the L2 cache decreased the latency between the global GPU memory and the registers. The performed MD simulation using the global memory and registers showed that the average acceleration relative to the CPU reached 80 times for single-precision calculations. Usually, the shared block memory gives much better results for this kind of calculation. We have found that using the shared memory gives acceleration over 116 times in comparison to the CPU. It is about 49% faster than using the global memory and registers. It is shown here that the performance of generally available graphics cards for double-precision calculations is significantly lower than for single-precision calculations. The recorded double-precision acceleration relative to the CPU in our experiment averaged 6 and 7 times for the global and shared memory, respectively. We performed these calculations on two different CUDA enable device systems.`