Abstract
In our current work we investigate the possibility of using modern AMD APU architecture in scientific and technical computing. The architecture combines both a CPU and a GPU in a single Accelerated Processing Unit, which theoretically allows for shortening the time of exchanging the data between the two hardware units. This capability solves the problem of performance bottleneck related to the exchange of data between the CPU and GPU memory. Due to the structure of this architecture, it can be considered as a natural evolution of the concept presented in the IBM PowerXCell processors that have been tested during our past research (Krużel & Banaś, 2013). As reference systems we use both a system based on similar AMD architecture and a specialized Nvidia Tesla Accelerator card. Moreover, due to comparable characteristics of the CPU and GPU parts of APU we have run our computations on both hardware units separately to see the difference in performance. For testing we used our previously developed finite element numerical integration algorithm implemented in OpenCL programming framework. This algorithm has been tested with various organizations of memory and computing techniques to fully check the hardware capabilities of the APU architecture, both in terms of data exchange and calculations acceleration. Our research brings an answer to the question whether this architecture is the right future for scientific computing and whether in the next few years will be able to play a significant role in many areas of computational science.