GPUs (Graphics processing units) are commonly used in high-performance computing to improve the execution time of parallelizable programs. SeisSol, as such a program, can currently only use GPUs manufactured by NVIDIA. In this thesis, we extend the GPU-related libraries to work additionally on GPUs manufactured by AMD. We examine the performance of the extended libraries on the AMD Radeon Instinct MI50 using the Roofline model and compare it to results measured on the NVIDIA Tesla V100. Furthermore, we investigate using multiple AMD GPUs, which communicate via OpenMPI. The Radeon Instinct MI50 achieved a good performance but was slower than the NVIDIA GPU. We detected a problem in the communication between the GPUs, which is restricted to collective communication. This does not affect SeisSol since it uses mainly point-to-point communication.