On the Use of Remote GPUs and Low-Power Processors for the Acceleration of Scientific Applications