An evaluation on using GPU coprocessing for software radios on a low-cost platform

The presented study explores the characteristics of signal processing for software radio on heterogeneous x86/GPU system architectures. Special attention is set on the question whether the use of the GPU as a signal coprocessor helps to reduce the actual load of the x86 host processor. We focus on low-cost platforms with a chipset-integrated GPU next to the application processor, since they are coming close to embedded needs. As a case-study, we evaluate a complete software defined radio being capable of standard-conform, real-timed on-air radio reception of Digital Audio Broadcasting. We present the obtained benchmark results for the compute kernels which were ported to the GPU subsystem, but also compare them to an implementation optimized solely to the x86 host processor. When outsourcing the DAB SDR kernels to the GPU coprocessor, it becomes apparent that GPU housekeeping overhead today introduces more load to the host CPU than the CPU would spent for actually computing SDR kernels by itself. This is verified by a detailed system-wide analysis, treating also all use case related aspects beside the actual signal processing kernels.