The AMD APU (Accelerated Processing Unit) architecture, which combines CPU and GPU cores on the same die at a low power budget, promises a significant advent in GPU computing, in particular to applications which performance is bottlenecked by the low PCI communication rate. However APUs are slowly heading towards a perfectly unified CPU-GPU memory space. Besides, APU integrated GPUs are also less powerful than discrete GPUs. In this paper we therefore investigate the interest of APUs for scientific computing by evaluating and comparing the performance and the power efficiency of an AMD APU (code-named Kaveri) against a discrete GPU (code-named Tahiti) and an hexa-core AMD CPU (Phenom CPU). For this purpose, we rely on two common applicative benchmarks in single precision, namely a matrix-matrix multiplication and a 3D finite difference stencil, that are optimized and tuned in OpenCL. Our results confirm that APU integrated GPUs outperform CPUs and may outperform discrete GPUs for medium-sized problems as well as for applications with high communication requirements. Moreover, our study shows that APUs can be up to 20% more power efficient than discrete GPUs.