Architectures and limits of GPU-CPU heterogeneous systems