Hardware accelerators are becoming a highly appealing approach to boost the raw performance as well as the price-performance and power-performance ratios of current clusters. In this paper we present a strategy to retarget PLAPACK, a library initially designed for clusters of nodes equipped with general-purpose processors and a single address space per node, to clusters equipped with graphics processors (GPUs). In our approach data are kept in the device memory and only retrieved to main memory when they have to be communicated to a different node. Here we benefit from the object-based orientation of PLAPACK which allows all communication between host and device to be embedded within a pair of routines, providing a clean abstraction that enables an efficient and direct port of all the contents of the library. Our experiments in a cluster consisting of 16 nodes with two NVIDIA Quadro FX5800 GPUs each show the performance of our approach.