In this proposal, GPU-accelerated applications are enabled to seamlessly interact with any GPU of a cluster independently of its exact physical location. This provides the possibility of sharing accelerators among different nodes, as GPU-accelerated applications do not fully exploit accelerator capabilities all the time, thus reducing power requirements. Furthermore, decoupling GPUs from nodes, creating pools of accelerators, brings additional flexibility to cluster deployments and allows accessing a virtually unlimited amount of GPUs from a single node, enabling, for example, GPU-per-core execution s. Depending on the particular cluster needs, GPUs may be either distributed among computing nodes or consolidated into dedicated GPGPU servers, analogously to disk servers. In both cases, this proposal leads to energy, acquisition, maintenance, and space savings. Performance evaluations employing the rCUDA Framework, developed as a result of the research conducted during a 4-year predoctoral period, demonstrate the feasibility of this proposal within the HPC arena.