The availability of low cost high performance hardware such as graphical processing units (GPUs) has been the driving force for considerable change in both traditional high performance computing (HPC), and more recently the data centre. Vendor independent methods of accessing such hardware include the recently standardised OpenCL framework. However, using GPUs as accelerators for certain computational algorithms requires careful provisioning of infrastructure in order to balance cost, power consumption, and device utilization. In this paper we describe vOpenCL, an architecture for provisioning OpenCL capable hardware. vOpenCL works by transparently offloading processing from traditional cluster nodes, to a remotely located farm of OpenCL capable processors such as GPUs. vOpenCL differs from existing provisioning solutions since it provides mechanisms to optimize the data transfers between components. This is achieved through the use of a module/plug in which fully exploits the Infiniband interconnect architecture. We provide full analysis of vOpenCL and all of its plugins, in a range of scenarios including from within virtualized environments. For certain algorithms, we demonstrate that the Infiniband plug in can provide significant performance benefits.