Graphic Processing Units (GPUs) are currently widely used in High Performance Computing (HPC) applications to speed-up the execution of massively-parallel codes. GPUs are well-suited for such HPC environments because applications share a common characteristic with the gaming codes GPUs were designed for: only one application is using the GPU at the same time. Although, minimal support for multi-programmed systems exist, modern GPUs do not allow resource sharing among different processes. This lack of support restricts the usage of GPUs in desktop and mobile environment to a small amount of applications (e.g., games and multimedia players). In this paper we study the multi-programming support available in current GPUs, and show how such support is not sufficient. We propose a set of hardware extensions to the current GPU architectures to efficiently support multi-programmed GPU workloads, allowing concurrent execution of codes from different user processes. We implement several hardware schedulers on top of these extensions and analyze the behaviour of different work scheduling algorithms using system wide and per process metrics.