Maximizing Efficiency with GPU as a Service