DirectML reuses GPU memory instead of returning it. So prints "practical" GPU memory utilization too.