I think the slow responses are the real generation rate when you time share the GPU. If you are willing to let GPUs occasionally sit idle though, you can probably get a faster generation rate by dedicating more of the GPU time to that one client.