That also very much depends on model size. For 70B+ models, while the tok/s are still fast enough for realtime chat, it's not going to be generating faster than you can read it, even on Ultra with its insane memory bandwidth.