Very low performance on video-to-video(not live mode) function on RTX 5070 Ti (But with strange ultra performance route?)

[Derhub66]

RTX 5070 Ti Intel i7 14700KF SSD + 128GB RAM

Performance is terribly slow with only average of 1.5s/frame processed with CUDA execution provider . cuDNN 8.6.0.163 CUDA 11.8 Python 3.10

BUT! Here's a weird part, I tested processing the video by enabling "Map face" first, then assign the selected faces and process to final video I was able to do it blazingly fast, about 50 to 70 times faster than processing the video normally through pressing "Start" with no change of quality. So that might be something worth looking into.