Inference device & Custom camera placements

[cyhasuka]

Hi there,
Thank you for open source your impressive work!
I have read your paper and understand that the model was trained for six days using 64x H20 GPUs. Additionally, inference for 16 frames × 8 views takes 4 seconds (about 4 FPS).

I would like to know:

1. What hardware was used for the aforementioned 16-frame inference? Single H20 GPU?
2. I note that Prediction quality on custom surround-view data (single frame, 6 cameras) #4 mentioned certain custom camera placements (resembling a bus configuration) can cause external parameter calculations to fail. Is this a widespread issue? If my cameras have non-overlapping fields of view or blind spots at certain angles, will this negatively impact external parameter estimation?