The CSV dumps generated by Netbox have the columns in random order that changes at every dump, making it very hard to compare the files without actually loading the CSVs.
I didn't check the rows order but that should be consistent too.
Another difference is generated in the dcim.device_types.csv where the front_image and rear_image fields, when set, have a temp_url_sig and a temp_url_expires GET parameters in the URL that changes frequently.

Two consecutive dumps should generate identical files.
Some of the above might depend on Netbox upstream and not the way we do the dumps, in that case please check with upstream what's their position and eventually open related issues in their bug tracker please.