I can’t really explain these specific observations as it would require us to study individual requests.
And explanation probably has more to do with complexity of sentinel-1 processing steps than with how specifically Sentinel Hub implements them.
I suggest you first establish a 1-1 comparison, i.e. use SNAP to produce exactly the same configuration of the results as with Sentinel Hub, then compare those. If you find significant differences (there is always rounding happening here, so…), please provide specific examples here and we might be able to compare them from IT point of view.
Best,
Grega