I'm wondering if it is possible that evaluation by value would not terminate but normal evaluation terminates.
Because in normal evaluation we will evaluate until we reach a normal form meaning that we cannot evaluate further, but in by-value, we evaluate everything except the last lambda and we evaluate all the arguments even the ones that are not used.
So I don't think it would be possible because normal evaluation is more eager than by-value.
Yes, it is possible that evaluation by value (which fires a $\beta$-redex only if the argument is a value, i.e. an abstraction or a variable) does not terminate but normal evaluation (which fires the leftmost-outermost $\beta$-redex) terminates.
Consider for instance the term $(\lambda y. I) (\Delta \Delta)$, where $I = \lambda x.x$ and $\Delta = \lambda x. xx$. We have: