To answer the question you asked: Yes, replacing nested capturing subgroups with non-capturing parens will improve the performance of your regexes.

In a rather crude test, almost exactly 1 order of magnitude improvement for each pair of capturing parens that you replace. That is m[((?:xxx)(?:yyy))]g runs almost exactly twice as fast as m[((xxx)(yyy))]g; and m[((?:xxx)(?:yyy)(?:zzz))]g runs almost exactly 3 times faster than m[((xxx)(yyy)(zzz))]g;; and so on.

In that test, I was assigning all the captures to an array, which obviously results in differing numbers of values in the arrays.

If you are only refering to the captures you are interested in by their aliases ($1, and ignoring $2 and $3 etc.), then the impact of using the capturing parens that you don't need is less (around 15%), but it is still significant.

The upshot is: only capture what you need, and use non-capturing parens anywhere else.