A metric quantifying how many of the available parallel streams a model layer actually uses in practice.