The difficulty in synthesizing sound is getting the models right. You can't simulate each atom so you need a simplifying model that allows you to reduce the work. And that model has to be accurate in the areas where it matters.
While moving stuff to a GPU gives more computing power (but in a more constrained fashion than a CPU) and certainly helps, the models aren't there yet.
The people researching physical modelling continue to make progress, but I think that if you put state of the art in a game, you'd perhaps have more lively sounds, but they'd probably sound worse than sampled sounds.