If you start doing synth modeling (which I think is one of the reasons Rudess has problems with the Kronos, which models some of its sounds) on any CPU, you can drag polyphony down quite quickly. It really doesn't matter what CPU you have in there, a few stacked, modeled synth sounds, and you are going to reach your limit.

TBH, there are still considerable drawbacks to CPU generation of sounds, especially live, as voice count is dynamic depending on what VSTi's you use, and when the ceiling is approached, FAR more nasty things happen than mere note (or voice within a note) dropout. Your entire CPU can drop out..!

At least arrangers haven't generally gone this route so far. I KNOW that, no matter what sound I am using, I get that full 128. You can't say that with certainty on a CPU generated instrument.