I dont believe this happens on Win 8.1. Take a look at Civilization V on Win 8.1Pro 64bit and FX8350.
It's not that it definitely is or isn't happening, but without clear long-term pegging of a few cores, it's hard to determine with just the performance graphs (if the same data were stacked in the same graph, it would be a lot better).
If there is reason to keep more cores active, not doing so, and parking, will lead to lower performance than bouncing around. If there aren't enough time slices to keep all those cores very busy, then not using them will lead to higher performance. Making that decision correctly is not simple in the first place, doubly so with MT being common, and doubly so again with Turbo features, and then once again for the same scheduler trying to be good at making CPUs conserve battery power. The only scheduler I am aware of that comes close to doing it ideally (and which has commonly been used for Android) has a problem scaling much past ~10 logical cores (CK's BFS patches for Linux), and isn't stuck with Windows' compatibility concerns.
Win 8 is better than Win 7 at it, which was better than Vista, which was better than XP, which was better than 2K, which was better than NT4...but none are anywhere close to perfect, and won't be any time soon. Each iteration can get better, due to having more cores and cache on newer sockets to actually test against, where provisions for more were speculation-based in the prior OS version.