Maybe it's bug, maybe feature ;) On my i7 it works excactly the same way. Looks like boinc is asigning one pcu core/thread to gpu automatically. It's my guess.

That makes no sense, because WCG GPU already allocates a whole CPU.

In other GPU-projects sometimes the GPU allocates e.g. only 0.2 CPU and then in your case you would have running 8 CPU's + 1 GPU. The 0.2 CPU is 'stolen' from the other 8 CPU-cores.

Then it would make sense to free 1 core to speed up the GPU. On WCG it doesn't. Faster CPU's even have some spare time, because during the GPU-part of a HCC-task, the CPU is hardly used.

Come to think of it, it does not matter now does it, on an HT machine. Whether the core is permanently allocated or not, the spare cycles are being used [I happened to experiment yesterday on an 8 core machine with 8 concurrent, then suspending them one by one. The temps on the cores only started dropping below 4 concurrent tasks, PLUS, the tasks that did run progressed noticably faster... the TTC went down in a hurry] So, for the HT crunchers, if the GPU segment runs, the cycles go towards the CPU... having the cake and eat it. Does that pass like a sound theory?

As for the 87.5%... the percent always rounds down to the nearest core, so if you enter 99 percent on an octo, there still will be only 7 used... 87.5%.

Set preferences to use 100% of the CPU cores. The HCC GPU tasks will use 1 core by default leaving the other 7 to run CPU tasks. If you set your preferences to use 87.5% the GPU task will still take 1 by default leaving only 6 available to run CPU tasks.