OK seems to be working correctly now, dropped all the GPU's back to stock clocks/volts and modified the app_info.xml and changed all the CUDA section in it to read:
<avg_ncpus>0.200000</avg_ncpus>
<max_ncpus>0.200000</max_ncpus>
and
<coproc>
<type>CUDA</type>
<count>0.5</count>
</coproc>
And also found a 64bit version of WinAFC (which was why I couldn't assign the 64 bit CPU client) that lets me keep one CPU core free for GPU use and the other 7 for CPU tasks.
Hasn't rebooted yet and GPU's are showing 60-90% useage across the board.
So 7 cores running 64bit optimised app, and 8 GPU Fermi clients on the one box, pulling 930W from the wall
Probably going to run out of work fast tho. And/or have to run reschedular on a regular basis.
Will give it a few days to settle RAC and see how we go, its not going to be pumping Seti on a permament basis but I and I know a few others were interested to see what she could do.
If its stable like this for 24 hours I'll look at tweaking again.