SETI for CUDA on Linux vs Windows

Well, the question posed last week regarding the big difference in performance between my CUDA card under 64-bit Ubuntu and cards under Windows has been answered.
I finally spotted an instance where two 8600GT cards have received the same work unit. The machine (AMD 64 X2 6000+) running 64-bit Windows finished the work unit in 156.45 seconds while my machine (AMD 64 X2 4200+) running 64-bit Ubuntu finished in 2968.83 seconds. That is a difference of 94.37%!
Looking at the CPU bench-marks, the Ubuntu machine is actually quicker so that shouldn't influence the result. The only notable issue is that the Windows machine has double the memory, both RAM AND GPU memory.
It will be interesting to see if this major difference still exists when SETI release their own Linux client...


After some time searching forums for ideas, the following two concepts were suggested by a user on the lunatics.kwsn.net forums:

1. 256M of GPU memory is right on the border and this may cause the work unit to fall back to CPU computation. I'm not convinced about this as I have watched the GPU temperature and there is no evidence that it stops crunching.
2. The linux CUDA application does use 100% of a core and I was just thinking that the "0.04 CPUs" line was a complete joke. The time reported could well reflect the "wall time" for the work unit, Windows on the other hand does use a small percentage of a CPU core (I believe..?), so the time reported to complete a work unit isn't necessarily anywhere close to "wall time". If this is the case, a Windows machine essentially has another core compared to the Linux machine and will have a higher credit average ...Grrrr!