How to choose the best Nvidia driver for CUDA application (PTX run-time compilation)?

Hello, everyone!

I would like to know, how to choose which Nvidia driver would be the best to achieve maximum performance in application using PTX run-time compilation?

I can’t share kernel code, but I can say that kernel was compiled from *.cu file to *.ptx file with CUDA 5.5 Toolkit. Only integer instructions are used inside kernel (hash calculation).
Environment: Windows 7 x64, 16 GB RAM, Nvidia GTX 680. Test application is x64.

Experiment results (the higher performance is better):
GeForce 327.23 Driver - WHQL has performance 19600
GeForce R331 Game Ready Driver (331.82 - WHQL) has performance 10800.

It’s almost 50% of performance decrease!
I thought it’s because 331.82 driver is “Game Ready”, but description of 327.23 driver also tell us it is “Game Ready” driver.

So, I’m very confused. Do I need to install only NOT Game Ready drivers? Or it is a bad driver optimization which will be fixed in near future and I always should install the latest driver?