I have a Reference XFX card (925/1375)

If I let cgminer compile the kernel at default clocks, and then apply overclock, it reaches ~700kh/s

If I let cgminer compile the kernel overclocked, it won't go past ~500kh/s, with same settings/clocks

so I guess those ghz editions have to be downclocked to default speeds at first, and then overclocked after cgminer compiled the kernel

I ran a diff on both compiled kernels, and found binary differences.