K20:Lower performance with a higher Driver Verion ?

I installed my CUDA toolkit by CUDA-5.0(Driver version 5.0 && Runtime Version 5.0) on K20.
For new characteristics, I installed cuda-9.0 tookit nowadays. However, I got a LOWER performance on my project with driver version 9.1 and runtime version 9.1 than before!
Meanwhile, I tried to install driver version 9.1 with runtime version 5.0, but I got a lower performance than before unfortunately.

Why does it happen? Why does the higher version get a lower performance?

CUDA 5 didn’t officially support Kepler (sm_35) devices like the K20. This support arrived with CUDA 6. Of course properly built code on CUDA 5 should JIT and run on CUDA 6.

It’s certainly possible that two different CUDA versions can give two different levels of performance with a particular code. It would be impossible to answer why without a full test case. Even if you provided a full test case, I personally would not spend the time to investigate behavior of a CUDA toolkit version that is about 6+ years old. You’re welcome to file a bug at developer.nvidia.com if you wish.