train speed slow after upgrade cuda


I am using pytorch to train and test on CNN model. Previously, I am using cuda 9.0 and pytorch installed through conda (also for cuda 9, linux, py3.7) to train the model, everything works great.

Recently, I upgrade my cuda to 10.0, during the time, upgrade driver version to 418.67, also upgrade the pytorch to cuda 10 version. However, the train and test went very slow. For one image, the predict time changed from 0.05s to 0.34s.

Did anyone met the same situation and know the root cause or have some hint?

My GPU is: Quadra P600.


I am using 1080Ti and recently updated cuda. The speed gets much slower. I don’t know what to do. Did you figure out something?