Some conclusions about the installation of cuda and cudnn and tensorflow

CUDA

1. Before installing CUDA should first know the native graphics driver version number, the latter has a corresponding download CUDA graphics driver version number should correspond well.

2. While the low version can be installed on cuda higher than its driver, but this is not recommended because of the installation of other software version is rather abruptly gpu will be a problem.

  The machine is driven such as 435, corresponding to the best version of the upper right corner cuda can be viewed by nvidia-smi. As shown below

 

 I installed this drive is 435, the upper right corner also showed the highest requirement of 435 cuda version is 10.1. So I can install the drive less than 435 but less than cuda 10.1 in. But this is not recommended to install, it will go wrong. Or as recommended to install.

 

CUDNN

Cuda installing the necessary equipment cudnn, cudnn also pay attention to its version number at the time of the download, it is also behind the corresponding version of cuda.

 

tensorflow installation

When gpu version installed tensorflow to note the following versions of the corresponding question. Of python, gcc, cudnn, cuda are all required

 

Guess you like

Origin www.cnblogs.com/answerThe/p/12643530.html