Will cuDNN v6.0 work with TensorFlow now?

Question

Do we need to specifically use cuDNN v5.1 (as suggested) for TensorFlow, or will the latest version (v6.0) work? Is there backward compatibility in cuDNN versions?

Answer 1

No, cuDNN 6.0 is not supported in the latest version 1.2. But there is hope: official release notes report the following:

TensorFlow 1.2 may be the last time we build cuDNN 5.1. Start with TensorFlow 1.3, we will try to build all our finished binaries with cuDNN 6.0. Although we will try to maintain source compatibility with cuDNN 5.1, this will be the best effort.

, , 1.3 cuDNN 6.0. , 7.0 .

Answer 2

(21 2017 .) Tensorflow v1.3, , REQUIRES cuDNN v6.0 v5.1. libcudnn.so.6, libcudnn.so.5

P.s. , cuDNN 5.1. , . v1.2, :

https://www.tensorflow.org/versions/r0.12/get_started/os_setup#virtualenv_installation

Answer 3

.

tensorflow 1.2.1, cudnn 6.0 . pip

pip install tensorflow
pip install --upgrade tensorflow 

.