TensorFlow officially only supports GPUs with "NVidia Compute Capability" of 3.5 or higher. In other words, basically it only supports GPUs of Titan or higher. This is explained in "Optional: Install CUDA (GPUs on Linux)" at https://www.tensorflow.org/get_started/os_setup.html#installation-for-linux as follows.
TensorFlow GPU support requires having a GPU card with NVidia Compute Capability >= 3.5. Supported cards include but are not limited to:
- NVidia Titan
- NVidia Titan X
- NVidia K20
- NVidia K40
Therefore, for example, if you install using the TensorFlow binary for GPU currently distributed (12.7.2015), an error will occur if you use an unsupported GPU such as GTX 960.
However, you can install TensorFlow for GPUs with "NVidia Compute Capability" of 3.0 by using the "unofficial settings" described in "Enabling Cuda 3.0" on the same page.
Note that there are many incompatible libraries just because they say "unofficial", and most of them don't seem to be test-supported. This can be confirmed by the following precautions that appear during installation (step 2 of "Procedure").
WARNING: You are configuring unofficial settings in TensorFlow. Because some external libraries are not backward compatible, these settings are largely untested and unsupported.
Follow the steps below to actually install.
Follow the steps below to build from source according to the official website (https://www.tensorflow.org/versions/master/get_started/os_setup.html):
./configure according to" Enabling Cuda 3.0 "
--The same content as above is detailed at https://github.com/tensorflow/tensorflow/issues/25#issuecomment-156234658.The author confirmed the operation on Geforce GTX 960, Ubuntu 15.04 with this setting. (I checked the operation by executing https://github.com/woodrush/neural-art-tf.)
Recommended Posts