What is the difference between tensorflow-gpu and tensorflow?
Question:
When I see some tutorials regarding TensorFlow with GPU, it seems that the tutorial is using tensorflow-gpu instead of tensorflow.
The only info I got is the pypi page where it doesn’t cover much information.
Where the official web says that the tensorflow already packed with GPU support.
So are there any differences between the two libraries?
My hypothesis is in the early version tensorflow doesn’t have native GPU support they create separate libraries, and the tensorflow-gpu is still updated for older users who already use tensorflow-gpu.
[Update]
Thanks to the comment and the answers. I just finish installing several prerequisites related to Nvidia and use the plain TensorFlow and now I need to reinstall the CUDA version since the latest CUDA not compatible with the latest tensorflow. It’s true that the setup can be a pain in the arse.
Answers:
The main difference is that you need the GPU enabled version of TensorFlow for your system. However, before you install TensorFlow into this environment, you need to setup your computer to be GPU enabled with CUDA and CuDNN.
Support for TensorFlow libraries for hardware type:
tensorflow
tensorflow-gpu
cpu-only
yes
no (~tf-like)
gpu with cuda+cudnn installed
yes
yes
gpu without cuda+cudnn installed
yes
no (~tf-like)
There are not many differences between the two libraries. The core syntaxes will be the same, if you have installed tensorflow-gpu in your python(or conda) environment, then the inference will simply run on the GPU. Along with tensorflow-gpu packages, CUDA toolkit for python will be automatically installed if you are using conda environment.
When I see some tutorials regarding TensorFlow with GPU, it seems that the tutorial is using tensorflow-gpu instead of tensorflow.
The only info I got is the pypi page where it doesn’t cover much information.
Where the official web says that the tensorflow already packed with GPU support.
So are there any differences between the two libraries?
My hypothesis is in the early version tensorflow doesn’t have native GPU support they create separate libraries, and the tensorflow-gpu is still updated for older users who already use tensorflow-gpu.
[Update]
Thanks to the comment and the answers. I just finish installing several prerequisites related to Nvidia and use the plain TensorFlow and now I need to reinstall the CUDA version since the latest CUDA not compatible with the latest tensorflow. It’s true that the setup can be a pain in the arse.
The main difference is that you need the GPU enabled version of TensorFlow for your system. However, before you install TensorFlow into this environment, you need to setup your computer to be GPU enabled with CUDA and CuDNN.
| Support for TensorFlow libraries for hardware type: | tensorflow | tensorflow-gpu |
|---|---|---|
| cpu-only | yes | no (~tf-like) |
| gpu with cuda+cudnn installed | yes | yes |
| gpu without cuda+cudnn installed | yes | no (~tf-like) |
There are not many differences between the two libraries. The core syntaxes will be the same, if you have installed tensorflow-gpu in your python(or conda) environment, then the inference will simply run on the GPU. Along with tensorflow-gpu packages, CUDA toolkit for python will be automatically installed if you are using conda environment.