How can I make tensorflow run on a GPU with capability 2.x?

Question

I've successfully installed tensorflow (GPU) on Linux Ubuntu 16.04 and made some small changes in order to make it work with the new Ubuntu LTS release.

However, I thought (who knows why) that my GPU met the minimum requirement of a compute capability greater than 3.5. That was not the case since my GeForce 820M has just 2.1. Is there a way of making tensorflow GPU version working with my GPU?

I am asking this question since apparently there was no way of making tensorflow GPU version working on Ubuntu 16.04 but by searching the internet I found out that was not the case and indeed I made it almost work were it not for this unsatisfied requirement. Now I am wondering if this issue with GPU compute capability could be fixed as well.

Solution 1

Recent GPU versions of tensorflow require compute capability 3.5 or higher (and use cuDNN to access the GPU.

cuDNN also requires a GPU of cc3.0 or higher:

cuDNN is supported on Windows, Linux and MacOS systems with Pascal, Kepler, Maxwell, Tegra K1 or Tegra X1 GPUs.

• Kepler = cc3.x
• Maxwell = cc5.x
• Pascal = cc6.x
• TK1 = cc3.2
• TX1 = cc5.3

Fermi GPUs (cc2.0, cc2.1) are not supported by cuDNN.

Older GPUs (e.g. compute capability 1.x) are also not supported by cuDNN.

Note that there has never been either a version of cuDNN or any version of TF that officially supported NVIDIA GPUs less than cc3.0. The initial version of cuDNN started out by requiring cc3.0 GPUs, and the initial version of TF started out by requiring cc3.0 GPUs.

Solution 2

Sep.2017 Update: No way to do that without problems and pains. I've tried hard all the ways and even apply below trick to force it run but finally I had to give up. If you are serious with Tensorflow just go ahead and buy 3.0 compute capability GPU.

This is a trick to force tensorflow run on 2.0 compute capability GPU (not officially):

1. Find the file in Lib/site-packages/tensorflow/python/_pywrap_tensorflow_internal.pyd (orLib/site-packages/tensorflow/python/_pywrap_tensorflow.pyd)

2. Open it with Notepad++ or something similar

3. Search for the first occurence of 3\.5.*5\.2 using regex

4. You see the 3.0 before 3.5*5.2, change it to 2.0

I changed as above and can do simple calculation with GPU, but get stuck with strange and unknown issues when try with practical projects(those projects run well with 3.0 compute capability GPU)

Comments

[mickkk]

Now I wonder why I was able to run mxnet on GPU using cuDNN though... In principle you could not even install tensorflow GPU on the last Ubuntu LTS..

[Robert Crovella]

cuDNN won't work on a cc2.1 GPU. Perhaps mxnet has a gpu-enabled path which does not require cuDNN. This would seem to be the case here. Note that GPU support is claimed for cc2.0 and greater, but that it uses "CUDNN to accelerate the GPU computation".

[Marcin Tarka]

I strongly advise not to do so. After aplying this trick on my laptop with GeForce 800M results where incorrect.

[Tin Luu]

Yes, it's sad to find out that. My GPU is also found to function incorrectly with complex model (strange bugs), while with the same model (same code), it can run smoothly with GPU 3.0

[mayank]

Thanks guys for reporting back the issues in your experiment above. It helps me to simple let it go and understand that I have to get a new GPU if i want to run TF. :) @TinLuu, please consider editing your answer to reflect issues so that others who might skip these comments do not go that way either!

[Tin Luu]

Thanks for your suggestion! I have updated the answer so that one can easily make decision

[JarsOfJam-Scheduler]

@RobertCrovella warning: the first two links are 404