I remember when I read about eGPUs for the first time. The symbiosis of having a light weight laptop at university or on the go, but still having a desktop like power horse when having some spare-time at home sounded like a dream. But everything faded into obscurity because I almost lost full interest into gaming the last years.

But this changed, since I'm spending a lot of time in deep learning since about two years. And it's well known that taking advantage of a GPU boosts training time by a huge margin. That's why I tried to get access to a high-performance graphics card in order to be able to train non-trivial networks and so some more serious research.

At first, I had a look at some offers in the cloud. I did not try out a GPU-enabled instance on AWS, because the use a billing based on a hourly rate. This means that you have to pay for a full hour, even when you just run a simple example for 1 minute. Next, I checkout out the 300$ free credit on Google compute engine. Unfortunately, the offer was limited to a 8-core CPU instance, while no GPU instance was available. Last but not least, I checked out FloydHub, which actually worked quite well. The free version comes with 10 hours of GPU access on a Nvidia Tesla K80. However, this is a dual-GPU and you only get access to one of them, so the performance is actually worse than it looks like on most benchmarks.

After checking out some of these cloud platforms, I was still curious about how an eGPU performs with TensorFlow. There was basically no benchmark available regarding eGPU setups for Machine Learning. Most or even all of them are focused on pure gaming. And that's why I thought: well, then do it yourself!

Of course, there were two big bullet points which made me hesitate some more days:

• high GPU prices due to crypto-currency mining
• no official support and only very few resources for most eGPU systems on Linux

To compensate the former, I ordered a used one on Ebay. Regarding the latter one, I simply wanted to give it a shot...

Ultimately, I ended up with the following components:

• MSI GeForce GTX 1070 Ti Armor 8GB
• Sonnet Breakaway Box 350

TL;DR

Yes, it's worth it!

Installation

While the setup of the eGPU on Windows is literally plug & play, there is much more to do on Ubuntu 17.10. In this post, I wanted to share how I achieved to run a simple TensorFlow r1.6 example on the eGPU. The test-system in my case was the ThinkPad X1 Carbon 6th Gen.

1. UEFI/BIOS security settings

On my system, I had to disable some security features in the UEFI/BIOS settings. I had to disable the following, otherwise my eGPU was not detected:

• Secure Boot
• Thunderbolt Security Level

2.  Install NVIDIA graphics driver

Next, the graphics drivers have to be installed. These available driver versions for linux can be found here. Interestingly, I installed the latest version 390 first, but after installing CUDA, the version 387.34 has been installed out of the sudden. I'm not sure whether CUDA is bundled with an appropriate graphics driver, but at least the version number changed after the installation. Nevertheless, I performed the following commands:
Replace the 390 with the version you would like to install. To verify the driver installation, you can use the following command:
The output should be non-empty in case of success. A restart after the driver installation might be required.

To be able to use the nvidia-smi command, I added the following lines to my .bashrc script:
Again, replace the version number with the version you installed previously.

The NVIDIA system management interface might still fail with an error. This is because you have to manually switch to your dedicated GPU first. This can be done using the following command:
As far as I know, this is kind of the Linux counterpart to NVIDIA Optimus on Windows. Enter the nvidia-smi command again, and you should get the anticipated output.
 

Remember: The prime-select command has to be executed every time you reboot your system and plug-in your eGPU via Thunderbolt 3.
 

3. Setup TensorFlow with GPU support

The installation requirements for TensorFlow with GPU support can be found here. However, I found the installation is still a bit tricky, especially because it is important to install the exact CUDA and cuDNN versions. That's why I want to share the commands and links I have used nevertheless. In my case, I installed TensorFlow r1.6 via pip, which requires CUDA 9.0 and cuDNN 7.0. In my first try, I installed CUDA 9.1 and cuDNN 7.1, and it did not work. Importing TensorFlow in Python resulted therefore with a warning that the appropriate libraries could not be found.

3.1 Download and install CUDA 9.0

CUDA version 9.0 can be found here, while the latest version is available here. I installed the main binary first, then followed by both batches. To install the binary, I used the following commands:
Furthermore, add something like this e.g to your .bashrc script:

3.2 Download and install cuDNN 7.0

Navigate to the cuDNN page, register, and download the appropriate cuDNN version. In my case, I installed version 7.0.5, which was the latest version of the 7.0 release. Afterwards, you should set the CUDA_HOME environment variable, such as in your .bashrc script:

3.3 Install GPU-enabled TensorFlow

I usually recommend to use a virtual environment for every project:

Hell yeah, we are ready to rumble...!

 

[출처] http://bsautermeister.blogspot.jp/2018/04/how-to-setup-egpu-on-ubuntu-for.html