How to set up an Nvidia GPU with CUDA


(Samir Moussa) #1

I’ve recently posted a blog post on how to setup your own Nvidia GPU with CUDA. Might be helpful if you’re building your own box!

Battling through an Nvidia GPU setup in the name of machine learning.


#2

does TF / PyTorch work with cuda 9 ?
IIRC they need cuda 8 ?


(Samir Moussa) #3

I’ve been using PyTorch extensively with CUDA 9. I’ve also run a few examples with Keras having TF as the backend. No issues so far!


#4

thats nice to know. Thanks for informing.

Btw , on a tangent.
Are you able to get virtual gl set up on a headless server ?

A lot of DL stuff requires Open GL , and i just cant seem to get that to work.
Perhaps that could be your next post! :wink:


(Samir Moussa) #5

Hmm interesting. I’ve not tried any of that yet. Sorry :sweat_smile:


#6

A big problem with jupyter is that , once you sever connection to it, the outputs wont be displayed. This is because the outputs are sent to a browser to be rendered, then that is sent back to server for storage.

So i was trying to set up a browser on my headless setup , and then use VNC to control it. Effectively that would be a proxy browser that would always maintain connection to the jupyter server, and I could disconnect.

But open gl acceleration seems to not work for this


(Samir Moussa) #7

So you’re saying that even if you’ve started a jupyter notebook server and connected via an SSH tunnel, then using the client browser to use the notebook, you don’t get any output? What are you trying to output? But if you use OpenGL, rendered graphics are started in a new window and your VNC cannot view that?


(Matthijs) #8

I think what he means is that, if the connection between the client and server is broken for some reason (client went to sleep, for example), then you can re-connect to the notebook. But if the notebook is still producing output, you will not see that new output (it will not be sent to the client). This is a known limitation of notebooks.


(Brian Muhia) #9

If you build pytorch from source (in a conda environment) you should be able to use CUDA 9 and CUDNN 7. This is my current setup as I take the part1v2 course.


(Erin Pangilinan) #10

Just curious, what did you want to use virtualgl for specifically?


#11

Browser with opengl acceleration.
Some RL stuff that requires games to run on open gl