39 thoughts on “Cloud GPUs Tutorial (comparing & using)

  1. Do not bother with Linode. They do not even have any GPUs available. Such a waste of time. Here's an email I received from their support.
    "At this time, we do not have any GPUs available. Additionally, we typically need to see at least three months of billing history before granting GPU access."

  2. I tried requesting access to GPUs on Google Cloud and was denied about 2 minutes after submitting my request, despite having my google account for ~18 years. AWS it is I guess….

  3. Did anyone try a game streaming provider, which provides a full desktop environment, to train NNs?
    Eg shadow offers a GTX 1080 comparable card for only 13€/15€/month
    The card isnt the fastest, but compared to the cloud gpu providers they seem (for me) to be very cheap. For longer training sessions i could use a raspberry pi to keep the session open.
    Also i could obviously use that provider to play some games 😀

    Did i miss something or why is nobody doing this?

  4. This is the best videos on this topic available! Thank you so much! I just tried this out. Not sure it was like this a year ago but now you can just "upgrade" & "downgrade" by clicking on a button.

  5. Maybe the transfer rate you are used to is 100 mega bits / sec, and this is just 12 mega BYTES per second. I believe a megabit is 1/8 the size of a megabyte, so 12*8 ~= 100 mega bits / sec.
    Again, could be wrong, but I think the standard is Mb is megabit, and MB is megabyte.

  6. Here you can see all the services for different places: https://www.linode.com/global-infrastructure/
    If I see correct, only Newark has GPUs at the moment.

    Update: I just had contact with their support, Mumbai also has GPUs

  7. I’m completely new to training models on a gpu. So is the purpose of this to come up with your model architecture on your local machine… over like a Jupiter notebook. Then when you are ready to pre process data and train your model…you move your info to the GPU and do those actions on the GPU? Then once your model is trained you shutdown the GPU and can use the trained model elsewhere?

  8. Fantastic tutorial, thanks!!! Just one question: when you tested on modelfit.py, you mentioned that we may want to do those np.array data preprocessing on another server. Could you provide a tutorial on this topic? The reason I am asking is because data preprocessing is a big part of deep learning exercises. And a lot of times I am using paperspace GPU servers to handle it (I don't know other efficient approaches)

  9. cudaGetDevice() failed. Status: cudaGetErrorString symbol not found

    Can you help me with my error? I'm trying to follow your tutorial with the cats-and-dogs dataset and i get this error when i try to run the CNN. I have a 1050 GTX

  10. rsync with compression is faster than scp.
    Also, your transfer speeds seem to be limited by the read speed of the spinning rusts.

  11. For the scp you were using the public ip of the server so the traffic had to go through more hops than if you were using internal ips, which I guess that you could have created as the servers are in the same lan.

  12. 😀 I am using google colab, it is pain to use but is free and you can even mount google drive on it + it comes with pre installed machine learning libraries and stuff. And tesla gpus

  13. Linode has internal IP’s. Under the network tab in dash. For me at least It works way way faster when I do scp.

  14. What about spell.run? You can just code locally, but when you want to run your code on the cloud, you just add "spell run" before your "python script.py" and it will only charge you for the run.

Leave a Reply

Your email address will not be published. Required fields are marked *