Is PyTorch fully functional without cuDNN?

Hey folks,

Currently I’m an intern at a company to do some research into the application of machine and deep learning. I’m in talks with my supervisor to get a more powerful machine at my desk, but I fear that it may not come through and I would be resigned to using whatever I have available right now. As such I’m trying to figure out my options in case I can’t get a proper rig going.

The problem is as follows: the company has given me a standalone PC with a 2GB NVIDIA Quadro 4000 graphics card. However, this card has compute capacity 2.0, which means that I am able to use CUDA but not cuDNN. I get that my code will be slower without cuDNN, but I can’t quite determine if I would run into problems when I would try to implement a more complex network without cuDNN (e.g. incompatibilities with network layout or certain functions). Would I be able to do everything using PyTorch without cuDNN, or do I really require cuDNN for certain functions?

Thanks in advance!

1 Like

hi Alex, with card graphic compute capacity 2.0, pytorch won’t run with cuda so it very slower training big data
i think only theano run cuda library but it don’t run cudnn library (library for depth neural network)
i also have nvidia 635M, compute capacity 2.0, should when run training example in github very slow
can i make friends with you?

3 Likes

Thanks đàm!

For those of you finding this thread in the future: I found the following blog post on GPU selection. TL;DR Don’t cheap out like I wanted to - invest in a card that’s suitable for the job. Convolution operators are performed using cuDNN, so it makes sense to require compute capacity 3+, and since Kepler cards are allegedly quite slow you’ll be looking at Maxwell or more recent cards.

And of course you can be my friend đàm! :wink:

my twitter is https://twitter.com/tai94bn
can you add ?