Nvidia RTX Series

Usage & Issues
#1

Hi!

I have a question related to the usage of GPUs in the process of learning and analyzing in DeepLabCut.

For this task, the authors recommend NVidia GeForce GTX models, which are compatible with Cuda 9.0 and Tensorflow up to 1.10.

However NVidia discontinued the GTX line in November of 2018 and the GPUs are often not available anymore! The last remainders are sometimes sold for more than their successors currently cost new.

Recently the new RTX line was introduced by NVidia. To my knowledge the new GPUs only support Cuda 10 and Tensorflow above 1.10. (1.12?)

Is it possible to run DeepLabCut with the new models as well (e.g. RTX 2080, RTX 2080 Ti?) or will it be possible in the near future?

I tested DeepLabCut with the CPU so far and it works great on my data. Now I want to speed up the process, however I have trouble in obtaining the GTX 1080 (Ti) GPUs…

Thanks’ a lot!

Btw, are any workshops planned in the near future?

Cheers :wink:

#2

Hi Peter,

I have a RTX 2080, I had some problems at the beginning but now it works well for me to run DLC.

Windows 10
CUDA 10
cuDNN 7.3.x
Tensorflow 1.12 (built from source)

Cheers!

#3

Thank you for the reply! Okay this are some good news, so I will be able to use the RTX. Still have to learn how to build tf from source, however I think there are some nice tutorials out there.

#4

Yes, there are some great tutorials. If you are using Windows, here is the link to download the tf 1.12 Wheel:
https://storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-1.12.0-cp36-cp36m-win_amd64.whl

#5

RTX and Titan series are all fine. The GPU just must be CUDA compatible, so unfortunately not AMD GPUs.

#6

I’ve been running DLC1.1 with driver 418 + cuda10 since mid January using tf-nigtly-gpu (no need to compile). Works without issues as long as you don’t get card that dies in a weekend