CUDA dependence in DLC

Given the difficulty of obtaining the 3000 series NVIDIA GPUs, the rumored performance parity by AMD’s latest Radeon GPUs and the lack of price drop for previous generation cards :

(A) Is there some custom part of DLC that is natively written in CUDA, explaining its reliance on NVIDIA GPUs?

(B) Is it possible that DLC might run on an AMD card with a TensorFlow port [1,2]?

© Are there other CUDA reliant libraries in DLC besides TensorFlow?

references :

[1] https://github.com/ROCmSoftwarePlatform/tensorflow-upstream
[2] https://rocmdocs.amd.com/en/latest/InstallGuide.html

Sorry for the the delayed response. Indeed the only aspect of DLC that requires CUDA/NVIDIA is TensorFlow (see here for more details on NVIDIA and TF: https://www.tensorflow.org/install/gpu).

Unfortunately, I have no experience with (trying) to use TF on Radeon cards. But maybe somebody else, has…

The 30x series indeed starts at ~500$, https://www.nvidia.com/en-us/geforce/news/geforce-rtx-3070-out-now/
(they are great cards!). The previous series are still available too, and are a bit cheaper… Perhaps a great alternative is using Google Colaboratory (you can use GPUs for free up to 6h at a time, or get a pro account for 10$ a month: https://colab.research.google.com/signup). Examples for how to use DLC on Colab are here: https://github.com/DeepLabCut/DeepLabCut/tree/master/examples

Sorry that I don’t have a more helpful answer