Welcome to our Community
Wanting to join the rest of our members? Feel free to sign up today.
Sign up

Question Anything wrong in AMD GPUs ?

barackobama

Active Member
Jul 25, 2020
61
3
2,300
Are there any disadvantages in buying AMD GPUs ?


Like hardware wise Amd RX 5700XT is little better than 2070 super. It wont have cuda, but using prebuilt libraries like tensorflow/ Pytorch would the performance be same ?


 
Last edited:
  • Like
Reactions: pr0ph3t
Abhiram Shibu Abhiram Shibu I mean this

According to here
Machine learning researchers do not program GPU code anymore than data scientists write BLAS libraries.

They find new ways to implement either existing networks or to create new ones.

What frameworks like Tensorflow do is to provide a set of primitives that can be used to construct neural network models. And the guarantee that this will run on a given hardware with acceptable performance.

You do not need to write a single line of CUDA or OpenCL in your day to day as an ML researcher or “developer” in fact if you end up doing so something went terribly wrong along that way.

ML devs/researchers will not write libraries or optimize GPU code that is well outside of their scope of work and capabilities.

Think of Tensorflow as Matlab, you aren’t going to expect someone who is using Matlab to optimize the execution of transcendental algebra on a given architecture or to add new functionality.

AlphaZero was written without using a single new line of CUDA or OpenCL or anything of the sort, their neural network was built on-top of Tensforflow and the magic of Alpha isn’t even its DNN but rather the Monte Carlo search tree algorithm they developed (which runs on normal CPUs and implemented in Python) which finds the best move out of a pool of moves learned by their trained network.

What ML researchers need is a working TF, Pytorch, Caffe2, etc. not CUDA or OpenCL support or anything of the sort as they would never touch it nor should they.
 
  • Wow
Reactions: pr0ph3t