Ml-agents: how fast is it to use gpu compare to use cpu instead?

[tanakataiki]

mlagents v7 has gpu option in Learninig Brain but in previous version gpu is useless and not reccomended to speed up.
but to search into inference device in LeaningBrain.cs eather compute fast or CS sharpfast is selected .
So I assume if CPU mode Script excution time is shorten while GPU mode enhance backpropagate time.
Am I right about it?

Thanks

[ervteng]

Hi @tanakataiki, two different things. In previous versions, inference _only_ supported CPU. In 0.7, inference supports both CPU and GPU. Typically, for small networks, CPU is OK and not worth the overhead of sending the data into GPU memory and taking it out again. The mode in the script _only affects inference._

For training, (e.g. for backpropagation), the version of ML-Agents is irrelevant, as it is entirely done in Tensorflow. You can use either Tensorflow or Tensorflow-GPU with ML-Agents; both will work.

[tanakataiki]

@ervteng
Does inference mean the out put of net work and net work is also constructed in C#?
I am wondering what gpu helps so far.
As for the back propagation does gpu helps for training large net work such as 1024 or 512 of two or more layer?
Thanks

Edited

I understood that it's only vaild for inference not training that is enabling nothing to do with training since there are no validation for mlagents. So far if the net work isnt large enough and agent isnt too many enough then cpu is enough.

Am I right about it?

[MarcoMeter]

Inference means that you are not training your model any further. Instead you use it for production (i.e. producing actions in real-time).

GPU training needs lots of data. As PPO generates data on the fly, the overhead, caused by the interface between CPU and GPU, slows done the training process.

[tanakataiki]

Thanks @MarcoMeter
I closed this issue

[lock[bot]]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.