Muti-GPU support for TensorFlow in M1

Question

mmenendezg OP

Created Dec ’21

Replies 1

Boosts 0

Participants 2

Hello,

I have read in this Apple's post that the option for Multi-GPU is currently not supported. I would like to know if this means that currently is working on only one GPU core, and if this means the performance for the training of the models will improve in future updates.

Thanks in advance.

Answered by Frameworks Engineer in 698348022

Hi @mmenendezg, We utilize all the cores of GPU in M1 machines. The GPU appears as a single Device in TensorFlow which gets utilized fully to accelerate the training. Multi-GPU is referring the something else when multiple GPU cards are attached to a single Host which is different from utilizing all the cores in a single GPU. In essence, all GPU cores would be accelerated on M1. If you do see some network where its not the case, please provide an example so we can reproduce the issue.

Boost

Answer 1

Frameworks Engineer OP

Apple

Dec ’21

Accepted Answer

Hi @mmenendezg, We utilize all the cores of GPU in M1 machines. The GPU appears as a single Device in TensorFlow which gets utilized fully to accelerate the training. Multi-GPU is referring the something else when multiple GPU cards are attached to a single Host which is different from utilizing all the cores in a single GPU. In essence, all GPU cores would be accelerated on M1. If you do see some network where its not the case, please provide an example so we can reproduce the issue.

1