MATLAB Deep Learning Container on NVIDIA GPU Cloud for NVIDIA DGX

I read the post, 'MATLAB Deep Learning Container on NVIDIA GPU Cloud for NVIDIA DGX'.

It says host machine and client machine and seems to be needed the internet connection

for downloading MATLAB containers.

But in my environment, the network connection is not available.

Could you give me the way how to set the 'standalone' environment of MATLAB deep learning in NVIDIA DGX systems?

I am also wondering about :

Can I just install MATLAB in Linux version on DGX system (with DL, Parallel Comput toolboxes) to use MATLAB deep learning?

Thank you

Answer 1

[Jason Ross]

You would just install MATLAB on the system and it should find the GPUs and work, since the DGX is really just a Linux system.

[DaeYoung]

According to your answer, I summarized what I have to do :

1) download the MATLAB (2020b) iso file for linux OS.

2) transfer the iso file to DGX system (DGX station) and install

3) The installed MATLAB automatically finds the GPUs in DGX system. Then, the command 'gpuDevice'

will exhibit the message about the information on GPUs MATLAB can exploit.

4) Use deep learning toolbox and others for creating deep learning network models.

Are those things right?

I have additional two questions :

a) I have several network models and matlab codes (m-files) used in Window OS. Are they executable in

MATLAB for linux OS without any actions?

b) Can MATLAB with a version of 2021a be executable in DGX system? The reason why I am wondering

about it is there is no container package in the relevant NVIDIA homepage.

I am really appreciated with your interest in my question. Thank you!

[Jason Ross]

1,2) Yes, you need to download the full install since you can't get to the Internet.

3) assuming that "nvidia-smi" works from a command line this should work. I would expect if the DGX system ships with some version of Linux, nVidia has also installed the relevant drivers for the GPUs in the system.

4) yes

a) a qualified "yes". Just plain old code is fine and will execute. The caveat is if your code includes some sort of OS-dependent actions, like file paths that you have not abstracted using fullfile/filesep or some sort of system() commands that are unique to that operating system ("ls" vs. "dir", etc).

b) I suspect that this qualification and release is underway and it's just taking some time to make it into NVIDIA's systems. 2021a can be installed as a full application in the same manner as 2020b.