2 GPUs in one computer

[eric_s]

I've been wanting to get into AI for quite some time now but haven't had the time until recently. I've been trying to get tensor flow and Pytorch to work with my Radeon 6650xt but found out that ROCm isn't official compatible with it. I have an old Quadro card and know that I can at least get some version of cuda on it so I can get started (if this works then I'd like to buy something like an Nvidia Tesla card which has no ports). I'm curious if I can use two GPUs in my one system with the 6650xt for gaming and the Quadro card for AI stuff. Is there a way that I can select the Quadro card to run simulations without having any displays plugged in or is there a way that I can run it through my CPUs integrated graphics?

 

Here are the specs for my system:

 

CPU: Ryzen 7600x

Motherboard: ASRock B650 PG Lightning

RAM: 32gb DDR5 Trident Z running at 6000mhz

Power supply: Thermaltake GF1 850 Watt Gold

Storage: 1tb Samsung 980 NVMe SSD

GPU: Radeon 6650xt

GPU 2: Nvidia Quadro 4000

 

Quick note: I did all of the calculations on power draw and I expect to be more than good especially since I don't ever plan on pinning all three chips at the same time.

 

Please let me know what you think!

 

Thanks I'm advanced!

[TheJudge91]

Maybe not the answer that you are looking for, but if you are starting using TF and Pythorc and you want just "play" with AI why don't you use google colab? 

Pytorch start

https://colab.research.google.com/github/omarsar/pytorch_notebooks/blob/master/pytorch_quick_start.ipynb

TS start

https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/quickstart/beginner.ipynb

 

 

Anyway...

You should have env varaibles called CUDA_VISIBLE_DEVICES that mask the GPUs that python can use

You have to use it in some of these ways:

eg: Let python see only the device 2 (1 index)

1) CUDA_VISIBLE_DEVICES=1 python myscript.py 

 

2) in the myscript.py

import os

os.environ["CUDA_VISIBLE_DEVICES"]="1"

 

3) export the env before run the script

export CUDA_VISIBLE_DEVICES=1

python myscript.py

 

Of course you can set more then one GPUs 

 

Hope it helps!

 

[eric_s]

6 hours ago, TheJudge91 said:

Maybe not the answer that you are looking for, but if you are starting using TF and Pythorc and you want just "play" with AI why don't you use google colab? 

Pytorch start

https://colab.research.google.com/github/omarsar/pytorch_notebooks/blob/master/pytorch_quick_start.ipynb

TS start

https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/quickstart/beginner.ipynb

 

 

Anyway...

You should have env varaibles called CUDA_VISIBLE_DEVICES that mask the GPUs that python can use

You have to use it in some of these ways:

eg: Let python see only the device 2 (1 index)

1) CUDA_VISIBLE_DEVICES=1 python myscript.py 

 

2) in the myscript.py

import os

os.environ["CUDA_VISIBLE_DEVICES"]="1"

 

3) export the env before run the script

export CUDA_VISIBLE_DEVICES=1

python myscript.py

 

Of course you can set more then one GPUs 

 

Hope it helps!

 

Hey TheJudge91,

 

Thanks for the quick response! I have considered using Collab and am not against it but I would eventually like to get it working on local hardware.

 

That makes sense. Do you think this solution will work in WSL2 running Ubuntu 22.04 LTS? Would I need to and an environmental variable to Windows?

 

Thanks a bunch for the advice and help!

[TheJudge91]

1 minute ago, eric_s said:

Hey TheJudge91,

 

Thanks for the quick response! I have considered using Collab and am not against it but I would eventually like to get it working on local hardware.

 

That makes sense. Do you think this solution will work in WSL2 running Ubuntu 22.04 LTS? Would I need to and an environmental variable to Windows?

I don't have any experience on Ubuntu 22.04 but i'm pretty sure that it will not cause problems.

 

The option .1 e .3 in windows should be with the set command

1W) 

set CUDA_VISIBLE_DEVICES=1 & python my_script.py

3W) 

set CUDA_VISIBLE_DEVICES=1

python my_script.py

 

The option .2 is OS agnostic because part of the script and i suggest this kind of programming. You can create a .config file that you can load in the script.