2x3090 RTX or 1x4090 RTX for training a model?

Hi! I have a doubt regarding the combination of GPUs for training the models. I would like to use two 4090 RTX, but I saw that they do not have the NVLink capability. I do not know if it is possible to train a model by using two detached 4090 RTX since they do not have this capability. If I am not wrong, 3090 RTX has this capability (thanks to NVLink), so I suppose it is possible to combine them and use the two-combined GPUs for the training of the model.
If it is not possible, is it better to use 2x3090 RTX with NVLink or only one 4090 RTX? Otherwise, do you have any other proposals?

Thanks :slight_smile:

There’s no need for the NVLink to train models using 2 RTX 4090. Even for the 3090s, the NVlink will speed up p2p communication a bit only, not resulting in a big performance gain.

I’m using 2x 3090 with the NVlink, but considering a single 4090 since it simplifies things.

If you have the budget for 2x 4090, this is the correct choice today.

1 Like