Question What NVLINK should be used for 2X3090 Palit 2.7 slot on a X570 MB?

Twenkid

Prominent
Sep 11, 2021
5
0
510
I work remotely with the machine. These are the cards AFAIK:

According to this scheme:


And these measurements from my colleague and the pictures:

It seems the distance between the connectors is about 80 mm, which is like 4-SLOT? That was my guess according to other pictures of sample configurations, such as:

It also looks like 4-slot?

However my colleague told me that a technician from his computer shop suggested that it was 3-Slot, because the distance between them was 2-spacers (but between which parts).


I guess the technician is wrong, but what's the truth?

Thanks

 

Eximo

Titan
Ambassador
4 slot looks right to me. Even that other picture shows dual two slot cards with a 2 slot gap, so looks like they are using a 4-slot bridge there as well.

If it were three slot spacing you wouldn't be able to see that 1x slot between the cards.

Each card is 2.5/3 slot, you have a gap of one slot, 3+1 = 4.

Now I know Puget says it is supported, but there aren't a whole lot of SLI capable motherboards these days.
 
  • Like
Reactions: Twenkid

Twenkid

Prominent
Sep 11, 2021
5
0
510
4 slot looks right to me. Even that other picture shows dual two slot cards with a 2 slot gap, so looks like they are using a 4-slot bridge there as well.

If it were three slot spacing you wouldn't be able to see that 1x slot between the cards.

Each card is 2.5/3 slot, you have a gap of one slot, 3+1 = 4.

Now I know Puget says it is supported, but there aren't a whole lot of SLI capable motherboards these days.

Thank you for the comment and reasoning! My colleagues who are with the machine also realized it looks more like 4-slot and they have ordered one.
 

Twenkid

Prominent
Sep 11, 2021
5
0
510
Unless you are doing ridiculously complex 3D modeling even then it will be questionable benefit, dual cards will do nothing except heatsoak the other.

The purpose for now will be ML and robotic simulations with ROS. The double RAM may be of use for running some huge models or working with many models at once for different tasks. Re 3D-modeling - right now it is not an application except for simple simulations, but I guess in the future it could be for 3D-modeling and game engines for high quality simulations as well, and I guess Omniverse or Unreal Engine 5 may make a good use of the cards.

Currently it is with Gazebo, which alone is known not to use the GPU much for physics simulation and here I agree that we didn't need even a single 3090, however probably we may evolve into Nvidia's Isaac libraries which utilise TensorRT etc., we may dedicate each card for different tasks, e.g. one running simulations, the other one training or running DL models etc.

I also don't know about whether the SLI would give much of an advantage, I personally suggested that we should better wait until we see what we can do without it.

You're right about the heat, my colleagues added water cooling for that reason.

Regards
 

TRENDING THREADS