Multi GPU setup

   2670   7   2
User Avatar
Member
37 posts
Joined: Oct. 2009
Offline
I'm building a new rig and am trying to understand the most optimal way to build in multiple GPU's for sim and rendering. I understand only one GPU can be used for OpenCL, and one is used for viewport. So I'd like to get some advice if it's possible to mix and match different cards to do this, or if they need to be relatively similar or the same card.

Ideally I'd like to use my 2 old cards, and then eventually add something beefier as well:

660ti 3gb - viewport card
1070ti 8gb - for sim and rendering
Titan RTX 24 gb - TBD on this last card, but it will be something big for sim and rendering

I'm being told on the Tom's hardware forum that using these cards will basically result in bottlenecking everything because the 660ti is so old. And in general, a rig with multiple GPU's will always get throttled down to the slowest GPU.

My original thinking was that each card would be used totally independently. But I'm really not a hardware expert. So any advice on the most optimal way to build a multi GPU rig would be greatly appreciated.

Is there a way to mix different GPU's or will that throttle everything down? For example, if the above combo wont work, would combining the Titan RTX with a newer card like a 2060RTX 6gb work? Use Titan for sims, and the 2060 for viewport?

Thanks!
User Avatar
Member
73 posts
Joined: March 2020
Offline
I think your answer is going to depend on which GPU renderer you are using. If this is about Houdini/Mantra only, I doubt the GPU config is going to make a whole lot of difference. I have a dual GPU setup, so I set up Houdini to use one GPU for the display and the other for OpenCL, and the GPU that gets used for OpenCL hardly ever goes beyond 25%ish utilization. By contrast all my CPU cores are maxed out even when doing a simple Mantra render. In my situation, the two GPUs aren't talking to each other, so I don't see why one might throttle the other. I can see bottlenecking happening if you are constantly shuttling data between your GPUs, which isn't something that happens in a vanilla Houdini/Mantra setup.
User Avatar
Member
818 posts
Joined: Jan. 2018
Offline
Get rid of the old GPU…it's not going to benefit you nearly as much as you think to have a separate GPU for display vs. a single one for both display and computing.

Also, be aware that at the moment the Houdini Sparse Solver doesn't use the GPU, so personally if I was in the market for a new system and I was planning on doing a lot of sims, I would opt for a faster CPU like a 16 or even 32 core Ryzen and at least 64Gb of fast RAM.
>>Kays
For my Houdini tutorials and more visit:
https://www.youtube.com/c/RightBrainedTutorials [www.youtube.com]
User Avatar
Member
7 posts
Joined: May 2020
Offline
More GPUs the better. If you want full performance out of all GPUs make sure to have a Motherboard that supports multiple simultaneous 16X PCI-E ports otherwise most boards will automatically force all slots to work at 8X mode once an additional GPU is installed.
i7 4Gz 32 Mb | RTX 2080 Ti | Houdini FX 18 | Fusion Studio | Sonar Platinum | Zbrush 2020 |
User Avatar
Member
4 posts
Joined: March 2017
Offline
Ascensi
More GPUs the better. If you want full performance out of all GPUs make sure to have a Motherboard that supports multiple simultaneous 16X PCI-E ports otherwise most boards will automatically force all slots to work at 8X mode once an additional GPU is installed.


Just curious if this motherboard would work with multiple gpus?
https://rog.asus.com/ca-en/motherboards/rog-zenith/rog-zenith-ii-extreme-model/spec [rog.asus.com]

Does Karma XPU work with multiple different model gpus from different manufacturers ( for example Asus ROG Strix 2080 Ti + Evga 3080 Ti )?
Edited by stephentywong - Jan. 20, 2022 03:44:47
User Avatar
Staff
187 posts
Joined: May 2019
Offline
stephenvfx
Does Karma XPU works with multiple different model gpus from different manufacturers ( for example Asus ROG Strix 2080 Ti + Evga 3080 Ti )?

Yes it should
User Avatar
Member
24 posts
Joined: March 2013
Offline
zicónico
I think your answer is going to depend on which GPU renderer you are using. If this is about Houdini/Mantra only, I doubt the GPU config is going to make a whole lot of difference. I have a dual GPU setup, so I set up Houdini to use one GPU for the display and the other for OpenCL, and the GPU that gets used for OpenCL hardly ever goes beyond 25%ish utilization. By contrast all my CPU cores are maxed out even when doing a simple Mantra render. In my situation, the two GPUs aren't talking to each other, so I don't see why one might throttle the other. I can see bottlenecking happening if you are constantly shuttling data between your GPUs, which isn't something that happens in a vanilla Houdini/Mantra setup.



hello how are you, could you give me the configuration of the houdini ENV, I have two RTXA 6000 and I would like to leave one for rendering and another for simulations with OPENGL I have hudini 19.5
Edited by olympusRaul - March 23, 2022 12:04:47
Artist CGI
User Avatar
Staff
2408 posts
Joined: Sept. 2007
Offline
olympusRaul
I have hudini 19.5

I hope that's a typo. Did you mean 19.0 or 18.5?
Chris McSpurren
Senior Quality Assurance Specialist
SideFX
  • Quick Links