Small GPU Render farm

[sheehan_partners_nm]

Hi,

GPU rendering is getting better and better and faster and faster, so I was thinking to invest in a mini GPU render farm vs CPU. I saw the list of unsupported features https://docs.chaosgroup.com/display/...orted+Features and I am fine with that since I barely would use any of the unsupported ones for architectural visualization.
As a new user of Vray RT, I wanted some suggestion for the render farm from who has more experience.

1- I would use a single box with 4 gpu in SLI mode. Is SLI fully supported nowdays, and problems with that?
2- Should I go for Quadros or 4x1080GTX (maybe Titans?)
3- Not clear how Ram is handled...for a big scene, I would have 4x 8GB of VRAM for each card. I think that's plenty to handle every scene. How is the the traditional Ram used during GPU rendering?

The final goal is to use GPUs in production mode, for final renderings and not for realtime previews.

Thanks you all in advance.

[urbanite]

I am not RT expert but

1. no benefits from SLI, also might be problematic
2. I would go for TITANS because they have more RAM on board but they are much more expensive than gtx being only few percent faster (considering the same architecture)
3. RAM does not sum up so in your case the scene would have to be not bigger then 8GB

[sheehan_partners_nm]

I didn't knew that Ram does not sum up! That's a pity! Also SLI is a problem? that's bad....does not make sense to buy separate motherboards and towers just to have single GPUs mounted. Beside all the hardware problem, that would ask for rendering nodes too! I'm not sure about this, there are plenty of videos around and post here in the forum with 4 GPUs connected. Hope someone from the developers team could brighten our minds
Thanks for the response!

[yavor.rubenov]

Hello,

About SLI - there are no benefits and no need for the SLI connections between cards - meaning you don't need the separate SLI cables between cards. Having more than one GPU in your system is supported and V-Ray RT GPU can use them.

Best regards,
Yavor

[sheehan_partners_nm]

Hello,

About SLI - there are no benefits and no need for the SLI connections between cards - meaning you don't need the separate SLI cables between cards. Having more than one GPU in your system is supported and V-Ray RT GPU can use them.

Best regards,
Yavor

Yavor, are you saying that I can simply plug 4 1080 in my machine without having to make them communicate with bridges between each other? I just saw that Nvidia only allows 2 way connections now.
Also I read that Pascal architecture allows now to add/stack the memory between cards? Is that true? Will 4 GPUs provide 32gb of memory ?

Thank you!

[yavor.rubenov]

Yes - simply plug the video cards and don't add any bridges. The bridges themselves are used only when gaming.

About Pascal and stacking - there is a special technology - NVLink that allows GPUs to communicate with each other faster but it will only be available in the P100 chip which I think is still not available.

Best regards,
Yavor

[sheehan_partners_nm]

Yes - simply plug the video cards and don't add any bridges. The bridges themselves are used only when gaming.

About Pascal and stacking - there is a special technology - NVLink that allows GPUs to communicate with each other faster but it will only be available in the P100 chip which I think is still not available.

Best regards,
Yavor

All clear, thank you a lot!

[Alan Iglesias]

Yavor, are you saying that I can simply plug 4 1080 in my machine without having to make them communicate with bridges between each other?

Currently there is no reason for them to "communicate". Each GPU stands on its own as a separate rendering entity and is controlled entirely by Vray RT/GPU. If you have four GPU cards in one computer, each one will contribute the to the rendering. You will likely find that the GPUs will scale pretty much in a linear fashion.

...Also I read that Pascal architecture allows now to add/stack the memory between cards? Is that true? Will 4 GPUs provide 32gb of memory ?

Since each GPU stands on its own when rendering the scene, it requires the entire scene to be loaded in its native RAM where it can access it directly. Very efficient. Very fast.

-Alan

[sheehan_partners_nm]

Thank you Alan, starting to understand now how GPU rendering works, looks very interesting. This ram limit is a big minus though...It is very easy now days to go over 8gb or Ram with big scenes. I did another test and my 1080 couldn't handle it, crashed and crashed. I guess is yet early to invest in GPU only for production mode.

[Alan Iglesias]

Thank you Alan, starting to understand now how GPU rendering works, looks very interesting. This ram limit is a big minus though...It is very easy now days to go over 8gb or Ram with big scenes. I did another test and my 1080 couldn't handle it, crashed and crashed. I guess is yet early to invest in GPU only for production mode.

Well, yes, the whole scene has to fit on each GPU's RAM - but the good news is that compared to Vray Advanced rendering on the CPU, all things being equal any particular file should take much less RAM to render on the GPU. This is because ONLY the rendering file is sent to the GPU whereas in normal CPU rendering you are already taking up potentially plenty RAM with your applications, OS, loaded file, etc., etc.

Since my current GPUs have only 3GB RAM on them, I often use normal rendering for ArchViz production, but the GPUs are still fabulously useful when setting up lighting, materials, composition, etc.

And for product visualization, it's pretty much GPU all the way...

Best of luck,

-Alan