is it worthed to install more than 2 GPU ?

From what i know so far in gaming world , the 2 gpu is the most optimal setup (a lot of game even doesn’t works well with multi gpu). How about for 3d app (blender and other 3d) ? is the same rule applied ? Maybe the more gpu added will increase the render speed , but the enhancement graphic is not linear , but the price is always linear. So how effiecient it is to have more than 2 GPU ? Any has experience with 3 or 4 gpu , how many percentage the enhancement from 2gpu to 3gpu ?

I plan to use 3 x 1080 ti but may cut to 2x if it’s not good option, any advices will be great. Thanks a lot.

The following was on our systems for rendering animation

1 GPU in 1 machine = 100%
2 GPUs in 2 machiens = 200%
2 GPUs in 1 machine rendering the same frame (using blenders multigpu support) = 140%
2 GPUs in 1 machine rendering different frames = 189%

this drop off would be less on more powerful cpus as it would have been limited via CPU usage at bvh building / post processing times.

totally worth it. if you can buy 3 1080ti then DO IT!

the biggest things to consider:

does your PSU have adequate power (at least 1.5x, 2x preferred, of total load)?

is there enough room in your case? is it well vented?

SLI or Crossfire is quite inefficient, having one GPU per tile is far better.

not really relevant to your situation, but the whole scene needs to be able to fit in the each cards ram.

The improvement is definitely linear, if you are not using SLI, as Daedalus said. That is really only applicable for games. With multiple GPUs you can have each GPU rendering a separate tile so 3x 1080s are 33% faster than 2x 1080s.

I disagree… It is not 100% faster. As i broke it down in my post above… using multiple cards on separate tiles is not the most efficient method. Maybe for still images, but for animations it definetly isnt.

That was a slight inaccuracy on my part, I apologise. Theoretically, you should get a linear relationship if you aren’t limited by CPU. On that matter, what would you consider the most efficient method? I suppose it would be faster to use separate systems on Network Render?

I somehow cna’t agree with you. The 140% improvement doesn’t look good to me at all. When I went from 1 to 2 GPU’s I saw about 185% improvement. but i’ll double check. Currently I’m running 4 RX 480s so i’ll post each improvement using blenders classroom scene.

what setup do you have? and what scene did you use? and what settings (tiles) did you use?

The 185% improvement seems realistic. A 200% improvement is pretty much an unattainable ideal but the improvement should be better than 140% if there is no significant CPU bottleneck.

My system is a GTX 1070 + GTX 1060 6GB and 6700K. I normally use 256 tile sizes with the Auto Tile Size default addon.

It was a production scene, 1280x720, 640x360 tile sizes 100samples, 2x gtx1080, core i7 860. Old cpu yes but the principle still is the same, If gpu’s are idle, then it wont be accelerated.

Larger renders / smaller tile sizes will keep the gpus active for longer / limit the problematic tiles. but again, it is best to run multiple blenders with a single gpu on each then one blender with many gpus.

Thanks for the feedback. I hope I’ll have time to render the classroom scene from blender across my GPU’s, especially 1 to 2 GPU’s.

But yeah, tiles are a pain to setup to utilize the GPU’s correctly. I was hoping that recent work with Cycles would resolve that since they were keen on that. But as you said, you have 2x GTX 1080s, so unsure if CUDA was also improved when it comes to tiles.

I do have to agree with your statement about each gpu rendering single scene. but only when working on an animation sequence. Even if my 185% ratio would be accurate, that would leave 15% unused. and when you have a 10 minute animation to render, ever second counts.

And a question for metacla.

Are you more, rendering full animation, or rendering still frames?

So quick … long… render of the Blenders Classroom scene. (all default settings - except tile changes when CPU rendered)

Blender 2.79 RC1. Each result is average of two renders of the same scene.

Single Xeon E5-2687w - 22m 43s (8 core 16 thread) (16x16 tiles) (second CPU in another rig)
1x RX 480 - 9m 32s (with monitors connected)
1x RX 480 - 8m 56s (no monitors connected)
2x RX 480 - 4m 30s

Scaling wise, we are close to almost 200% scaling.

I’ll test more when i get my main render farm build with 4 GPU’s

Note, this is just one scene and there could be others were results would show other scaling.