I planning to build a PC with 2 GPU for Blender and i can’t find complete explanation about hardware compatibility. If i choose budget motherboard with 2xPCI-E x16, something like these - GIGABYTE B360 HD3 (rev. 1.0) and 2 GPU: GTX1070 and GTX1080, for example. Will these configuration work for Blender, or better choose high-end motherboard with SLI support?
Depends what do you mean. I don’t think blender is going to ever benefit from SLI. Blender will however use more than one compute device for rendering. That has nothing to-do with SLI however.
Yea SLI isn’t for cuda based stuff. They say it actually renders slower when multiple cards are in SLI mode.
Sli works with identical cards on specific motherboards and it only applies in gaming (and not every game afaik).
It has no meaning whatsoever for 3d apps like Blender. It won’t improve viewport performance and furthermore, 2 gpus connected via sli bridge will render slower than two gpus without it.
So,
- if you plan to use 2 gpus in sli for faster viewports, don’t do it. It won’t work.
- If you plan to use 2 gpus for rendering, then pick the right motherboard and don’t use sli to bridge them. Just install the cards on the appropriate slot.
1 Like
Ok, all i need what would motherboard have 2 PCI-E x16 slots and budget motherboard will work fine for Blender, right?
you don’t even even necessarily need two 16x slots. Since the data is loaded to the card before rendering you won’t loose much in speed.
Hi, most budget boards reduce PCI-E performance if you use both slots.
My board go down to x8.
Make sure yours not go under x4 for example and you don´t have problems.
Even x1 is fast enough to fill your GPU RAM in seconds, ~1GB per second.
Cheers, mib
1 Like
Yes, but in gaming terms at least, no gpu currently fully saturates PCI-e 3.0 x16. PCI-e 2.0 x16(= pci-e 3.0 x8) should still suffice even for a 1080Ti(perhaps just barely though).
So as @mib2berlin mentioned just make sure the secondary slot can offer the 8 lanes and you’ll be fine. In case of intel consumer motherboards the secondary pci-e slot will go through the chip-set AFAIK and share lanes with other peripherals such as m.2 nvme drives. In AMD Rysen boards case the secondary slot also connects straight to the cpu and only the third slot goes through the chip-set. Thats talking about consumer boards.
x8 on the second PCI slot support only High-End motherboards which support SLI as well:) Blender - you’re ruining me…
Maybe, but you need neither one of those things. Your card will fill just fine with even a x1 slot. we are talking about fractures of a second here. And that only before your scene actually starts rendering. Once it’s loaded into the memory of the card the speed of the slot has absolutely no impact.
And to the SLI thing. Even when the board supports it you don’t have to use it. You actually can’t use it since you plan to use 2 different cards.
1 Like
Yes it would only be really noticeable in a real time renderer.
1 Like
Ok, i think SLI motherboard and 2 identical graphics cards would be best choice. If i need gaming i switch to SLI, if i want to work - turn off SLI.
HI. it is best to install more than one GPU, using one for display and the other(s) for rendering.
https://docs.blender.org/manual/en/dev/render/cycles/gpu_rendering.html
I have a Workstation with one 1070 on it. And I have three other gtx connected trough pci-e risers. Rendering lighting fast. 
The problem is that Blender viewport is using (checked in HWMonitor) the gtx1080 pluged trough the riser to update/shade the viewport. Like so viewport is very VERY SLOWER than if I disable the 1080 in the device manager and Blender start to use the pci-e 16x 1070.
I tried to disable all other CUDA devices exept the 1070 on user preferences but Blender is still using the 1080 unless I disable the gpu trough device manager.
I think Blender is just using the device it think is the faster available and not taking in account the riser that slow it down for the realtime.
You have more than one GPU? A you happy with it? Does make sense build multiple GPU system for Blender? What motherboard do you have?
Yes I have set it up using pci-e risers. Now I have 1080, 1070, 1060 and a 970 gtx all rendering together as CUDA devices and each GPU I add I would say render times drops proportionally.
Mobo is PC Mate b350.
I made a quick benchmark here for you with 200 samples (I would use 2000+ for the final render). The scene is an interior living room complete with lights, textures and hdr sky:
Just the GTX1070: 3m58s
1080+1070+1060+970: 1m26s
I found the solution as you can see here:
Thanks, very helpful information!