Multi GPU on Motherboard

Ey, mates.
Imma buy myself a new computer, with a Ryzen 3950x CPU.

Lookin at x570 motherboards, and here’s a very basic thing I don’t understand:
Currently I only have one video card, but In the future, when there’s a substantially better GPU out, instead of taking out the old one, It makes sence to keep it and get extra blender render power out of it.
I don’t quite get it how to know whether a motherboard supports multiple graphics cards or not for blender. Some of them have multiple PCIe slots, but that doesn’t mean it will work, right? How do I know?
For example:


Two potential slots for a GPU, but as far as I understand, supports a GPU only on one of them, right? But that’s for gaming only, maybe - No SLI support. Is SLI required for Blender?
What would happen if I placed a second GPU to the lower PCI slot of this motherboard? Would it simply not be recognized as a GPU? Would it be unstable? Slower? How much?

Thanks :slight_smile:

hmm…
Watched a video.

Apparently, The value I’m looking for is “Lane speed”.
The top one supports 8x, the bottom supports 4x. (Not sure why they both have 16x on their label… That might just mean the form factor…)
As I understood, a GPU requires 8x.
But… Using the 3950x unlock PCIE 4.0, which doubles the lanes… So I could potentially use two GPU’s on this motherboard, because of my CPU…

Still, i have a feeling the future Nvidia Ampere cards would only work on the top slot, because they’re rumored to use PCIE 4.0. … so will require all 16x?..
Do i understand this correctly?

If you want multi GPU motherboard you need to check couple of things. Two most important are PCIe implementation - you want as much x16 slots as possible. Second thing is power delivery - this is more general thing as component quality vary little between manufacturers. It’s considered a good practice to check out before purchase thou. Stable power is more important for CPU than GPU as the VRMs for CPU are located on a motherboard itself. VRMs for the GPU are on the GPU board, and the GPU take power both from MB ad directly from PSU. In recent motherboards for multiple GPUs you can find additional sockets on the motherboard for powering more than 2 cards.

According to official specification: https://www.asus.com/us/Motherboards/TUF-GAMING-X570-PLUS-WI-FI/ your motherboard have 2 PCIe x16 slots that means you can safely put up to 2 GPUs.

As far as I know SLI is not required for Blender. I’m currently rendering on 2 AMD cards.

The last thing to remember is to check your PSU wattage - compare it with combine draw from both GPUs + CPU under full load, and add a safety buffer for the rest of the components. Be sure that your PSU can safely handle the draw. You DONT WANT to cheap out on PSU while playing with multi-GPUs.

Current card does not saturate PCIE 3.0, 2080 super is close (2080 super 15.5GB/s, PCIE 3/16x does 15.75 GB/s ). Nvidia Ampere should though.

The difference is actually negligible for gaming.
https://www.gamersnexus.net/guides/2488-pci-e-3-x8-vs-x16-performance-impact-on-gpus

Found a test with 2060 supers:

PCIE doesn’t matter much for rendering either. The scene will mostly be loaded in the GPU at the start then rendered with minimal PCIE use.

Now, for a 3950x, I’m not even sure there is a motherboard without PCIE 4.0 available. If you want to go expensive, you could even get a PCIE splitter from one 16 to 4x4. But as silex said, don’t skip on the PSU and the cooling.

SLI is useless for blender if I remember.

Each of those slots is a x16 slot in terms of the pin count. Hence the label.

The second slot is fed through the chipset hence x4 speed. For offline rendering that is fine, btw. But since it is PCIe 4.0 x4 it is actually equivalent speed to PCIe 3.0 x8. So it’s even less of a problem.

On x370 and x470 boards the second slot was always PCIe 3.0 x8 anyways, and actually since both came from the same cpu lanes, in case of 2 GPUs they were both running at x8 speed. With this board using the x4 lanes from the chipset even in dual gpu config the first slot remains at full x16!

edit: note that I’m commenting on the specific TUF board you were looking at.

edit2: I actually think the spec sheet might be just confusing. the chipset x4 likely goes for the nvme m.2 slot. So similar to the previous chipsets the x16 slots likely share the same connection from the cpu. in that case dual gpu likely does similarly split the x16. But it’s fine, since it’s PCIe 4.0 so more than fast enough.

1 Like

In today, nearly all motherboards support multi GPU. But powerful GPUs takes up a lot of space, and easily one GPU takes up another PCI-Express slot, and you can not use second GPU. For this reason, check two PCI-Express slot space is enough, or motherboard have 3 or more PCI-Express slot.

Yes, to be on the safe side make sure to get a GPU that isn’t too thiccccc.

I would like to add that if you wanted to use multi GPU, blower style cooler is usually better for cooling instead of the usual axial style cooler, particularly if the case airflow is not good or the spacing between cards is tight.

If you are still using 1 GPU then axial style cooler is better, but when you add another GPU maybe it’s worth considering the use of blower style cooler for GPU at the top PCIe slot.

Thanks, guys.

I’ve noticed, that on many cards, the small PCIe slots can easily get obscured and made useless if a GPU is inserted, which is a bummer…
I may want to get extra expansion cards, like an HDMI capture card and firewire input… And I do have a RTX 2080 super, which is thicccc.

My Current choice is ASUS ROG STRIX X570-E GAMING ATX MB

I couldn’t find anything wrong with this card, that matters to me. Good vrm, wifi6, not obscured PCIe small slots… cool stuff. 315 EUR. fair enough.
If the future ampere cards will require faster than 16gb/s, I guess it’s either one Ampere card, or two slightly bottlenecked cards, for which I will not see a difference, right? Still, this ain’t happening any time soon. Should be good for a while.

PCIe 4.0 x16 has the bandwidth of 31.5GB/s, even if the card ends up using x8 at hanlf that and is able to saturate a little more than 16gb/s the bottleneck shouldn’t be too high.

As I understood,
The top slot is 16x lane on Pcie 3, and turns 32x with PCIe 4.
But when two cards are connected, they both become 8x on PCie 3. Both are 16x with PCie 4.
Or maybe that’s just valid on some other motherboard I watched a video review of… xD

PCIe 4 is and remains PCIe 4, even when split up between two slots.

The X1 X4 X8 and X16 are the PCIe lanes. as in how many lanes the socket is connected to.

Both the first and the second x16 slot are connected to the same group of lanes coming from the CPU, if you populate both the slots then the x16 lanes get split into two x8 lanes but they are still PCIe gen4 with gen 4 speeds.

PCIe gen 4 x16 lanes is just twice as fast as gen3 is what it means. It’s actually still just 16 lanes.