Why doesnt VRAM stack when using multiple cards?

[PlatinumSun]

You know, Valve and God have alot in common.

[Air]

Because the memory on each card has to be exactly the same, limiting VRAM to the lowest common denominator. For example, if you somehow had a 4GB card and an 8GB card in SLI/Crossfire, you'd only have 4GB of VRAM available to games.

[Bad 💀 Motha]

It works that way by design; ask AMD & NVIDIA. We're just telling you how it works.
Nothing we can do about that.

NVIDIA SLI has been dropped for future GPU Families anyways cause we're at a point now where a single GPU can handle all the work just fine without need for this multi-GPU non-sense; which doesn't work properly half the time anyways. If you need more performance, buy a single stronger GPU.

[oobymach]

I believe vram is cloned in an sli setup, I see no reason it couldn't stack, it's theoretically possible someone just has to write the code for it. I think it stacks in GTAV.

[Monk]

Multi gpu hasn't been dropped yet and likely won't.

However the reason it won't work is a driver / instruction issue, using the combined Vram was one of the key features dx12 was supposed to bring way back when the 900 series cards launched, however we still seem to be waiting for that ...
On the plus, cards are now offering more than enough memory for it not to be an issue.
One thing I'll agree with is you should always buy a single stronger card than two weaker ones these days, making sli only really an option for those who want more than a single strongest card can offer.

Originally posted by oobymach:

I believe vram is cloned in an sli setup, I see no reason it couldn't stack, it's theoretically possible someone just has to write the code for it. I think it stacks in GTAV.

It shows the two cards total but usage is doubled so you don't get the use of both cards Vram, my 980's showed as 8GB by 1080ti's show as 22GB, but only 4/11 usable.

[vadim]

Originally posted by Sword of Damocles+5:

I just wanted to know if anyone new why VRAM doesn't stack.

Because each GPU has immediate access only to its own VRAM. Exactly the same as you cannot stack RAM if you have 2 PCs in your home. Even if they are in the same local lan segment.

[MancSoulja]

It's because both GPU's need access to the same memory. If GPU 1 needs to draw a frame but the textures it needs is in GPUs 2s memory, the process of transfering that to the other GPU's memory would slow down the game.

[Carlsberg]

Originally posted by Bad_Motha:

It works that way by design; ask AMD & NVIDIA. We're just telling you how it works.
Nothing we can do about that.

NVIDIA SLI has been dropped for future GPU Families anyways cause we're at a point now where a single GPU can handle all the work just fine without need for this multi-GPU non-sense; which doesn't work properly half the time anyways. If you need more performance, buy a single stronger GPU.

Multi gpu SLI has been dropped, (3 and 4 way SLI), 2 way SLI is still current and is still supported. I have been using SLI setups for last 8 years and have had no issues in any game i have played. You should google your info before posting.

In answer to the question, SLI setups use alternate frame rendering like one does all odd frames and one does all even. Each gpu has to maintain its own frame info, textures etc and then output a complete frame, technically i guess you are still only using one card at a time but its doing half the work so output is much faster on screen. So, cards do not stack or share their memory but use it each for their own output.

[igloosfolly]

DX12 can stack ram not the full amount of each card but a portion of it. It is up the Dev's to code it in. The process is not SLI or Crossfire. I always find it interesting that people will spend money on a faster hard drive or delidding or super fast ram and then not spend the money to SLI . Hey a few more frames is a few more frames right. While a single card may be fast enough if the game supports SLI it can be faster. DX12 the sixteen year old elusive dream.

[Bad 💀 Motha]

http://thetechaltar.com/is-multi-gpu-dead

Here is a quote from that article:

Originally posted by TheTechAltar:

NVIDIA, AMD Taking a Break from Multi-GPU
We’re seeing both manufacturers slowly shift their stance away from multi-GPU, whether it’s the gradual decline in production of dual-GPU on a single card aimed at gamers (they’re still somewhat produced by AMD, but the focus is definitely shifting away from gaming and gamer oriented budgets, NVIDIA hasn’t made one since the disastrous Titan Z), or the slow reduction in supporting technologies such as NVIDIA dropping generic support for up to 4 cards down to 2 and with their new Titan V which launched late last year, dropping SLI and NVLink altogether (although admittedly this is more of a pro card than previous Titans).

Arguments abound, chief among them that the real reason these edge case pushing scenarios are disappearing is that with the advent of DirectX 12 and the API natively supporting the possibility of multi-GPU implementations without the need for the traditional manufacturer backed offerings, the traditional view of multi-GPU is no longer required. A GPU power utopia of logic which just magically matches the relative performance levels of any given combination of cards will just sort out any issues which occur and lead us all to the promised land of consistent and smooth high frame rate gaming performance…

So as you can see, SLI falls off when you go looking at GPUs like TitanV

You can most likely expect to see that in full-enforcement when the GTX 11xx series gets here.

At the very least, what we do know for sure based on previous steppings and changes; is at the least, NVIDIA SLI is no longer something aimed towards a feature that Budget-Gamers can really utilize, case-in-point "GPUs in the 10xx family below GTX 1070 do not support ANY kind of SLI, period"

Now in the future, it is very possible for an API such as DX12 to pickup the slack and move multi-GPUs forward in that regard, however CFX/SLI still would require the hardware-makers to allow this to even function in the first place. I suppose moving forward (just my guess here) I would hope that NVIDIA just gets rid of any requirements for an SLI-Bridge and allows certain GPU steppings to be allowed to take advantage of a multi-GPU config; case-in-point would be with regards to AMD and it's current and also last few Gens of GPUs, where a physical CFX-bridge was no longer required and they could potentially run under a CrossFire (CFX) config based on the Motherboard, and then supported GPUs, and then also supporting Drivers and Games. Overall I think the makers as a whole need to do something at both the GPU manufacturing as well as Driver stages of all of this, to rid certain un-useful requirements from the overall picture, to avoid hefty requirements in the first place, allowing a Motherboard + Game + API to then pick up this slack and allow for multi-GPU configs to then work as intended.

[Monk]

So your evidence is the cut price scientific professional titan Volta along with an editorial piece from a site I've never heard of and rumours ... They dropped it from lower tier cards as it's an enthusiast thing, as you've said, it makes more sense to have a single more powerful card.

I don't see sli vanishing from the high end any time soon unless it swaps to nvlink.
It's missing from titan Volta largely to keep its top end professional cards relevent seeing as it's a 3rd the price of them already, which, do support multi gpu, the titan V is the bargain priced cheap professional card after all.

[igloosfolly]

Well for 3000 buckaroos lets hope we dont need a second card. But 1070 and 1080 has the bridge and all the better motherboards support sli. Plus my first observation DX12 the sixteen year old elusive dream. Only the shadow knows or the janitor at Nvida and he aint talking. SLI has always been a enthusiast product or a way to skip a generation. DX12 has been a selling point on the side of every box since Fermi and the present gives us a handfull of games.

[TehSpoopyKitteh]

There would be no point to it since each video card's GPU is only doing part of the workload. Also, because of the way PCIe works, if you run multiple VRAM but identical GPU cards, they will go off the lowest card's vram. So one card with 8GB and one with 4GB would limit them both to 4GB.

[Bad 💀 Motha]

Originally posted by The Spoopy Kitteh:

There would be no point to it since each video card's GPU is only doing part of the workload. Also, because of the way PCIe works, if you run multiple VRAM but identical GPU cards, they will go off the lowest card's vram. So one card with 8GB and one with 4GB would limit them both to 4GB.

That is simply based around a non-VRAM-stacking factor.

But yea it would cause possibly more slow-downs if the VRAM stacked and acted as "one" cause then there might come a need for GPU #1 to draw from GPU #2's VRAM and that means delays; however you look at it.

[TehSpoopyKitteh]

Originally posted by Bad_Motha:

Originally posted by The Spoopy Kitteh:

There would be no point to it since each video card's GPU is only doing part of the workload. Also, because of the way PCIe works, if you run multiple VRAM but identical GPU cards, they will go off the lowest card's vram. So one card with 8GB and one with 4GB would limit them both to 4GB.

That is simply based around a non-VRAM-stacking factor.

But yea it would cause possibly more slow-downs if the VRAM stacked and acted as "one" cause then there might come a need for GPU #1 to draw from GPU #2's VRAM and that means delays; however you look at it.

A better way to illustrate the problem is probably to look at RAID 0 and how it works.

One thing I forgot to mention. The GPU's would all have to be connected to one slot simultaneously out of the entire PCIe bus to be able to stack RAM. 3DFX Got around that limitation with the Voodoo 2 series, if I am not mistaken, by having multiple 3D and 2D GPU's on the cards.

[igloosfolly]

Well anything is possible as we see CPU able to utilize more PCI lanes something that in the past was not there. Stacking Vram was not possible on older tech. These days there is talk about mixed brand GPU working together on the same board. We have also seen two 1060's working together without a SLI bridge in Ashes of Singularity.