A Question I Had No Answer For.... SLI...

[Kunaak]

I build and sell custom PC's to people in my free time.
today I got a question that I just had no idea how to answer.

"if I get 2 6800 ultras, each has 256 megs right, so does that mean I'd have 512 megs of videocard ram?"

I had never thought of that...
so I have no idea how to answer it.
but it did bring up a interesting question to me.
if it does mean you get 512 megs of videocard ram...
will it be just as good as a single card with 512 ram (when it happens)
will games be able to use the 512, like doom 3?

how would it affect benchmarks?

kinda embarassing to be asked something, and draw a complete blank....

[r3b0rN]

always happens to me sure some1 as respected as u would be a little peeved tho

thats interesting tho, id like to know... i bet it depends on the application?

[LardArse]

If it is using the good old 3dfx SLI technology it should only be 256 meg, just like the Voodoo 2s and 5s.

[Minnyboy]

Originally posted by Kunaak
I build and sell custom PC's to people in my free time.
today I got a question that I just had no idea how to answer.

"if I get 2 6800 ultras, each has 256 megs right, so does that mean I'd have 512 megs of videocard ram?"

I had never thought of that...
so I have no idea how to answer it.
but it did bring up a interesting question to me.
if it does mean you get 512 megs of videocard ram...
will it be just as good as a single card with 512 ram (when it happens)

I ain't 100% sure but I don't think it'll be as good as a 512MB card

will games be able to use the 512, like doom 3?

I think Doom 3 would be able to see it as a single 512MB Card

how would it affect benchmarks?

I think it'll be very good for benchies unless idiots start using that Null PS

kinda embarassing to be asked something, and draw a complete blank....

[WeaponX]

I would have to say yes. Graphics cards are moving towards that magical number of 512mb, it would only make sense that SLI (a next generation card) would utilize all of it's available memory. Just my 2 cents.

[Dissolved]

depeneds how there ran..

Alienware method devides the screen up. so one card does top half of the screen and the other does the bottom..

the nv way im not to sure of cuz frankly i havent cared since still in testing and wont be out for months..


My best answer would be no, since the power would get slipt between the 2 cards and ran in parelle?

4am, i guess i need sleep.

[zakelwe]

It will act as if a 512MB card as far as I can tell. If the frame buffer requirements are X then both cards have to do be able to hold 0.5 of X as they draw half the screen. So each 256MB card holds half of a maximum of 512MB of textures before texture grabs have to be done over the PCIe bus. Same as a 512MB card drawing the whole screen.

Regards

Andy

[slavik]

actually, I think that both cards will hold a copy of the texture ... so it's more like a RAID1 HDD array ...

[zakelwe]

Originally posted by slavik
actually, I think that both cards will hold a copy of the texture ... so it's more like a RAID1 HDD array ...

Well if one is drawing the sky and one is drawing the ground then it makes sense to have one having 256MB of sky textures and one having 256MB of ground textures. If they both have 256MB of everything then this is very inefficient, so my assumption is that they have a 256MB framebuffer of what they are working on.

I'm not 100% on all this though it has to be said, good question posed.

Regards

Andy

[zakelwe]

Well I have found out what the answer is. Through some really nitty gritty journalism worthy of a Pulitzer prize I sent an email to Dave Baumann at Beyond3d.com ( who actually did all the hard work ) and sent back this

http://developer.nvidia.com/object/sli_faq.html#10

within 20 minutes of me emailing him .. so thanks Dave

"Is the frame-buffer memory shared, i.e., can I access a 512MB frame-buffer if I have two GPUs with 256MB each?

No, each GPU maintains its own frame-buffer. Rendering data, such as texture and geometry information, is duplicated. "


Damn, I was wrong again ...

Regards

Andy

[i_jester]

The SLI that will be working with the nVidia chipsets is exactly what Zakelwe said. Each card will draw 50% of the screen so two cards will act as a double fast, double memory, card. Of course rarely do the two halves of the screen carry the same amount and type of information, so performance will not be twice as fast, but should be in the 160-180% as fast range.

Also, it is designed for the PCI-Express bus (lga775) chipsets.

[i_jester]

nVidia SLI Article

[TM0D]

The cards do a 50/50 split of the workload, but not necessarily the screen space.

[Noktar]

I don't know exactly.
But NOW SLI is the only way to get 512mb of Video Ram, i think in Benches you don't get much more points, but in Games like FarCry an Doom3 you can play at higher res. (D3 -> Ultra High), I think SLI (two Cards) is a bit like HT, each Card do half of the work so you get more fps !

[Kunaak]

zakelwe

I need a dumbed down version of your answer, it's too early in the morning for me to think

[craig588]

All textures and geometry information about the scene is held in both cards memory. It is not split up. With 2 256M cards you will still only have 256M of video memory avalible. I would imagine that with a 512M card a 256M card you could only access 256M of memory too because each card needs to have the same data in its memory and the 256M one would fill up and the drivers would prevent more data from entering either one.

[thephenom]

With the huge memory bandwidth between vid card RAM and core, I doubt 512MB will ever be used up to the max in the near future.

And we should all take into consideration that 2 X 6800 Ultra doesn't mean it's 2 times faster than 1 6800 Ultra due to some inefficiency.

[i_jester]

After reading through the nVidia SLI developers manual, more of this becomes clear.

The textures and geometry are sent to both cards. The frame buffer will not be shared so the 256megs of memory will still be what applications see from the outside. However, all processing will be dynamically shared between the two (or more - the specs call for up to 8 eventually) so the brunt of the realtime 'work' will be doubled in theory.

According to nVidia, the small amount of overhead to run the SLI and send the duplicate textures and geometry result in a 10 - 15% hit - meaning the real world results should be that each card will add somewhere in the neighborhood of 85% more processing power.

That is my understanding.

[craig588]

I wonder whats stopping them from making a chain of like 8 of them and using 5 power supplies? Does it become a nightmare to program drivers to dynamicly allocate the propor resources from each card to the correct areas of the scene?


(Completely off topic regarding your sig: I love the red green show.)

[-=TriX=-]

I think they should make it so it allows you to use all the memory on both video cards. Or will we buy one graphics card with no RAM to run sli? =\

And Kunaak, I hope you just didn't stand there dumb-founded; what line did you use to cover your ignorance of this little detail of sli?

[craig588]

Thats not possible. It needs the ram to hold the data, its not doing it to piss you off. It wouldn't be able to proporly render the scene without it, its how videocards are designed.

[drunkenmaster]

though doom 3 info said 512 mb card to run at ultra settings, aren't loads of people running ultra settings fine? albeit probably wihtout max aa and af res.

not payed too much attention to doom 3 after i'd got half way through --------bored much?

Its useful, its simply too expensive, we still don't know what the overhead will be from nvidia and manu's for the sli capable versions, i'd assume more, we hope we'll see gt's and the like but i've not read we will yet. Also the boards are likely to be fairly expensive, but not as much more as first thought, originally people assumed needing a 2nd 16x slot to run them, which takes up space on chipset which means everything must be manufactured especially. By the looks of things it will be 1x16x pci-e slot split externally to the chipset which will mean no extra costs making extra chipsets and the boards should be able to run chipsets same as other boards, same pinout = same pcb design in that area, less redesigning.


I really hope dual card rigs have to enter a separate 3dmark section as this is a step to far from most peoples affordability.

[slavik]

what if u have a slope so that there is ground and sky in both halves of the screen???

Originally posted by zakelwe
Well if one is drawing the sky and one is drawing the ground then it makes sense to have one having 256MB of sky textures and one having 256MB of ground textures. If they both have 256MB of everything then this is very inefficient, so my assumption is that they have a 256MB framebuffer of what they are working on.

I'm not 100% on all this though it has to be said, good question posed.

Regards

Andy

[saaya]

yes and no.

its not like the old sli, afaik each card still uses its own memory wich would mean 256mb, but since one card will only render the bottom and the other will render the top it would mean 512mb effectively, 256mb upper textures, 256mb lower textures.

since the boarder where the image gets split keeps moving all the time up and down and the cards have to load complete textures even if only one edge of it is drawon on their side of the image id imagine it will be like having something in between 256mb and 512mb video memory.

im not sure, nvidia COULD have reduced the textures each cards loads, so each card only loads the textures that will show on the part of the screen output they are responsible for, BUT this would need more tweaking and make sli more complicated so maybe they just ignored this...

but it will defentiely not be like you had 512mb of video memory!

[zakelwe]

Just to reiterate, both cards contain all the screen in the framebuffer, not just top or bottom half, so it is exactly like one 256MB card.

Regards

Andy