Vega's SLI Scaling Thread

V

Vega

Supreme [H]ardness
Joined
Oct 12, 2004
Messages
7,147
The test settings:

SLITestSettings.jpg



Resolution 4800x2560 (3x 30" LCD Portrait). Portrait bezel correction is not working properly in nVidia driver release 270.51 so the standard resolution is used. 2x 3GB GTX580's as baseline as two cards minimum for Surround. The computer: http://hardforum.com/showthread.php?t=1588587

The numbers:

SLIScaling.jpg



F1 2010 in-game Benchmark: Quad SLI 75% over dual cards.

Video of SLI: http://www.youtube.com/watch?v=Nqdc66a0l3U
Video of Tri-SLI: http://www.youtube.com/watch?v=5h6nil795M8
Video of Quad-SLI: http://www.youtube.com/watch?v=KWZqS1wsGTc

Lost Planet Benchmark B: Quad SLI 77% over dual cards.

DCS:A-10C "Spring Hill" simulation replay: Quad SLI 88% over dual cards.

AvP Benchmark automatic settings: Quad SLI 93% over dual cards.

Crysis 2 "Single player static entry": Quad SLI 59% over dual cards.

Batman in-game Benchmark: Quad SLI 76% over dual cards.

Metro-2033 "Single player static entry": Quad SLI 100% over dual cards.


"Single player static entry" means the single player campaign was loaded and character was sat sitting in the game world unmoved from original position during all tests. The FPS was then annotated with FRAPS.

All benchmarks were tested with the exact same settings with a reboot in-between to reconfigure SLI bridge.

Without VRAM restrictions (3GB GPUs) and full 16x PCI-E bandwidth available to all cards, there was not a single instance that Quad-SLI did not garner an appreciable performance advantage over Tri-SLI. In a number of instances, the performance gain transitioning from Tri-SLI to Quad-SLI was actually greater than the transition from SLI to Tri-SLI.

All games saw ultimately a performance increase going from SLI to Tri-SLI, and then also from Tri-SLI to Quad-SLI. Some games had just incredible SLI scaling such as Metro2033, Alien vs Predator, and DCS:A-10C. 90-100% performance gains over a dual card setup is pretty much unheard of. I attribute these large gains to no VRAM restriction and the extreme demand of this resolution.

The phrase "Quad-SLI does not scale in games" is ancient history.

Thanks for viewing!
 
Good work. But it's clear that the 4th card is useless. But at least you did some good tests. :)

Now go out and buy 4X 6970 and then do the same tests. I think you will be surprised with the latest 11.4 drivers. :)

It would be interesting.
 
LEVESQUE said:
Good work. But it's clear that the 4th card is useless. But at least you did some good tests. :)

Now go out and buy 4X 6970 and then do the same tests. I think you will be surprised with the latest 11.4 drivers. :)

It would be interesting.
Click to expand...

I would hate to say it, but they wouldnt be as fast as the 580 3gb in quad-sli.

580's only drawback is the memory limitations, which Vega fixed by getting the 3gb models.

Overall nice results Vega, those Metro 2033 numbers are amazing.
 
i think 4.8ghz on the cpu is way too little for quad sli. bump it to 5.5ghz or more and you'll see a trumendous difference. you should cherry pick a cpu that can do it with acceptable voltage
 
I saw 29fps in SLI and was like "what the hell!? then I remembered what resolution this was at :D

Do quake!!! (cant remember if that even works on SLI, it did with voodoos...or did it...:confused:)
 
I was about to go 6970+6990 then i seen the 3gb back in stock and after seeing vega benchies on here i'm glad i i pulled the trigger on 2 gtx 580 3gb Palit instead of ati.Will have mine tommorrow going buy metro 2033 keep hearing great things about this game and i'm seeing it's now playable at high res.
 
ryu4000 said:
I was about to go 6970+6990 then i seen the 3gb back in stock and after seeing vega benchies on here i'm glad i i pulled the trigger on 2 gtx 580 3gb Palit instead of ati.Will have mine tommorrow going buy metro 2033 keep hearing great things about this game and i'm seeing it's now playable at high res.
Click to expand...

So you spent more money to have less performance?

Makes sense to me!
 
Extremely impressive results Vega, thanks for taking the time to test and post this :cool:.
 
Nice to see the review - man I wish NV didn't screw the pooch by shipping the 1.5GB 580s.
 
DASHlT said:
So you spent more money to have less performance?

Makes sense to me!
Click to expand...

no less noise and plus ati doesn't play to well with most games look at rift for instace 30fps when nvidia cards hitting 80+ and also i rather not hear jets taking off and heating issues.
 
JayteeBates said:
Nice to see the review - man I wish NV didn't screw the pooch by shipping the 1.5GB 580s.
Click to expand...

Well to be fair, unless you're running an eyenfinity setup, it doens't really matter. 1.5GB is more than enough memory for a single 2560x1600 display. It's even enough for 3x1600x900 that a some people run eyenfinity at. (2700x1600 in portrait mode, so almost identical to 2560x1600).
 
interesting results Vega. very impressive scaling in A-10C, AVP, and Metro 2033.
 
drooooooooooooooooooool

Thanks god I can't stand multi-monitor gaming or I'd probably shell out the cash too :D
 
ryu4000 said:
no less noise and plus ati doesn't play to well with most games look at rift for instace 30fps when nvidia cards hitting 80+ and also i rather not hear jets taking off and heating issues.
Click to expand...

Not sure about Rift.

With my setup below with max details I hover around 45-60fps?

Maybe you need to check thoise results again.
 
So essentially, you would have to double the number of cards to get 2x the performance.

To get twice the performance of 2x, you need 4x.
To get twice the performance of 4x, you need 8x (theoretically).

Also, you should have done a benchmark with 1 card just for reference.
 
DASHlT said:
I would hate to say it, but they wouldnt be as fast as the 580 3gb in quad-sli.

580's only drawback is the memory limitations, which Vega fixed by getting the 3gb models.

Overall nice results Vega, those Metro 2033 numbers are amazing.
Click to expand...

zerounleashednl said:
I would love to see the vram load numbers (gpu-z or msi afterburner) that goes with these results! :cool:
Click to expand...

You can see some VRAM usage numbers in the F1 2010 videos. Virtually all of these games use more than 2GB VRAM. The 1.5GB GTX580 would slow down tremendously and the same would happen to the 69xx series as it too runs out of VRAM.
 
DASHlT said:
So you spent more money to have less performance?

Makes sense to me!
Click to expand...



my man, hd 6970 is way behind gtx 580. trust me i've tested both of them. hence the price difference. if you are looking for the ultimate performance gtx 580 is the way to go
 
LEVESQUE said:
Good work. But it's clear that the 4th card is useless. But at least you did some good tests. :)

Now go out and buy 4X 6970 and then do the same tests. I think you will be surprised with the latest 11.4 drivers. :)

It would be interesting.
Click to expand...
What? All of his results show increased scaling from 3x to 4x cards. How is the fourth card useless?:confused:
 
Thanks for the great benchmarks. It's refreshing to see someone really expose the potential of the 580 without it's VRAM limitations. If only NVIDIA had done it right from the beginning, at the $500 price point, then a lot more people would have jumped on board.

Right now the palit is going for $628 on amazon, which is high, albeit worth it and perhaps non-issue for a lot of people who plunked down $500/$1000/$1500 for the 1.5gb setups. Would have been nice to see a 3GB reference card from the getgo...still would be nice to see one now from Nvidia.

Now, if we can finally get a test setup for benchies using a similar setup to yours (little/no CPU bottleneck and resolution appropriate for the mammoth setup) using triple SLI of the 1.5gb versions against its 3gb counterpart and 6990+6970 and 6970x3's. that would be a sweeeeeet review.
 
You can see some VRAM usage numbers in the F1 2010 videos. Virtually all of these games use more than 2GB VRAM. The 1.5GB GTX580 would slow down tremendously and the same would happen to the 69xx series as it too runs out of VRAM.
Click to expand...

Ýeah I see them now... great benchmarks! :D
 
LEVESQUE said:
Good work. But it's clear that the 4th card is useless. But at least you did some good tests. :)

Now go out and buy 4X 6970 and then do the same tests. I think you will be surprised with the latest 11.4 drivers. :)

It would be interesting.
Click to expand...

3x to 4x Performance Increase based on Vega's numbers:

F1 2010 21.7%
Lost Planet 30.8%
A-10C 24.7%
AvP 40%
Crysis 2 17%
Batman 28.9%
Metro 2033 36.4%

Average Increase 3x to 4x 28.5%, almost linear scaling.

You couldn't be more wrong if you tried, and you try REALLY hard.;)

Impressive numbers! You're getting close to 100% scaling on average.
 
Were you able to max out the settings at these resolutions as well? (e.g. max antialiasing, all options set to highest, etc?)

This is an incredibly impressive rig. I mean yeah it's expensive, but you obviously put a lot of thought and care into it to get the best results possible.

Would you ever try putting in 4x 6990's?
 
Cerulean said:
So essentially, you would have to double the number of cards to get 2x the performance.

To get twice the performance of 2x, you need 4x.
Click to expand...

The scaling is what we're examining and it looks impressively linear for that setup. Most SLI testing I've seen the benefits of adding the 3rd and 4th gpu are quite terrible.

Also, you should have done a benchmark with 1 card just for reference.
Click to expand...

Not possible for that setup. NV Surround requires at least TWO cards in SLI mode to work.
 
avatar[djedi];1037106297 said:
Were you able to max out the settings at these resolutions as well? (e.g. max antialiasing, all options set to highest, etc?)

This is an incredibly impressive rig. I mean yeah it's expensive, but you obviously put a lot of thought and care into it to get the best results possible.

Would you ever try putting in 4x 6990's?
Click to expand...

You cant do 4x 6990's, that would be 8 gpu and 4 is max. He did have 4x 6970 and it did well against the 1.5GB 580's (VRAM limit) but the 3GB's put the hurt on the 6970's. Some games were limited by the 2GB VRAM and he even found limitations of the CFX bridges that caused other issues. There were the displayport issues he had.
 
LEVESQUE said:
Good work. But it's clear that the 4th card is useless. But at least you did some good tests. :)

Now go out and buy 4X 6970 and then do the same tests. I think you will be surprised with the latest 11.4 drivers. :)

It would be interesting.
Click to expand...

Yeah he did the whole 6970 thing. I doubt new drivers will fix the limitations of 2GB VRAM, or the CFX bridge bandwidth issues, or the DP issues.
 
"Single player static entry" means the single player campaign was loaded and character was sat sitting in the game world unmoved from original position during all tests. The FPS was then annotated with FRAPS.
Click to expand...

Do you mean no new information was loaded for this whole test? for exampel, you loaded the level up and you only looked at one place the whole time?

Don't really get your wording on it
 
Lost Planet
A-10C
Batman
Metro2033.

In all those games, Tri-SLI is ALREADY over 60 fps. So the 4th card is useless in those 4 games. E-peen only. Who cares about 200 fps? 60 is enough to play a game smoothly.:)

F1 2010. 10 fps from Tri-to Quad, but, both still under 60 fps. Meh.

AVP. 6 fps from Tri to Quad, again, both under 60 fps. Re-meh.

Crysis 2. 9 fps from 53 to 62. Ok. Maybe better. Maybe. But 9 fps?

So 3 games only were UNDER 60 fps with Tri-SLI, with 2 still under 60 fps with Quad.

I don't know. But 600$ for the card +150$ for the waterblock, plus 2 months waiting for the blocks, plus 2 months working on a crazy climatisation system, plus applying 4 inches of sealant-goo for hours over every parts of his set-up.

Was it really worth it for 3 games? and 2 of those are still under 60 fps with Quad-SLI? All the hours lost, all those troubles for 3 games?

4th card is useless., and scaling badly.
 
vengence said:
Well to be fair, unless you're running an eyenfinity setup, it doens't really matter. 1.5GB is more than enough memory for a single 2560x1600 display. It's even enough for 3x1600x900 that a some people run eyenfinity at. (2700x1600 in portrait mode, so almost identical to 2560x1600).
Click to expand...

This is totally true however the real point of 580's and 6970's is for multi mon goodness. There really is little difference between the high end cards and one step down at what I will refer to as "mortal" resolutions (I.E. a single screen 24" or less). So NV did screw the pooch on multi-mons with that 1.5GB.
 
JayteeBates said:
This is totally true however the real point of 580's and 6970's is for multi mon goodness. There really is little difference between the high end cards and one step down at what I will refer to as "mortal" resolutions (I.E. a single screen 24" or less). So NV did screw the pooch on multi-mons with that 1.5GB.
Click to expand...

But Nvidia has 3D pron. :) + You can always get a Palit 2GB Nvidia card, but I don't think they make a 580GTX model just a 560TI :rolleyes:

Yeap Nvidia scewed the pooch for the most part.
 
I will not knock this build, by all measures it is "Elite"

But the problem for us mortals is that the price/performance ratio for any of the 580GTX multi-monitors setups is to far out of wack to truely consider. The only time the price/performance scale falls in Nvidia's favor is in 3x30" Quad Card setups because at that point ATI has nothing that comes close to the performance of four 3GB 580GTX's. For ANY other price/performance situation, ATI's either come out the clear winner or the lines blur to the point that price creates the true winner.

Here's my point.

Taking the current Vid Card Prices: $630 for a 3GB 580GTX, $700 for a 6990, and $325 for a 6970.

$1250 will net you two 580GTX 3GB's, or one 6990 and one 6970 + $225. ATI clear performance win.
$1900 will net you three 580GTX 3 GB's or Two 6990's and $500. Toss up in performance but $500 in your pocket with ATI, giving the final nod to ATI.
But at $2500 you will have 4 580GTX 3GB's and there is nothing in the ATI arsenal to come close so winner has to be Nvidia by default.
 
I don't think that anyone has ever considered this a cost effective rig, that's not the point and that's way people are so interested in this build, what can you do when money and will power aren't limiting factors. The money is obviously much but that cooling solution is something I simply would never do as that's a lot of work.

The other thing that AMD doesn't really have is S3D. While cost/performance is the main thing that people are interested in obviously it's not the only consideration.
 
Last edited:
Putting LEVESQUE on ignore significantly improves your online experience and cleans up Vega's topic a bit. I highly recommend it.
 
Not a bad idea but he's spreading disinformation that needs to be challenged so as not to confuse people. He's just flat out lying/wrong about the scaling and that needed to be addressed. At best both CF and SLI scale LINEARLY and the benchmarks that Vega presented on average are giving an average of 28.5% better performance than 3x SLI out of the theoretical maximum of 33%, that's scaling of 85% efficiency in these benchmarks which nobody in their right mind would say is anything but great scaling. Just could let him flat out lie about something this important and basic.
 
Last edited:
That 4th GPU is kicking ass, and I love this near/sub zero rig. Gnarly mod man, come cool my house with this thing, Phoenix gets HOT!
 
You must log in or register to reply here.
Back
Top