SLI/Crossfire @ 2560*1440 and 3240*1920 (XBit Labs)

[3DVagabond]

The review is for 2x690's, but they compare the performance to 1x690 and 2x7970 (as well as single 680 and 7970 to show scaling).

A nice change is no lowres testing. 2560*1440 and 3240*1920 only.

XBit Labs

 

[RussianSensation]

Oh snap! HD7970 GE CF demolished the 690. The fair comparison should have been 2x 680s in 4GB SLI against HD7970 GE CF I feel, but that costs $1100 vs. $900 for HD7970 GE CF:

Multi-GPU
2560x1600 AA
GTX690 = 73.9
HD7970 GE CF = 87.8 (+19%)

3 Monitor 1920x1080 AA
GTX690 = 43.5
HD7970 GE CF = 56.5 (+30% faster)

Single-GPU
2560x1440 AA
GTX680 1137 = 44.3
HD7970 GE = 47.9 (+8%)

3 Monitor 1920x1080 noAA
GTX680 1137 = 42.8
HD7970 GE = 50 (+17%)

GTX690 Quad-SLI scaling is terrible with 3 monitors. 2 GPUs continues to be the sweet spot.

 

[BallaTheFeared]

Where multi-gpu is concerned, fps isn't the most important factor.

Looking at the scaling I wonder if the nvidia cards were operating in PCIe 2.0 or 3.0 mode.

http://www.youtube.com/watch?v=S0-xcxAvu54&feature=plcp

Edit: Video cuts, too long anyways...

Graph instead.

 

[cmdrdredd]

Where multi-gpu is concerned, fps isn't the most important factor.

Looking at the scaling I wonder if the nvidia cards were operating in PCIe 2.0 or 3.0 mode.

http://www.youtube.com/watch?v=S0-xcxAvu54&feature=plcp

It isn't? Then why bother doing SLI at all?

 

[3DVagabond]

Oh snap! HD7970 GE CF demolished the 690. The fair comparison should have been 2x 680s in SLI against HD7970 GE CF I feel:

2560x1600 AA
GTX690 = 73.9
HD7970 GE CF = 87.8

3 Monitor 1920x1080 AA
GTX690 = 43.5
HD7970 GE CF = 56.5 (+30% faster)

But a lot of that has to do with 2 games (Dirt Showdown/Sniper Elite V2). NV's performance is dreadful in those 2.

GTX690 Quad-SLI scaling is terrible with 3 monitors. 2 GPUs continues to be the sweet spot.

I was more noticing how "broken" crossfire apparently is. I hadn't noticed how much better nVidia support is on those 2 new titles. /sarc

Seriously though, quad scaling is typically not good. Whether it's because of CPU or bandwidth bottlenecked, or just because it's so niche they don't waste a lot of resources optimizing for it.

Good to see, for the most part, 2x gpu's are enough for Eyefinity/Surround.

How much different are 2x680's compared to the 690? I seem to recall when the 690 came out that there wasn't an advantage to 2x680. It's been a while though and my memory isn't what it used to be.

 

[cmdrdredd]

I believe the 690 has more limited TDP and a 680 can have higher limits. This gives higher boost clocks.

 

[RussianSensation]

Looking at the scaling I wonder if the nvidia cards were operating in PCIe 2.0 or 3.0 mode.

They are using 6-core 3930 @ 4.6ghz on X79 platform. I doubt that there is a PCIe 2.0 x8 vs. 2.0 x16 vs. PCIe 3.0 x16 bottleneck.

The Intel board they are using is rated to support PCIe 3.0 x16 for 2 slots I believe:

"For expansion cards there are three PCI Express 3.0/2.0 x16 slots with the third slot electrically designed to support PCI Express 3.0/2.0 x8 speed, two PCI Express 2.0 x1 slots and one regular PCI slot." ~ Intel DX79SI

Here are their GPU-Z screenshots for

2-Way SLI

4-Way SLI

I am going to guess that 2GB of VRAM on the 690 isn't enough for triple-monitor gaming.

For example, from the review:

"The test from Sniper Elite V2 seems to require a lot of graphics memory and high memory bandwidth. Particularly, it needs over 2.8 gigabytes of memory at 3240x1920 with enabled antialiasing, which explains the failure of the Nvidia-based solutions."

 

[BallaTheFeared]

It isn't? Then why bother doing SLI at all?

Having more than one gpu can cause a phenomenon called micro stutter, where actual reported fps don't reflect what the end user is actually experiencing.

Some cards/drivers/vendors can have more of it than others, which could lead to reportedly higher fps feeling as if they're less than what is actually be reported by programs like fraps.


Say I had two cards that reported an avg fps of 70, but they had a high rate of stutter so on my end it felt like 40.

Another set of two cards reports having only 60 fps, but does not suffer from stuttering so it feels like 60 fps.

In this scenario fps is clearly not the most important factor to overall experience and is something that has been address time and time again on this forum.

 

[BallaTheFeared]

They are using 6-core 3930 @ 4.6ghz on X79 platform. I doubt that there is a PCIe 2.0 x8 vs. 2.0 x16 vs. PCIe 3.0 x16 bottleneck.

I am going to guess that 2GB of VRAM on the 690 isn't enough for triple-monitor gaming.

For example, from the review:

"The test from Sniper Elite V2 seems to require a lot of graphics memory and high memory bandwidth. Particularly, it needs over 2.8 gigabytes of memory at 3240x1920 with enabled antialiasing, which explains the failure of the Nvidia-based solutions."

3240x1080 is smaller than 3600x1080, yet their results match up pretty close with quad 680 performance in BF3 using a 2.0 bus.

Those same cards jump from 97 fps to 140 fps by switching the bus to 3.0 in the registry, on a x79 platform no less.


Seems reasonable to assume it's possible the review was done with quad sli operating over 2.0 bandwidth.


Edit: Not that 2.0 vs 3.0 would affect a clear and present vram limitation such as in your example, however in the other tests where vram was not thunderstruck, it could explain some of the scaling issues present.

 

[SolMiester]

Oh snap! HD7970 GE CF demolished the 690. The fair comparison should have been 2x 680s in 4GB SLI against HD7970 GE CF I feel, but that costs $1100 vs. $900 for HD7970 GE CF:

Multi-GPU
2560x1600 AA
GTX690 = 73.9
HD7970 GE CF = 87.8 (+19%)

3 Monitor 1920x1080 AA
GTX690 = 43.5
HD7970 GE CF = 56.5 (+30% faster)

Single-GPU
2560x1440 AA
GTX680 1137 = 44.3
HD7970 GE = 47.9 (+8%)

3 Monitor 1920x1080 noAA
GTX680 1137 = 42.8
HD7970 GE = 50 (+17%)

GTX690 Quad-SLI scaling is terrible with 3 monitors. 2 GPUs continues to be the sweet spot.

Ah, so they finally have new drivers for CF that work across multiple screens...

 

[RussianSensation]

Ah, so they finally have new drivers for CF that work across multiple screens...

30% faster than GTX690 for $900 is better than just working OK. :thumbsup:

The same can't be said for Quad-SLI GTX690 scaling. But most people probably wouldn't be that stupid to spend $2k on 690s for games and their 2GB VRAM bottleneck.

 

[moonbogg]

690SLI is pure retardation. 2gb Vram is ludicrous for all that GPU power + quad scaling sucks anyway. A single 690 is a 1080p card. 7970 is better for higher res, otherwise 600 series 4gb cards. The Moon has spoken.

 

[boxleitnerb]

Would have been nice if they had included 3-4 GTX680 4GB.
While this test shows that AMD has taken the lead in terms of fps, this is only half of the truth. Without a comparison (measured with a high speed camera or subjective impressions) of playability/smoothness, the usefulness of this review is limited. AFR-fps != non-AFR-fps

Also note:
Even a 3860X@4.6GHz doesn't always necessarily exclude a CPU bottleneck. Especially with 4 GPUs. Power consumption measurement would also have been nice. The 7970GE CF uses close to double the amount of power of the 690.

For that res I would choose 2x4GB GTX680s any day over the 690 and overclock them as well as the 7970GE is overclocked.

 

[3DVagabond]

Would have been nice if they had included 3-4 GTX680 4GB.
While this test shows that AMD has taken the lead in terms of fps, this is only half of the truth. Without a comparison (measured with a high speed camera or subjective impressions) of playability/smoothness, the usefulness of this review is limited. AFR-fps != non-AFR-fps

Also note:
Even a 3860X@4.6GHz doesn't always necessarily exclude a CPU bottleneck. Especially with 4 GPUs. Power consumption measurement would also have been nice. The 7970GE CF uses close to double the amount of power of the 690.

For that res I would choose 2x4GB GTX680s any day over the 690 and overclock them as well as the 7970GE is overclocked.

The actual subject of the review was gtx-690's in SLI. It wasn't intended to be an all encompassing SLI vs. Crossfire review.

I found it interesting that the badly broken crossfire that everyone is being warned to avoid doesn't seem to be too broken after all. I find it just as intriguing that when the objective measurements don't follow our beliefs we refuse to accept it without further proof. I hate to say it, but the last thing we could trust is anyone's subjective assessment. Until tests are done double blind and repeatable, they are totally worthless. People, even testers, just have biases. It's only human.

Edit: Sorry, 7970GE crossfire uses 2x the power of a 690? Are you sure about that?

 

[boxleitnerb]

I didn't dispute the measured fps, where did you get that? I merely noted that this is only a part of the final experience with AFR.
How do you propose one should measure/judge the smoothness/microstuttering? Just because there seems to be no objective measurement technique as of yet doesn't mean this topic should be completely excluded from a review. It will affect the user after all (depending on sensitivity of course).

 

[3DVagabond]

I didn't dispute the measured fps, where did you get that? I merely noted that this is only a part of the final experience with AFR.
How do you propose one should measure/judge the smoothness/microstuttering? Just because there seems to be no objective measurement technique as of yet doesn't mean this topic should be completely excluded from a review. It will affect the user after all (depending on sensitivity of course).

Unless they can quantify measurements with what they see the measurements alone are worthless. That's beyond any expertise I have. So, it's hard for me to suggest any particular measuring technique. I already said how they would have to do a visual comparison, double blind. There is no other scientifically acceptable method of comparison. Neither the person administering the test or the subject have any knowledge of which card(s) they are viewing. The subject reports what (s)he is seeing. The differences, if any, are noted, and the tests are rerun. Typically 10 times. If the subject can't repeat their results then they are deemed to not be of any scientific relevance.

Now, they didn't say there was any issues. They do mention banding in Metro and say how good the surround experience was, overall. They appear to have watched the benchmarks as opposed to just letting them run and check the results. I'll ask them if they can specifically address any microstuttering or other shortcomings in the comments section of the review.

 

[boxleitnerb]

10 times, I doubt that any reviewer has that much time (and test subjects). I think it is better to have a possibly flawed observation than none at all. Because pure fps are of little help here as well.

Watching a benchmark and playing a game yourself is vastly different. If the (single?) reviewer is not susceptible to microstuttering, this would be useless anyway. We desperately need some kind of metric (not frametimes, they don't account for frame metering) to judge the quality of those measured fps.

 

[3DVagabond]

10 times, I doubt that any reviewer has that much time (and test subjects). I think it is better to have a possibly flawed observation than none at all. Because pure fps are of little help here as well.

Watching a benchmark and playing a game yourself is vastly different. If the (single?) reviewer is not susceptible to microstuttering, this would be useless anyway. We desperately need some kind of metric (not frametimes, they don't account for frame metering) to judge the quality of those measured fps.

Well, you are back to some sort of measurement to quantify the smoothness. To do that we first have to establish that the phenomena exists. That is done with repeatable double blind testing. Once we've established that it exists then measurements can be taken to correlate a cause to the effect. For something as basic as this, it wouldn't be difficult to do.

I assume the point of your posts is to cast doubt in the legitimacy of the review because of a lack of information. Very well, point taken.

 

[boxleitnerb]

Yes it is, I openly admit that.
Have you ever read that CF feels smoother than SLI? I haven't, and I have read many many reviews and user comments on this topic. The other way around, however, is mentioned quite often and consistently by different people/sites. This in itself would constitute some form of blind test, would it no? That microstuttering exists with AFR is a proven fact.

 

[3DVagabond]

Yes it is, I openly admit that.
Have you ever read that CF feels smoother than SLI? I haven't, and I have read many many reviews and user comments on this topic. The other way around, however, is mentioned quite often and consistently by different people/sites. This in itself would constitute some form of blind test, would it no? That microstuttering exists with AFR is a proven fact.

I'm not saying that it doesn't exist. There is enough evidence to say that it is likely a real occurrence. It hasn't been established though exactly what it is. I've read where some people say it's there all of the time with multi GPU setups. Both brands regardless of frame rates. I've read where other people have said it doesn't exist and it's merely imagined or people have been convinced it occurs through suggestion (If you've never witnessed it you'd be shocked at what the power of suggestion can make people perceive.). I've read where people have said that it requires twice the frame rates as single GPU to be as smooth. I've read where people have said that it occurs at less than 70 or 80 fps. So, exists is probably too strong of a word in this instance. First it would have to be made to occur under controlled conditions that could be repeated. Reports on the internet don't count.

 

[Skurge]

Techreport has some framtime measurements, if you read how they do the tests and what the numbers mean they are probably one of the best GPU review sites out there.

http://techreport.com/articles.x/23217/9

It's not SLI or crossfire, but if you look at BF3 (at my resolution). The 680 is ahead in FPS, but the 7970 is smoother, BUT if you look at the actual numbers the difference is so small is imoreceivable.

Looking once again at BF3 (some people think it's the only game on the planet), even with older drivers 7970 is faster and smoother. It's got a big FPS advantage, but not much of a smoothness advantage.
http://techreport.com/articles.x/22890/7

http://techreport.com/articles.x/22890/5
Looking at Skyrim, the 7970 is a lot less smooth. The new drivers improve FPS, but it remains to be seen if that makes it more smooth.

The other games they are pretty much neck and neck. So from that review it looks like SLI can be the smoother in general, but it was donw with older drivers.

And it looks like the case of how smooth SLI or Crossfire is a case by case thing. You can't really says SLI is smoother than CF. The only Evidence we have of how smooth either one is are the Techrepot articles. Unless you know of a website that does something similar or better.

 

[3DVagabond]

Techreport has some framtime measurements, if you read how they do the tests and what the numbers mean they are probably one of the best GPU review sites out there.

http://techreport.com/articles.x/23217/9

It's not SLI or crossfire, but if you look at BF3 (at my resolution). The 680 is ahead in FPS, but the 7970 is smoother, BUT if you look at the actual numbers the difference is so small is imoreceivable.

Looking once again at BF3 (some people think it's the only game on the planet), even with older drivers 7970 is faster and smoother. It's got a big FPS advantage, but not much of a smoothness advantage.
http://techreport.com/articles.x/22890/7

http://techreport.com/articles.x/22890/5
Looking at Skyrim, the 7970 is a lot less smooth. The new drivers improve FPS, but it remains to be seen if that makes it more smooth.

The other games they are pretty much neck and neck. So from that review it looks like SLI can be the smoother in general, but it was donw with older drivers.

And it looks like the case of how smooth SLI or Crossfire is a case by case thing. You can't really says SLI is smoother than CF. The only Evidence we have of how smooth either one is are the Techrepot articles. Unless you know of a website that does something similar or better.

I'm familiar with what they do and it very well might have a correlation with microstutter/perceived smoothness. It hasn't really been verified though.

 

[cmdrdredd]

3240x1080 is smaller than 3600x1080, yet their results match up pretty close with quad 680 performance in BF3 using a 2.0 bus.

Those same cards jump from 97 fps to 140 fps by switching the bus to 3.0 in the registry, on a x79 platform no less.


Seems reasonable to assume it's possible the review was done with quad sli operating over 2.0 bandwidth.


Edit: Not that 2.0 vs 3.0 would affect a clear and present vram limitation such as in your example, however in the other tests where vram was not thunderstruck, it could explain some of the scaling issues present.

Right but the idea is if you get fps that is not high enough with one card you can add a second and get playable fps. I get what you are saying tho...xfire may have higher fps but more microstutter so sli may feel smoother anyway.

 

[SolMiester]

The actual subject of the review was gtx-690's in SLI. It wasn't intended to be an all encompassing SLI vs. Crossfire review.

I found it interesting that the badly broken crossfire that everyone is being warned to avoid doesn't seem to be too broken after all. I find it just as intriguing that when the objective measurements don't follow our beliefs we refuse to accept it without further proof. I hate to say it, but the last thing we could trust is anyone's subjective assessment. Until tests are done double blind and repeatable, they are totally worthless. People, even testers, just have biases. It's only human.

Edit: Sorry, 7970GE crossfire uses 2x the power of a 690? Are you sure about that?

After 8mths, you would hope it wasn't still broken..!

 

[RussianSensation]

Hey Sol, you still gaming on a 9600GT? Or is that just a place holder in your signature?