CrossFire and SLI frame rates do not reflect reality because of lack of synchronization!

[Datenschleuder]

Hello!

The by far most commonly used SLI/CrossFire mode is alternate frame rendering, where one card draws all odd, and the other one all even frames.

So if a scene is for example rendered with 30 FPS, every frame takes ~33.3 ms to update and we should get something like the following frame times:

WHAT 30 FPS SHOULD LOOK LIKE

Frame, Time (ms)
1, 0.000
2, 33.3
3, 66.6
4, 99.9
5, 133.2
6, 166.5
7, 199.8
8, 233.1
9, 266.4
10, 299.7

This also is what 30 FPS look like on a single card setup.

But how does these 30 FPS look like when the scene is rendered using SLI or CrossFire?

WHAT 30 FPS DO LOOK LIKE WITH SLI/CROSSFIRE

Frame, Time (ms)
1, 0.000
2, 12.1
3, 68.3
4, 80.5
5, 136.8
6, 148.9
7, 205.1
8, 217.3
9, 273.7
10, 285.7

Every odd frame number takes much less than the expected 33.3 ms to update and every even frame takes much longer to update.

In this example, every even frame takes around 56.5 ms to update, which is almost 70% longer than expected!
The frame rate says ~30 FPS, but in reality every second frame is updated with a rate of only ~17.8 FPS!

This problem becomes very noticeable at low frame rates of course.
You can test this on your own quite easily by logging the frame times with a tool like Fraps.
Driver versions, SLI/CrossFire compatibility flags and the kind of the Windows OS has no impact on the problem.

The issue exists since SLI/CrossFire got on the market and has already been under discussion in German forums for a while and was reproduced by all SLI/CrossFire users there.
There is even a German speaking article that demonstrates the problem with a video.

I think that this is quite a scandal and that it is in the interest of the whole enthusiast market to make this public, so that ATI/Nvidia finally delivers a solution for this.

Thanks for your support on this!

[CP5670]

Yeah, this has been discussed here before. I've seen this effect in many games and it's one of the reasons I am staying far away from anything multi GPU in the future.

It seems that only some people notice this issue, however. I think more people would see the difference if they compared it to an identical framerate on a single card (which is not that easy to do in practice).

[Jax Omen]

I can't even tell the difference between an 8ms LCD and a CRT (as far as responsiveness), I don't think I'll need to worry about this

[Datenschleuder]

Yeah, this is discussed in extremely isolated spots in the web, but unfortunately 99% of the market doesn't know about this issue because major review sites like Anandtech do not report on this.

And this really has to change at last!

It is somewhat difficult to make out visually at frame rate over 30 FPS, but what you can easily verify this on your own by logging the frame times.
All SLI/CrossFire users are effected by this problem unfortunately.

[Pelu]

omg... i got crossfire... that means the rendering is crap with this issue.???

[Datenschleuder]

Originally posted by: Jax Omen
I can't even tell the difference between an 8ms LCD and a CRT (as far as responsiveness), I don't think I'll need to worry about this

This is no valid comparison.

People buy a second graphics card because they want improved performance and make their decision based on frame rate comparisons in reviews.
The problem is that these frame rates don't reflect reality at all because of this problem.
Or don't you agree that there is a huge difference between 30 FPS and 17 FPS?

So this is very significant to worry about!

[Jax Omen]

I don't think it's significant in anything but Crysis, as any modern dual-GPU setup will far, far exceed 30fps in every other game out there.

Hell, most good single cards can pump out 60+fps on most games at 1920x1200.

[Datenschleuder]

Originally posted by: Pelu
omg... i got crossfire... that means the rendering is crap with this issue.???

Yes. You can watch the linked video and I think you can clearly see the difference.
I am extremely frustrated by this as well because I spent a lot of money into my SLI hardware.

We were all deluded badly! :thumbsdown:

[Datenschleuder]

Originally posted by: Jax Omen
I don't think it's significant in anything but Crysis, as any modern dual-GPU setup will far, far exceed 30fps in every other game out there.

Hell, most good single cards can pump out 60+fps on most games at 1920x1200.

The problem exists with 60 FPS as well.
I don't know about you, but I didn't buy SLI cards so that I could get only ~35 FPS instead of the promised 60 FPS!

[CP5670]

Originally posted by: Datenschleuder
Yeah, this is discussed in extremely isolated spots in the web, but unfortunately 99% of the market doesn't know about this issue because major review sites like Anandtech do not report on this.

And this really has to change at last!

It is somewhat difficult to make out visually at frame rate over 30 FPS, but what you can easily verify this on your own by logging the frame times.
All SLI/CrossFire users are effected by this problem unfortunately.

It has been a while since I used SLI, but from what I remember it was actually more prominent at higher framerates (the perceived framerate was correspondingly high, but the disparity between that and what you would expect to see from the recorded framerate was larger). A lot of games never looked entirely smooth to me on AFR, even with framerates over 100.

I think this problem is related to the incompatibilities with vsync and triple buffering, since things looked normal if I was able to get those settings working in a game (although back then, getting both working right was a rarity in itself).

And yeah, it would be good if AT or one of the other well known sites investigates these issues in depth.

[Sylvanas]

I can't notice anything like this so it doesn't phase me....

[Datenschleuder]

Originally posted by: Sylvanas
I can't notice anything like this so it doesn't phase me....

You don't care that most of the money you spent on SLI/CrossFire is going down the drain because of this?

[Jax Omen]

It's only "going down the drain" if he can notice the difference. If he feels it performs better, then he got his money's worth.

Obviously many people feel SLI/Xfire perform better, as many people set up multi-GPU rigs over and over, as new GPUs come out.



Question: I don't have SLI/Xfire, so I can't exactly test this, but can't you still use SFR to do it, and avoid this problem entirely? Sure, it wouldn't give as high of a synthetic framerate as AFR, but wouldn't it fix this problem?

[Datenschleuder]

Originally posted by: Jax Omen
It's only "going down the drain" if he can notice the difference. If he feels it performs better, then he got his money's worth.

But I think most people don't agree with that.
At least I haven't bought SLI for a placebo effect.

Question: I don't have SLI/Xfire, so I can't exactly test this, but can't you still use SFR to do it, and avoid this problem entirely? Sure, it wouldn't give as high of a synthetic framerate as AFR, but wouldn't it fix this problem?

SFR generally performs very bad if it works at all.
It's often much worse even than single card mode.

[bfdd]

I understand what you're saying, but you're forgetting something, now that a single card doesn't have to produce two frames in sequence a single card in sli runs faster than a single card period. Comparing 30fps in a single card setup vs 30fps in an SLI setup isn't quite fair imo. Seeing as the SLI setup should get much more fps so it's not noticeable. Another thing, you're not getting 40% of the frames. that's why you normally see a 30-50% or higher improvement moving to a multi gpu setup

[apoppin]

Originally posted by: Datenschleuder

You don't care that most of the money you spent on SLI/CrossFire is going down the drain because of this?

What is in your rig ?


personally i think you are full of .. BS "theory"

it isn't how it works .. i have not time to explain and i would not bother if i did
--try it for yourself then blab here from experience
--i can see the differences - i also know the disadvantages - and my Crossfire is way faster than a single GPU .. and you have no clue

[img]i/expressions/rose.gif[/img]

the ONLY time Crossfire is *useless* - is if the game does NOT "scale" and perhaps when you are running way below 30FPS with Xfire - then who cares, as it is below the practical experience for "playing games" and you are just benching a slide show [anyway]; comparisons with a single GPU to Xfire - each way over 60FPS - is also of no consequence as far as i am concerned,

and your comparison is silly .. One 2900xt gets me 30FPS .. with Fraps .. adding my other 2900 to pair it in Crossfire and i *now* get 50 FPS .. also with FRAPS [for example]
--now what are you saying again?


i DO benchmarking as well as any of 'em
- what would you like me to test for you? .. i probably already have it

Be nice, Apoppin. You're starting to skirt the line on rudeness.

-ViRGE

[Pelu]

is this problem getting a relief if vsync and triple buffer is used?

[Sylvanas]

Originally posted by: Datenschleuder

You don't care that most of the money you spent on SLI/CrossFire is going down the drain because of this?

Thats hilarious. My bike has scratches on the paintwork- does that mean I cant get from A to B anymore?

Jax Omen:

It's only "going down the drain" if he can notice the difference. If he feels it performs better, then he got his money's worth.

But I think most people don't agree with that.
At least I haven't bought SLI for a placebo effect.

Oh right.....so if it is the placebo affect I'm not ACTUALLY getting better fps! Neither is Anandtech or any review site that bench multiGPUs. Have you asked 'most people' that have a multi-GPU setup for their opinion? Or just a small percentage of enthusiast who happen to browse your same forum and of that small % a few jump on the bandwagon or agree with your flawed logic......

Yes...... that represents and accurate portrayal of THE WORLD as most people in the world with CF/SLI HATE it! Thats why they bought it!

/end sarcasm

[v8envy]

Ok, think about this for a moment. If you're about to perform an action (let's say fire your weapon) right before video card number 1 starts rendering a frame. Let's also assume that single video card would render at 15 fps. What's the minimum latency until you see the result?

1000/15 = 66.7 milliseconds. The second card won't help here -- it needs to wait until card 1 does at least some work in order to not duplicate the work.

Now, let's say a single faster card was running at 20 fps. That's 50 ms, for a savings of 16.7 ms, or exactly the time which passes between screen refreshes on an LCD.

Worst case you could be looking at a 32 ms additional latency by going with 2x slower cards, although the FPS you see in FRAPs is 30. Also, best case you'd have 16.7 finishing just in time for the monitor to refresh. Hence, microstutter.

This isn't an issue for slower paced non-twitch games and/or running at very high frame rates per card. But if you are running at > 60 FPS on one card, why do you need 2? Since benchmarks tell you you're running at > 30 fps you may even turn on far more eyecandy than if you knew you only had 15 fps. Also, that kind of firepower is commonly paired with large monitors -- which have additional input lag and slow refresh compounding the problem.

If you don't play twitch shooters and don't notice microstutter then SLI + large LCD is not a bad way to fly. For the rest of us I'm glad there are single card solutions and CRTs.

[apoppin]

If you don't play twitch shooters and don't notice microstutter then SLI + large LCD is not a bad way to fly. For the rest of us I'm glad there are single card solutions and CRTs.

i do notice and i also play twitch shooters ..
... everything is compromise .. but the *cure* for most of us - except to buy a faster single GPU [which for example, could not happen if you had purchased a 8800GTX-Ultra back last - last November. You still can't buy a single faster GPU - it will be TWO years!! .. So ..
If you want "faster" then your choices are - (1) to lower your resolution, (2) sacrifice details - OR (3) buy a second one

Problem Solved for most of us by #3
- the least compromise

[img]i/expressions/rose.gif[/img]

[BFG10K]

I think that this is quite a scandal and that it is in the interest of the whole enthusiast market to make this public, so that ATI/Nvidia finally delivers a solution for this.

I don't think this problem can be "solved" as it's inherent to AFR. It's like asking SLI to solve the issue of not getting full scaling in every game; that's simply not going to happen.

What they could do is hold back the faster frames to make them more in line with the slower frames but I can see that getting very messy very quickly with questionable benefits at best.

Question: I don't have SLI/Xfire, so I can't exactly test this, but can't you still use SFR to do it, and avoid this problem entirely? Sure, it wouldn't give as high of a synthetic framerate as AFR, but wouldn't it fix this problem?

Yes, SFR fixes a lot of issues such as micro-stuttering (the issue being discussed here), input lag, and other problems inherent to AFR. The VSA-100 employed a form of SFR at the hardware level which is why it had consistently better compatibility and more consistent scaling than today?s solutions, all without the need for application profiles.

I say ATi should bring back super-tiling, the ultimate SFR implementation. Among other things it does automatic load-balancing at the hardware level which in theory could deliver scaling without the need for application profiles.

[lavaheadache]

I'm very sensitive to this as i dicussed in a topic not too long ago which was titled with Micro stutter... anybody else annoyed(too lazy to link it up). I think the problem can be remedied somewhat by using "smooth mouse" or maybe im a total idiot

[apoppin]

Originally posted by: BFG10K

I don't think this problem can be "solved" as it's inherent to AFR. It's like asking SLI to solve the issue of not getting full scaling in every game; that's simply not going to happen.

What they could do is hold back the faster frames to make them more in line with the slower frames but I can see that getting very messy very quickly with questionable benefits at best.

Yes, SFR fixes a lot of issues such as micro-stuttering (the issue being discussed here), input lag, and other problems inherent to AFR. The VSA-100 employed a form of SFR at the hardware level which is why it had consistently better compatibility and more consistent scaling than today?s solutions, all without the need for application profiles.

I say ATi should bring back super-tiling, the ultimate SFR implementation. Among other things it does automatic load-balancing at the hardware level which in theory could deliver scaling without the need for application profiles.

agreed .. what they do now works OK .. but if you are "sensitive" ...
. . . get an Xbox

i think they are working on all of this .. AMD needs to be competitive with more GPUs scaling better together than NVIDIA does. i see something new with every cat release; and i am not so sure it is strictly AFR anymore ... i am just starting to look at Cat 8.4 and it looks like there is also super AA with edge detect with some unreal AA possibilities .. AMD is gonna have to tap alternate forms of Multi-GPU rendering imo to stay competitive and i think they are doing just that; without announcing anything.
--the fact remains if you want FASTER - with the least compromise .. you need another GPU [or two]


look, i have played with Single GPUs and Crossfire .. there IS a definite advantage to playing with one - IF it is fast enough for you. BUT if you need to go faster, the disadvantages of Crossfire are far outweighed by the advantages and i can handle the little "oddities" for overall faster [and smoother - for me] FPS; and i won't compromise on detail.

[i can't speak for SLi personally]

[BFG10K]

agreed .. what they do now works OK .. but if you are "sensitive" ...
. . . get an Xbox

A better idea is to get a GT200.

[VERTIGGO]

So that's why it says 45-50 fps in crysis at certain settings but it still chugs like a frat boy.