PCGH: Micro stuttering on multi GPU solutions

[Fr3ak]

With this thread in German 3D Center Forum, the whole things starts and reached almost any IT website. The summary of the thread creator: Although SLI allows higher frame rates, the perceived outcome is often not better than the lower framerates of a solo graphics card. Meanwhile, it is clear that not only (triple) SLI, but Crossfire is also concerned. And as we said in our testing of multichrome, S3 doesn't make it better than AMD / Ati and Nvidia.

Source (english):
http://www.pcgameshardware.de/?article_id=631668

At the end of the article there is a video showing NFS Pro Street and UT3 on a HD3870 X2 with both crossfire enabled and disabled. I have to admit the video smoothens everything a bit. The difference is much bigger when you see it with your own eyes. NFS becomes almost unplayable when running crossfire. On UT3 you can see the stuttering much better. Though the game runs 10 Fps slower without CF, it feels way more smooth.

[Unbornchild]

Thanks Fr3ak.
Hmm... i was often tempted to get Crossfire, especially with the new 3870X2 card, but held it off till today. I don't regret it. Still sitting on my saved money for the convincing best solution. I'll only go Crossfire or X-fire when it gives a better result than a single card in all available games.

[EDIT]: and i mean the subjective result, the experience, feeling of the game running completely smoothly, during gaming. Not just the clinically measured fps.

[OBR]

micro stuttering is very funny naming ... as long how MultiGPU exists this is named input lag ...

[Fr3ak]

To make it clear, this article is not about CF performance which can reach only 60% of a single card in some games. The article is about the gaps between frames. Using a single card the gap of the frames is linear and only varies by a small amount that is not noticable. With CF or SLI on the other hand the displayed frame times vary a lot which looks to the human eye like stuttering although the Fps are way higher tan it feels like.

OBR: I think it's rather output lag, isn't it?

[Hyperhorn]

Sometimes I had strange feelings about my 7950 GX2 and UT3. Knowing about micro stuttering it's maybe time to re-investigate it. I guess it was because framerate dropped sometimes at ~40 fps, but the smooth-feeling was more affected (--> because total time intervals were larger) . The whole thing is troublesome, NVidia/ AMD have to find a solution.

[STaRGaZeR]

I've been saying this in the 3870 and 3870X2 threads. Glad to see it's actually confirmed and reviewed. Also, Fr3ak if you have the chance as a reviewer, please do some pressure if you can to ATI and NVIDIA. We need it.

And it has nothing to do with input lag.

[Apoc@]

Nice review.

[realsmasher]

which driver was used ?

i just ask, as crysis is mentioned on the website while already is proven, that crysis has no more "Micro stuttering" problems at all with the 3870x2.

don't know about the other games though.

but HOW do you get 25fps in UT3 ?
even with single 3870 you get 30+ fps in 2560*1600 EASY.

[AnarchX]

micro stuttering is very funny naming ... as long how MultiGPU exists this is named input lag ...

input lag != unsteady frame-times

[Fr3ak]

Testsetup being used for the video:
Cat 8.1, C2D E6700 @ 400x8, XP x86 SP2, Gigabyte X38, 2 GiB DDR2-800

The Fps are that low because of AA being used in combination with a high resolution. I didn't make the test myself, so I cannot tell you what exactly has been used. But what you can clearly see is that even 25-30 Fps on UT3 feel way less smooth than the 15 Fps with CF disabled.

We made sure it's not just a weird behaviour with that specific hardware configuration. It's the same on skulltrail, P35 and 780i using completely different components.

[saaya]

use fraps to look at the frame time
i started to work on an article about this years ago after ati just introed xfire but then never finished it cause i moved to the uk

multi gpu solutions lack frame syncing in afr modes, in many scenarios you get both cards finishing frames at almost the same time. so you do get more frames per second on your display, but the final result looks almost like showing each frame twice within a short period of time.

frame time looks a bit like this:

1:00:01 frame 1
1:00:04 frame 2
1:00:31 frame 3
1:00:34 frame 4

frame 1 and 2 and 3 and 4 are so close to each other they are almost identical and it would look almost the same to show frame1 twice and frame3 twice.

back then i told ati about this and they said they know about this and are working on tile mode and better syncing of the cards. but it seems most of the time cards still work in afr which gives better fps scaling and the syncing still doesnt seem to work out that well either.

from my experience it seems that the lower the fps the more apparent the flashing, the higher the fps the less you note it.
so xfire seems to make gaming look smoother at high fps, at low fps it actually makes it look worse tho.

EDIT: lol i just read the article, seems you already found the same things
check other games and benchmarks, some stutter much worse than nfs, i think in cod2 the frame times for 2 consecutive frames was almost identical.

[realsmasher]

Testsetup being used for the video:
Cat 8.1

ok, then i understand it.
8.1 had all those problems.

with newest driver at least crysis does not have this problem anymore.
i can link the frametimes if anyone wants to see.

[Oc-Ghost]

Secondary to primary frame buffer lag.. didn´t think it existed because of the bridge.
But then again thanks to serial connections, we have some lag.
Guess that´s the only edge parallel connections will ever have.

[Shintai]

GPUs is like women, you simply ask for too much trouble if more than 1

[Lestat]

GPUs is like women, you simply ask for too much trouble if more than 1

depends on how big your card is lol

[tombman]

EDIT: lol i just read the article, seems you already found the same things
check other games and benchmarks, some stutter much worse than nfs, i think in cod2 the frame times for 2 consecutive frames was almost identical.

Hi, i'm the original author of the thread "die sli Lüge..." at 3dc forums and i'm happy that this problem is starting to get spread across the web.

ATI and Nvidia obviously don't want the world to know, because it will severly hurt the glamour of multi-gpu solutions which is the future of highend graphics!
The solution to this problem would be to enable some kind of intelligent frame limiter, which can easily be done within the driver, but will cost a small amount of average frames per second. And losing framerates will hurt benchmarks and therefore reputation.

Some games, that have a built-in framelimiter, will show perfectly even frametimes no matter how much gpus you use, so there IS a solution to this problem. In fact, a user @3dc forums already programmed a litte tool that can limit the framerate in all d3d9 applications, but then suddenly "disappeared" and never came back.

All this tool did was to delay the output of frames depending of previous frametimes. It wasn't intelligent, so you could only lock to a particular fps rate, e.g. 30fps or 40fps. Making it intelligent, so it could adapt to whatever framerate necessary in realtime, would have been no big deal, but as i said, he never came back to finish his work.

I hope, that with enough pressure from all of the "freaks" and "nerds" on the internet, ATI and Nvidia will have to come up with a solution, namely a small switch in the control panels of the drivers- for example they could call it: "syncronize/harmonize frametimes for SLI/CF" [Yes] [No]

[Lightman]

vSync if working is sorting this problem in CF
At least it does on my RIG under XP x64

[L7R]

Great article! I found this issue immediately when I get my first CF system. It's not in every game but almost all new games suffers this issue somewhat. Crysis is best example with early drivers. It runs flawlessy and steady 28-35fps with single card and actually feels somewhat smooth. But when enabling cf I get 50-100(!)% increase but its like every other frame (or even many frames) are not being rendered. Framerate jumps between 20-100 and it stutters all the time. It's very hard to explain but if framerate is 30 with single card, then compared 30fps in CF it's very much unplayable in many games. What affects is your resolution: For example Oblivion, CF stutters more on 1680x1050 without AA and AF than 1400x900 AF8+AA even both have same framerates. VSync helps but not when fps drops under 60 or 30.

Another example is Quake 4. With CF it stutters badly even when framerate drops from 60 to 54! Fortunately those under 60fps moments happens only a few times. With single card you can't even notice those kind of drops.


Games which have been fixed with driver updates:
CoD4 with 8.2beta -- Fps is much of the same but feels very smooth even in high res.

Crysis with 8.2beta -- same here. Now it feels as smooth as fps counter shows.

Stalker with 7.11 -- Just GREAT! This was bad even though fps counter showed high framerates.


CF is fantastic if your system can run a game 60fps minimum so you can hit vsync on (like Half life 2 or Stalker with maxed out settings).

[tombman]

vSync if working is sorting this problem in CF
At least it does on my RIG under XP x64

Vsync cannot solve this, as it is a completely different problem. (i can give explanation if requested)

[MuffinFlavored]

GPUs is like women, you simply ask for too much trouble if more than 1

Thanks.

I thought multi-GPU was all about rendering the upcoming frames, but that would mean they frames would be ready one after another and there would be no stuttering.

I guess this is not how it is. How is multi-GPU then? All work to render one frame?

[Richie P]

VERY interesting. Thanks a lot for the insight and head's up

[foxmino]

Is the clear line scrolling down the monitor in some games the same thing as the stuttering issue or is this a v-sync problem?

[realsmasher]

well, that's a good question.

if you keep in mind, that the frame from the second card comes close after the first, this "line" should appear much more often if there is micro stuttering.

but anyway, you can avoid it with v-sync.(the line, not the stuttering)

[saaya]

Secondary to primary frame buffer lag.. didn´t think it existed because of the bridge.
But then again thanks to serial connections, we have some lag.
Guess that´s the only edge parallel connections will ever have.

thats not it, you could delay the output of a frame to compensate for that, but the problem is the time the frames were rendered at, so what time in the game they show. the frames were finished almost at the same time and showing almost the same thing in the game.

so delaying one of the two so the frame times look smooth will only hide the problem, but it wont fix it. you would still see almost the same frame twice and only then see a new frame, again twice. so it will look like its stuttering, you see almost no movement between frames 1 and 2 and then to 3 theres a big difference, 3 and 4 are almost the same again and 5 is a big difference again...

From what saaya is saying though, there is no sync'ing? That just seems rediculous.

i dont know if there is or isnt, but the the frame calculation is def not perfectly in sync as ideally you want a crossfire system to shoot one frame from card 1 and then 1 frame from card2. and not have frame 1 and 2 come at almost the same time. then you are shooting 2x as many "bullets" as with 1 "gun", but your not producing a steady "cover fire" which is what crossfire is all about

Hi, i'm the original author of the thread "die sli Lüge..." at 3dc forums and i'm happy that this problem is starting to get spread across the web.

ATI and Nvidia obviously don't want the world to know, because it will severly hurt the glamour of multi-gpu solutions which is the future of highend graphics!
The solution to this problem would be to enable some kind of intelligent frame limiter, which can easily be done within the driver, but will cost a small amount of average frames per second. And losing framerates will hurt benchmarks and therefore reputation.

Some games, that have a built-in framelimiter, will show perfectly even frametimes no matter how much gpus you use, so there IS a solution to this problem. In fact, a user @3dc forums already programmed a litte tool that can limit the framerate in all d3d9 applications, but then suddenly "disappeared" and never came back.

All this tool did was to delay the output of frames depending of previous frametimes. It wasn't intelligent, so you could only lock to a particular fps rate, e.g. 30fps or 40fps. Making it intelligent, so it could adapt to whatever framerate necessary in realtime, would have been no big deal, but as i said, he never came back to finish his work.

I hope, that with enough pressure from all of the "freaks" and "nerds" on the internet, ATI and Nvidia will have to come up with a solution, namely a small switch in the control panels of the drivers- for example they could call it: "syncronize/harmonize frametimes for SLI/CF" [Yes] [No]

interesting to see how many people noticed this but it never really became public

the tool your explaining doesnt solve the problem though, it will only have a cosmetic change in that the frame times will look fine, but as i explained above, the game will still stutter... the stuter would actually be even perceived worse cause you stretch the time between the two identical frames!

[DefStar]

Yes, the stuttering is quite obvious. When I had my 7950 gx2 I allways adjusted my overclock and settings thinking something was just off. I had never used sli before, and because I got the 7950 gx2 and a whole new system at the sametime it was frustrating trying to isolate the problem. Anyhow, I knew it was a side effect from sli even though nobody ever mentioned it because they were to arrogant and ignorant about having their oh so powerfull sli system. Anyhow I got lucky and did the stepup program for 8800 gtx but untill then I felt so burnt and so pissed that I fell for sli marketing it made me sick. I sware it was the most annoying thing It was just enough to make you think wtf, wtf is that, it seems like a lag spikes.