HD5970 Microstuttering tests

[annihilat0r]

Microstuttering is a giant multi-GPU problem and is why I had felt the games were smoother when I went "down" from a SLI rig to a single 4870 a year ago, even though the FPS of my SLI rig was faster. Unfortunately this big, big issue gets often overlooked. It's completely unrealistic to compare AFR frame rates with normal frame rates.

I was yet to see a microstuttering analysis for HD5970 but some guys at Donanimhaber (cenova and Source) did tests with a good selection of games: Crysis, Crysis Warhead, Batman:AA, MW2, GTA4, Dirt 2, Unigine Heaven, World in Conflict, Wolfenstein 2.

http://forum.donanimhaber.com/m_3702..._/key_//tm.htm

Out of these popular titles, the only game that wasn't reasonably affected by MS is Wolfenstein 2. All others are significantly affected.

In the graphs the X axis is the frame rendered and the Y axis is the amount of time in milliseconds that it took to render the frame at that exact time.

Crysis@1920 (1680 doesn't display much MS, weirdly)


Here you can see that FPS varies wildly between 40 and 75. Most probably, in these instances FRAPS showed FPS around 62, whereas it was 40 fps-smooth at that time. This is why you can't trust AFR's FPS.

Batman:AA@1920


MS kills this game even more, with FPS varying between 40 and 100(!).

MW2@1680 (1920 shows less MS in this game)


There is terrible MS half of the time, varying between 40 and 200. Don't forget that if momentary FPS goes 10-150-10-150-10-150 the smoothness won't be the average of that FPS. It will be the minimum. You will see 10ish FPS smoothness.

Dirt 2@1920 (In 1680 there is little MS.)


WTF. This could be the definition of MS.

GTA4@1920


Unigine@1920


And you thought it couldn't get worse than Dirt2.

BOTTOM LINE: Never compare dual GPU FPS to single gpu FPS. Back when I had a G92 SLI system I had concluded that the second GPU, while multiplying FPS by about 1.5-1.6, didn't contribute anything to the smoothness.

[Xoulz]

Thnx.. but that's been known forever! People who brag or use SLI for anything other than pushing extra pixels are using an epeen. (sorry) I have money, but never once considered SLI. That is why I was interested in Hydra. But I feel that technology is still a good year from tweaks in overtaking SLI/Xfire.

So in the mean time, single best single card has always been my choice. So where does that leave me this round?

[saaya]

awesome article!

[NickF]

This is why dual GPU cards should be communicating through an HT-esque bus, sharing memory and NOT using the same rendering ideas from the ATI Rage Fury MAXX of 10 years ago.

[jaredpace]

I had concluded that the second GPU, while multiplying FPS by about 1.5-1.6, didn't contribute anything to the smoothness.

I agree!

[zerazax]

How do the captions calculate out FPS compared to the ms response from frame rendering? That's what I'm confused by right now but that might be a wording issue

Also, did they provide baseline/control graphs? It's one thing to graph results, but what about a single gpu solution? AKA "normal" behavior? That would be more illuminating

[grimREEFER]

yea, there should be some type of control.
i think this may look similiar to the type of lag you get with vsync even on a single gpu.

[Humminn55]

Few questions:

First, as was pointed out, where are the single gpu control graphs?

Second, where are the FPS graphs? All I see is a graph showing a frame number and time interval, no fps charted.

Third, when Batman was run, why was it run with AA on, since it's well known that AA is intentionally crippled in that game for ATi cards, unless the game's profile was "fixed".

[saaya]

This is why dual GPU cards should be communicating through an HT-esque bus, sharing memory and NOT using the same rendering ideas from the ATI Rage Fury MAXX of 10 years ago.

its beyond me why amd hasnt equipped gpus with a HT interface yet...
lower latency than pciE = slightly better perf and faster frame times even at lower fps = better gaming experience even at low fps + the usage for dual gpu would make a whole lot of sense... they could use NUMA and save half the memory on the cards or share the entire memory of both or more cards instead of mirroring the data...

if i read the graphs right, frame time delta is displayed on the left, and frame number is displayed on the bottom.
average frame time delta is around 25ms, that would mean 1000/25=40fps

frames are rendered in batches, thats the problem with afr... too large batches without any syncing between the batches.
when the frame time delta, the time between one frame and the next, decreases, is below average, then the graph goes down, you see several frames within a short period of time, this part feels great and the game feels fluid. but then, the batch is done, and the card(s) start to render a new batch of frames, at this point in time the time from one from to the next increases and in many cases instantly jumps up, this is causing a spike in the graph going UP, and this is the part that feels terrible when playing the game, you only see a few frames in a certain amount of time, and the delay from one frame to the next gets worse and worse.

this can, and should, be fixed by ati, but it hurts performance... youll end up with a lower average fps, naturally...
games can fix this as well by managing the gpus properly and distributing the work more evenly, and making sure the stream of work to the gpu and from the gpu is even and not fluctuating... basically forcing gpus to only pre-render a low amount of frames SHOULD improve the situation a lot...

[Shadov]

Since when does a fraps graph represent the experiance of a human eye?

The intervals could be even longer, while you wouldnt even ackonwledge it, which has been proven before already many times.



Edit: All Im saying is I have played mentioned titles on an X2 card and havent noticed any MS.

[NickF]

its beyond me why amd hasnt equipped gpus with a HT interface yet...
lower latency than pciE = slightly better perf and faster frame times even at lower fps = better gaming experience even at low fps + the usage for dual gpu would make a whole lot of sense... they could use NUMA and save half the memory on the cards or share the entire memory of both or more cards instead of mirroring the data...

One would think after AMD acquired ATI, the 3870 X2 basically flopped, that ATI would have at the least put an HT link on the table for an upcoming generation. With the success of the 4870 X2, and the currently unchallenged performance of the 5970, I fear that ATI may have gotten lazy in this regard. I mean, its very plausible that they could do HT with a 6000 series card, but not likely.

[saaya]

Snnce when does a fraps graph represent the experiance of a human eye?

The intervals could be even longer, while you wouldnt even ackonwledge it, which has been proven before already many times.

the reason this is getting so much attention is that there IS a perceivable problem with multi gpu setups, if you cant see and feel it, good for you... but most people do...

the problem MIGHT be caused by something entirely different... its possible that the fraps graphs have nothing to do with this... but it must be a pretty huge coincidence then that the fraphs graph results match the perceptive problems with multi gpu... :P

One would think after AMD acquired ATI, the 3870 X2 basically flopped, that ATI would have at the least put an HT link on the table for an upcoming generation. With the success of the 4870 X2, and the currently unchallenged performance of the 5970, I fear that ATI may have gotten lazy in this regard. I mean, its very plausible that they could do HT with a 6000 series card, but not likely.

yeah, instead they added sideport, and never used it...

[NickF]

yeah, instead they added sideport, and never used it...

I thought sideport was just a pr stunt for low-end gpus

[Blacky]

Great article! and some people though this was "fixed" over a year ago! though I had an G92 SLi as well and the MS issues were awful, though I gave SLi another chance with GT200 and the fps were smoother and not bothersome!

[KingOfsorroW]

When I had 3x285 in tri sli there was no micro or macro stuttering - I can tell you that for sure. It's just that some games work better on a single card... I cannot tell anything about ati though for now - I simply haven't tried crossfire yet. But they say it scales marvelousely with no stuttering. this microstutter myth has been busted a numerous times already.

[saaya]

When I had 3x285 in tri sli there was no micro or macro stuttering - I can tell you that for sure. It's just that some games work better on a single card... I cannot tell anything about ati though for now - I simply haven't tried crossfire yet. But they say it scales marvelousely with no stuttering. this microstutter myth has been busted a numerous times already.

rofl, what? hahahah
man, have you ever benched vantage with your cards?
it stutters so bad with tri sli it almost gives me headaches...
you must be BLIND not to notice any stuttering with tri sli...
unless of course you play at normal resolutions and have so high fps that the stuttering doesnt matter cause it jumps from 180fps to 100fps

[Shadov]

rofl, what? hahahah
man, have you ever benched vantage with your cards?
it stutters so bad with tri sli it almost gives me headaches...
you must be BLIND not to notice any stuttering with tri sli...
unless of course you play at normal resolutions and have so high fps that the stuttering doesnt matter cause it jumps from 180fps to 100fps

I disagree saaya, some titles are more prone to this effect then others.

That way you CAN NOT make it a general rule for all titles and just support it by a fraps graph.

[trinibwoy]

Actually a fraps graph is far more useful evidence than people claiming to see / not see something. Microstuttering is a well understood phenomenon, people just stick their heads in the sand to avoid acknowledging that their multi-GPU setups are pumping out useless numbers.

[Lokinhow]

oh maan...

multi GPU is around for yeeears, but it doesn't work properly yet


ok, AFR is crap, so what would be the "right way" to do multi GPU?

[Final8ty]

Actually a fraps graph is far more useful evidence than people claiming to see / not see something. Microstuttering is a well understood phenomenon, people just stick their heads in the sand to avoid acknowledging that their multi-GPU setups are pumping out useless numbers.

You cant claim that all people notice Microstuttering & are ignoring it or lyingly.
They either notice it or they don't.

[Decami]

When I had 3x285 in tri sli there was no micro or macro stuttering - I can tell you that for sure. It's just that some games work better on a single card... I cannot tell anything about ati though for now - I simply haven't tried crossfire yet. But they say it scales marvelousely with no stuttering. this microstutter myth has been busted a numerous times already.

give me one link where it has been busted, the links don't exist, cause micro stuttering is a true problem. If you have heard it has been busted many times, thats straight here say bull crap from people in denial that there massive money spent isn't 100% reliable. I attempted SLI myself with G92 and it was a waste experience. and I am not shure why you said macro stuttering, thats not a term. I hear (no fact on this) that since GT200 with SLI this has been less of a problem, and more of a problem with Xfire, either way its still an issue. I'm awaiting an advancement in multi GPU options. hence hydra.

This is a big reason for me I am awaiting hydra. I thought the scaling would be more, but oh well, if it results in no micro stutter and no input lag issues, I'm game.

You cant claim that all people notice Microstuttering & are ignoring it or lyingly.
They either notice it or they don't.

this is entirely true as well though. same thing as FPS and the people that say you cant see above 40 fps, which is total non sense.

[Eastcoasthandle]

Folks, just don't read the OP but click on the link and read all the graphs. There you will find a trend where fluctuates are more pronounced at 1920 then at 1680. Thus it's easy to see that there is some sort of frame rate induced problem more so then anything else. It's clear to me that there is some sort of variant added to all this that for the most part not mentioned. For example, enhanced settings from CCC, lower 3d clock rate, etc.

Because this is more of a low frame rate issue the graphs have no meaning. Also, from what other's have said, there is no control results and no frame rate results. However, the current graphs used when providing both 1680 and 1920 per game clearly show low frame rates.

[Decami]

....ok and the fact they might be using more graphic enhancements still doesn't make the graphs mean nothing though. So what if they are, it means its still a problem with more enhancements, which, means its still a problem.

poor logic.

[Eastcoasthandle]

....ok and the fact they might be using more graphic enhancements still doesn't make the graphs mean nothing though. So what if they are, it means its still a problem with more enhancements, which, means its still a problem.

poor logic.

The poor logic is found when trying to reinvent a problem when the former is clearly flawed.

[Sentential]

You cant claim that all people notice Microstuttering & are ignoring it or lyingly.
They either notice it or they don't.

I agree with you 100%

Coming from a heavy SLI user microstuttering is a major problem; Its happened across multiple setups and multiple monitors. I have personally seen it on:

6800 GS
GTX260 SLI
GTX260 192SP SLI
8800GT (G92) SLI
8800GTS (G80) SLI
HD3850 Xfire

Sometimes its noticable other times its not; I personally think its linked to the saturation of the PCI-E bus; not neccessarly on the GFX end of things in terms of lane saturation but data saturation at the southbridge.

The worst games for microstuttering in my personal experience have been MMOs and games that have large open worlds that have to access heavy amounts of CPU/Mem/HDD loads.

Enclosed areas the microstuttering goes entirely away; however the moment you step outside its really pronounced especially as I get near the 40FPS marker.

Given my experiences; Once Fermi is released I plan on shifting my focus from SLI on my main gaming rig to:

1 GFX card
1 PhysX card
1 PCI-E Raid controller

(for the reasons mentioned above)