new AMD Catalyst driver from Alpha Micro stuttering

[bystander36]

I'd say you're gonna crack past 2GB at 4K, and if you add MSAA, way more than 2GB. It might still run fine (who knows what the driver is doing in the background to memory manage), but I wouldn't want to have my VRAM used at 99%... it's just not something that should happen on a $1K piece of hardware.

Titan will not risk bottlenecks, not now, and not for another 3 years IMO.

4k resolution is like 2 screen Eyefinity, screen size wise. 3-way 2D surround is usually never a problem with 2GB, so I wouldn't expect a major problem. But we'll have to wait and see.

 

[bystander36]

"Runt frames" is a nVidia term
I believe the algorithm has a general aspect to it but will need game specific tweaking for better perfomance

The term came from Nvidia, the definition did not. THG is using the default definition in their software, which is frames less than 21 rows, Pcper is using a different definition of frames less than 20% the average of the last 10 frames.

 

[ICDP]

And yet when I was running 2 7970s on Crossfire, neither vsync nor frame rate caps (nor any other of the myriad things I tried) ever 'fixed' micro stutter. Sure, they mitigated it noticeably. But that is not the same thing as saying that I was satisfied with the outcome. The smoothness was still orders of magnitude behind what a single 7970 was delivering.

Just because you couldn't get it to work or where unhappy with the results doesn't mean it doesn't work at all. Here are the results of Frame latency tests taken with different vsync/FPS caps.

Tomb Raider
Summit Forest level
2560x1600
Ultra Preset with Sweet FX enabled and Shadows set to ultra.
Crossfire 7950s @1100/1450

Vsync Off = unplayable. FPS actually looked nowhere near the 69 average reported.



Vsync off and FPS capped at 60. As can be seen the frame latency is excellent now, no micro stutters observed at all. Much better but the tearing was horrendous and for me unplayable because of it. One other major benefit was that the minimums have increased by ~20% and were so infrequent as to be unnoticeable.



Vsync on and FPS cap set to 59. I used the TR vsync setting to use triple buffering.
Absolutely perfect, no stutter and no tearing whatsoever. Utterly smooth throughout the test. I didn't observe any massive increase in latency. Once again the minimums are much higher than the vsync off test.


So as can be seen using these three tests it is very possible to eliminate microstutter, even without using vsync if you don't like the extra latency. I can assure you that the two results with FPS caps and or vsync truly run far faster and smoother than a single HD7950 even if it is overclocked to 1200/1700. So while it looks like I am giving up 15% performance I am actually getting smoother FPS with higher minimums. A far smoother and stutter free experience with some simple tweaks using RadeonPro.

 

[Black Octagon]

I did all of that, and again, while it mitigated the microstutter, overall smoothness was still no where near what a single 7970 gave me, and is still giving me.

What driver version did you use for those graphs, and is your software doing more of a FRAPS or more of a FCAT analysis in terms of the stage of the rendering pipeline that it is capturing?

P.S., I was gaming at 120hz, so vsync was even less useful than usual

 

[VulgarDisplay]

I did all of that, and again, while it mitigated the microstutter, overall smoothness was still no where near what a single 7970 gave me, and is still giving me.

What driver version did you use for those graphs, and is your software doing more of a FRAPS or more of a FCAT analysis in terms of the stage of the rendering pipeline that it is capturing?

P.S., I was gaming at 120hz, so vsync was even less useful than usual

Did you have over 120fps?

 

[ICDP]

I did all of that, and again, while it mitigated the microstutter, overall smoothness was still no where near what a single 7970 gave me, and is still giving me.

What driver version did you use for those graphs, and is your software doing more of a FRAPS or more of a FCAT analysis in terms of the stage of the rendering pipeline that it is capturing?

P.S., I was gaming at 120hz, so vsync was even less useful than usual

The driver version doesn't matter as I get the same results on older drivers. The results are from FRAPS/FRAFS and if the line is flat it = no stutter at all. I think the issue is you are running 120Hz and as most games don't reach 120+ FPS than using vsync wont help.

For example in Crysis 3 at 2560x1600 very high, I get an average FPS below 60Hz so I have to find my avarage FPS and use an FPS cap below that average. So if I get 47 average FPS I would set a 40 FPS cap which smooths performance considerably. As I mentioned in my test results above this has the advatnage of raising the minimums as well.

I am happy doing these tweaks but I can understand they aren't ideal for an out of the box experience. SLI does work much better than CF out of the box but it most certainly isn't stutter free. I personally hate screen tearing and even with GTX680s in SLI I was using FPS caps and vsync to elminiate tearing and MS. Because I alsways use vsync my experience of CF 7950s was/is no different at all compared to GTX680s in SLI, but the difference in cost was massive, hence the decision to change.

 

[UaVaj]

Input lag will go up, just like with any frame metering. That's why AMD said they will put in a slider to give the user choice as to how much smoothness versus input lag they want.

x2

neutering was the fix. good for amd. better neuter than broke.

 

[Haserath]

Does the prototype driver also include the memory rewrite?

 

[Will Robinson]

No,that will be released later this summer as far as I'm aware.

 

[ICDP]

Does the prototype driver also include the memory rewrite?

The prototype driver is also the memory rewrite, all recent drivers are simply individual game profile fixes. The prototype/memory rewrite driver will be released in the summer, though betas are usually out sooner.

I would guess we could see a beta of the prototype drivers in a month or so. Just a guess and I could be wrong of course.

 

[Black Octagon]

The driver version doesn't matter as I get the same results on older drivers. The results are from FRAPS/FRAFS and if the line is flat it = no stutter at all. I think the issue is you are running 120Hz and as most games don't reach 120+ FPS than using vsync wont help.

For example in Crysis 3 at 2560x1600 very high, I get an average FPS below 60Hz so I have to find my avarage FPS and use an FPS cap below that average. So if I get 47 average FPS I would set a 40 FPS cap.

I asked about drivers simply because this was at the very start of the year when we first moved beyond 12.xx versions and the 'frame latency' issue for DX9 was being targeted for the first time.

Yes, I was getting over 120fps as this was with 2 7970s and on not too demanding games. To be sure, vsync, frame rate caps, and experimenting with triple buffering, the problem could be noticeably reduced but never brought down to single Gpu levels.

It also varied from game to game. In Deus Ex: HR it was very minor and bordered on bearable. For Skyrim the result was always deeply unsatisfactory. As in, 75fps on a single 7970 felt smoother than 2 7970s (which, with vsync off, could drive a solid 175fps).

The other thing: I asked whether you're using FRAPS for your graphs because of Anandtech's recent interview with AMD, in which both parties seemed to agree that there was a need to measure frame times later in the rendering pipeline. Even if a FRAPS frame time graph is particularly bumpy (or smooth) they said that this does not necessarily correspond to what the end user actually sees. Hence the interest in FCAT right now.

Maybe different people just have different levels of sensitivity to microstutter...

 

[SirPauly]

Maybe different people just have different levels of sensitivity to microstutter...

Indeed and why I prefer the tangible data.

 

[ICDP]

Maybe different people just have different levels of sensitivity to microstutter...

Indeed and why I prefer the tangible data.

Which is why I provided 3 frametime graphs taken using different vsync FPS combinations with the express intent to show how these settings affect micro stutter. I mean is't not like I simply stated an opinion without hard information/data to back it up.

You can't claim crossfire stutter can't be fixed and when presented with hard data to show otherwise simply change your claim to, "maybe we have different sensitivity". Of course we have different sensitivity, but it is impossible for us to test which one of us is more sensitive. Likewise you can't choose to ignore my tangible data by stating "I prefer tangible data", that's called irony.

 

[SirPauly]

I don't disagree with you only agreed on the subjective and tolerance level may differ with MS.

 

[bystander36]

Microstutter is known to cause simulator sickness for some. I get nauseated the more and longer I experience microstuttering.

 

[ICDP]

I don't disagree with you only agreed on the subjective and tolerance level may differ with MS.

My apologies, of course that is true. My point is that at some point MS is indeed to all intents and purposes, gone and I posted graphs to back that claim up. Now many may counter claim that maybe I have more tolerance and I only think it is fixed. That is why I called it ironic to ask for tangible data only to dismiss that tangible data with a flipant, "Maybe different people just have different levels of sensitivity to microstutter".

 

[dguy6789]

still, it's hard to understand how so many users have been happy with CF without this for so long...

Because it's not a big issue in reality

 

[bystander36]

My apologies, of course that is true. My point is that at some point MS is indeed to all intents and purposes, gone and I posted graphs to back that claim up. Now many may counter claim that maybe I have more tolerance and I only think it is fixed. That is why I called it ironic to ask for tangible data only to dismiss that tangible data with a flipant, "Maybe different people just have different levels of sensitivity to microstutter".

According to THG's review, which has gamers test their systems, without knowing which was which, people still can tell the difference. Though they were not using the prototype driver. Later, when they tested the prototype driver, they noted an improvement.

So we wait for the prototype to be released and compare again. They should improve it more by then. In the THG and Pcper articles, it still had a lot more variance than Nvidia's frame metering, but definitely better than before.

 

[railven]

I've been enjoying all my games lately, with zero issues. If I have the highes tolerance level or magical 7970s that scale in games reviewers claim zero/negative scaling - I dunno.

But right now, I have zero complaints outside of WoW.

 

[bystander36]

Because it's not a big issue in reality

It's more of an ignorance is bliss type of deal. It is not so bad you notice it on its own most the time, but bad enough you can see a definitely difference in side by side tests. Though those with weaker cards in CF, have come to the boards complaining often, more so than those with strong enough CF setups to have reasonably smooth game play on a single card.

It is more of a, "damn! I am not getting the improvements I paid for," type of deal

 

[parvadomus]

Well, runt frames are a metric invented by Pcper. They give a lot of leeway, as it is defined as less than 20% of the average frame size from the previous 10 images. They could bump that up to 50% and call that a runt frame, which would make things look vastly different. They could go with THG's current definition of 20 frames or less, which would look different again.

The point is, evenly spaced out frames creates smoothness. At what point someone notices a difference is a little up in the air, and may vary, but the variance of frames is clearly a metric that matters too.

The end result is the test gamers choose SLI over it, so while it is improving, it isn't quite as good yet. That is the prototype driver, not due to be released for a few months. That said, by then it might be better. It should. The one thing that worries me is they are doing it on a per game basis, rather than a general algorithm.

I dont agree with this. Evenly spaced frames does not necessary means smoothness. Think that in any game, there is a step of simulation which may or may not be rendered. If it is not rendered by any cause, frame metering may in the worst case try to show frames at an even rate when in fact it should have not do it.
If that was a character animation, it would show even more stutter than without frame mettering.

l  l l l    l  l   l       < Frames sent to GPU at some game time
 x  x  x  x  x  x  x    < Frame time captured by fraps, when using a frame mettering technique evenly spacing them)
 c  c c c    c  c   c    < Correct frame time (FRAPS will show this like crap, but really is the smoother one)

As you can see, this is not as simple as you think. I think FCAT is a good way to measure FPS and runt frames, and smoothness to some point. FRAPS is the same, when you see too much variance its clear that game is not running smooth, but there is one threshold of variance from where you cant say for sure if x setup is smoother than y setup.
For now, I cant say that 7990 is worse or better (in smoothness) than GTX690 with the prototype driver, but if its throwing more FPS I would take it over GTX690.

 

[bystander36]

I dont agree with this. Evenly spaced frames does not necessary means smoothness. Think that in any game, there is a step of simulation which may or may not be rendered. If it is not rendered by any cause, frame metering may in the worst case try to show frames at an even rate when in fact it should have not do it.
If that was a character animation, it would show even more stutter than without frame mettering.

l  l l l    l  l   l       < Frames sent to GPU at some game time
 x  x  x  x  x  x  x    < Frame time captured by fraps, when using a frame mettering technique evenly spacing them)
 c  c c c    c  c   c    < Correct frame time (FRAPS will show this like crap, but really is the smoother one)

As you can see, this is not as simple as you think. I think FCAT is a good way to measure FPS and runt frames, and smoothness to some point. FRAPS is the same, when you see too much variance its clear that game is not running smooth, but there is one threshold of variance from where you cant say for sure if x setup is smoother than y setup.
For now, I cant say that 7990 is worse or better (in smoothness) than GTX690 with the prototype driver, but if its throwing more FPS I would take it over GTX690.

I understand that even delivery of frames is not the only thing that matters. I do understand the smooth delivery of simulation also matters. I even said right there that high FPS matters too.

The reality is, if you evenly space out the frames, the simulation also gets spaced out for any following frames. Though the corrected frame may be slightly off, depending on how much correcting they do.

That does not change that lack of variance means smoothness more so than high FPS.

If the variance didn't matter, then the gamer test group would not have picked SLI as looking smoother to them. Unless there is another factor that is also better on Nvidia SLI not tested.
http://www.tomshardware.com/reviews/radeon-hd-7990-review-benchmark,3486-18.html

 

[parvadomus]

If the variance didn't matter, then the gamer test group would not have picked SLI as looking smoother to them. Unless there is another factor that is also better on Nvidia SLI not tested.
http://www.tomshardware.com/reviews/radeon-hd-7990-review-benchmark,3486-18.html

But they picked SLI cause they draw conclusions from the 13.5 driver and not the prototype one. I would need to test myself both setups to see which one feel smoother, or watch a video of both setups running the same thing recorded with a high speed camera side by side.

 

[bystander36]

But they picked SLI cause they draw conclusions from the 13.5 driver and not the prototype one. I would need to test myself both setups to see which one feel smoother, or watch a video of both setups running the same thing recorded with a high speed camera side by side.

But the prototype driver adds time to the frames, which you were discrediting. If the prototype does help (the same gamers said it did btw), that backs up my theory of evenly distributed frames helps smoothness more so than observable FPS (though only to some extent).

I didn't bring up the prototype driver, because you were trying to discredit that type of fix.

 

[BrightCandle]

Bear in mind guys that frame metering only needs to delay a tiny amount most of the time. Once you have delayed one card enough such that it is offset by 8ms (@60fps) then you need only to move tiny amounts if the cards start to drift back together again. There should be no appreciable delay caused by frame metering except for one or two frames at the beginning. All other adjustments should be minor.