[AnandTech] AMD Comments on GPU Stuttering, Offers Driver Roadmap

[zalbard]

Ultimately AMD believes that it?s to the benefit of buyers and journalists alike to better understand just what is happening, why it?s happening, and just what the most common tools can and are measuring.

What follows is based on our meeting with some of AMD's graphics hardware and driver architects, where they went into depth in all of these issues. In the following pages we?ll get into a high-level explanation of how the Windows rendering pipeline works, why this leads to single-GPU issues, why this leads to multi-GPU issues, and what various tools can measure and see in the rendering process.

More HERE

[Rollo]

I sold one of my 7970s and went back to 680Sli in my main box due to this issue, and others. (E.G. groups of rocks flashing on an off in Crysis 3)

NVIDIA's drivers are just better, especially at multi GPU.

On the other hand, I stand by my earlier statements that the 7970s are the best deal I've ever gotten for a high end card. The XFX Black Edition I'm using in my second rig was $360 AMIR and came with two games I would have paid $50/ea for: Bioshock Infinite and Crysis 3. A card with comparable performance to the GTX680 for $260 is pretty freaking amazing in this era of $1000 high end cards.

[cobra_kai]

Good article and it's nice to see AMD admitting the problem and working hard to fix it. The information about FRAPs and its limitations is very interesting, hopefully we see new and better tools come out soon for measuring this kind of performance.

[Lanek]

They could be start by this :

HT ON ( core I7 + HT on )

HT OFF ( Core I7 without HT )


I dont tell you too what impact have core parking on some games . like BF3 under windows 7 / 8 .

A second solution for overclocker, is something we use in general naturally .. remove all "clocks" variation, CE states etc. 4.5-5ghz full speed all the time on all cores.

I will maybe advance myself but it look like some frametimes issue look simply linked to cores speed variation when HT is on / core parking / and the way windows handle threads in this case ( where you rarely goes over 25% usage on a core ). HT look to increase this situation. ( this dont mean there's no problem, but it increase the problem ) .

But reading this article it look like it really shown what i say when using fraps, you will not notice more or less "stutter", but fraps will record the input before enter the gpu, Direct3D and render pipeline and see this hicups of the HT and windows thread schedule.

[Kallenator]

Why was Fraps a popular tool for measuring FPS to begin with? Would not a more optimal tool for frame grabbing like Dxtory be a more suitable choice?

http://dxtory.com/v2-home-en.html


It's incredibly cool that AMD is so honest about this issue.

[Lanek]

I need agree with Anandtech for GPUview still.. i have use GPUPerfview ( AMD tool ) and Nvidia one + GPUview by the past, but im just a bit lost when it come to read correctly the data.

[flopper]

excellent article.

[Hell Hound]

If windows is part of the problem push for open gl development and linux support,part of fixing the problem would be to remove windows limitations.

[Lanek]

If windows is part of the problem push for open gl development and linux support,part of fixing the problem would be to remove windows limitations.

When we see the difference from a a game to another, i think the problem is more how developpers code their engine on this side.

[Hell Hound]

True but windows makes you work harder, the pre rendered frames cap plus

Complicating all of this is the fact that Windows is not a real-time operating system, meaning that Windows cannot guarantee that it will execute any given command within a certain period of time. Essentially, Windows will get around to it when it can. In order to achieve the kind of millisecond level response time that applications and drivers need to ensure smoothness, Windows has to be over provisioned to make sure it has excess resources. Consequently this is part of the reason for why the context queue exists in the first place, to serve as a buffer for when Windows can?t get the next frame passed down quickly enough.

[Lanek]

True but windows makes you work harder, the pre rendered frames cap plus

Agree. still we will need a complete pictures on Linux based system, different problem will ofc enter the equation. could be interessant.

Personally when i see all the "shot in the dark", aleatory results, differences from a game to another, from a cpu / system to another, some results look a little bit overstated by some peoples.

[Final8ty]

Hopefully they get it sorted by the time the 8 series comes out, i just seem to have a nose for skipping the problems and jumping in at the right time, a big tech change, avoid the first round.

[cobra_kai]

Common linux distros like Ubuntu, Debian etc. are not real time either.

[Hell Hound]

Amd supports Red Hat which seems to have real time computing.

[Hornet331]

Sorry but rtc OSes are not feasible for consumers. So blaming windows just for the sake of it is kinda of a joke.
Desktop computing is not mission critical and never will be aka the average users isn't interested in the possibility that one task is done under certain latency limitations, to make the one task run as low latency as possible, but many task are done in a fashion that work without noticeable perceived latency.

[Rollo]

It's incredibly cool that AMD is so honest about this issue.

Got to disagree there. How many years has it been going on and now they come clean when they have the problem largely fixed for single cards and a fix 4 months out for CFx?

I read an interview with Tom Peterson at NVIDIA the other day where he said they've been optimizing for smooth frame delivery for a couple years now. (likely why Scott Wasson at Tech Report looked into it in the first place)

So 2 years after their competitor did this, and a year and a half after the tech review sites started reporting on it, "they are "honest" and fixed it".

I think it would be "cooler" if they would have a. thought of it on their own like NVIDIA did b. at least started working on it a year and a half ago c. not released the "Never Settle" drivers which were widely reported to increase the problem.

I'm very used to SLi in my boxes and largely "it just works". CFx with 7970s was not that kind of experience for me so now I'm down to one 7970.

I'd say this to AMD:

I like your single card performance and am glad this is fixed.

I very much dislike NVIDIA charging $1000. for cards that depreciate very rapidly.

I'd be happy to go all AMD if they can get the 8000 series out the door for $700 or less. I'd buy at least two, if not three.

[Hell Hound]

Those things were mention in the article because there is a negative impact on how software acts in that environment,and was stated that more resources are needed because of it.So how is that blaming just to do it.Also there is plenty of capable people who want a gaming alternative to windows/osx.Dedicated gaming os like 360/ps3 will devastate consoles and may make them obsolete.

[Sparky]

Oh, so it was nvidia claiming to be "smoother" this time around?

(I wonder how many people will get the admittedly bad joke )

Anyway, I can't say I've really noticed much jitter or whatever with either my 4870 or my 7950

[zalbard]

(I wonder how many people will get the admittedly bad joke )

Haha, do you mean Anand's article about Phenom II?

[QuadDamage]

Haha, do you mean Anand's article about Phenom II?

Umm I don't think so.

[madcho]

Lanek >

Most gamers disable CE when they overclock. It would be nice to see CnQ effects when enable or disable on shuttering & micro shuttering.
And there is a patch for going windows in real time os. But never tryed. The problem with that, if your app crash, all sys crash too
and you should loose some performance, and comfort while using windows desk, apps switch should be less pleasant.

But i agree. Very old OS were real time or having a real time mode for gaming. And it was more sweet. They increased latency a lot with new rendering techs. I would prefer play an old carmageddon than a BF3 ... more fun, more fast ...

http://en.wikipedia.org/wiki/Carmageddon

old time ...

[Stukov]

Haha, do you mean Anand's article about Phenom II?

Ironically compared to my C2D (which I loved) the Phenom II while slower at some games, was much smoother, as is my IB (while also being faster). I'm pretty sure this had to do with HT and IMC at the time despite C2D having better CPU cores.

[Kallenator]

Got to disagree there. ...Snip

Companies of this magnitude are generally fairly closed-circuit in terms of releasing information about issues or even openly discussing them. You can of course argue that a fix is long over due or that bad choices had been made, this however does not change the fact that people have not been complaining about this in single-card setups which is probably because of it's negligible factor.
As far as dual-card setups are concerned I do not qualify to comment as I have very little experience here and the only negative factor I remember is the neglect of working profiles and/or support in games.

[BenchZowner]

They could be start by this :

HT ON ( core I7 + HT on )

HT OFF ( Core I7 without HT )

Frametimes testing - Air Superiority - B. Flats - 3770K @ 4GHz - GTX 680 @ 1181MHz - 1080p 2xMSAA 16xAF 2xTrMSAA - HT ON vs HT OFF



My testing obviously

[Lanek]

I know this difference from system.. Its why use fraps is so aleatory ..


Hitman absolution: Gameplay maps ( not benchmark )

Chinatown 1920x1080p .. the highest possible settings in game, but i dont use a single card, in my case, i use 2x 7970 @ 1050mhz with AFR.

This is the type of result you see on sites with a single HD 7970 GHZ ( techreport, HH etc )..


Now this is my result on the same game and same place ( this come from an old test i have done 1 month ago , maybe even more )
Catalyst 13.2 beta - 1920x1080P, all Maxed, using the exact same method ( i have scrupulously follow the HH frame times test in the game ) but with a large difference, who should bring extremely bad used, i use Crossfire enabled . no tricks on HT, no Coreparking disabled, ( CPU is just OC to a standard 4.5 - 5ghz if remember well ) Thoses tests have been conducted in January before 13.3 beta catalyst for Hitman frame fix.



This is with CFX and AFR ( who commit this alternate frame rendering time ( but not frequencies ).. The result on MY system, are better of what is reported on sites with a single card ? ( note, there's no monitor yet who can show this variation yet, or only on " extreme case" , they have an input latency higher of the variation time between frame render with their hardware ( 8-9ms in best case, 16ms on average display monitor / the 144hz Asus27 " TN is around 6ms in best case )

the average frametimes is of 7.1ms .. the 997th percentile is at 12ms . I prey anybody to imagine been able using a monitor today to see a stuttering following this graph.. this is just impossible ( in this limited case )