Nvidia drivers give lower performance than AMD's due to high CPU overhead/inefficient

[Skurge]

I am saying that they show reduced CPU performance in a 3dMark11 physics test. But when was the last time you ran hardcore physics effects on your CPU? The whole point of testing driver overhead is to test CPU dependence in real world games. I don't disagree that NV's cards have steeper CPU requirements. I just think their example is meaningless for drawing any real world conclusions for actual games. The article I linked on the other hand shows exactly how the driver overhead impacts performance in favour of AMD.

Ahh. I see now, you are saying that article has no real world relevance.

 

[notty22]

The article is also flawed because SLI was hacked to work in that driver version/ article. Unfortunately 3dmark 11 launched without sli support. The physics test, and the combined test scored lower when sli was 'hack forced' . On my rig by over 50%.
You can see thats fixed just compare combined score results from the top gtx 480/580 at futuremarks web scores and the authors.

 

[nitromullet]

No. AMD drivers are better, many studies have proven this without a shadow of doubt.

I've been on AT since 2004, have been a video card enthusiast the entire time, and I do not recall there ever having been a single study that has definitely answered the question of which drivers were better.

Perhaps you could share a few links to these "many" studies.

 

[Grooveriding]

The article is also flawed because SLI was hacked to work in that driver version/ article. Unfortunately 3dmark 11 launched without sli support. The physics test, and the combined test scored lower when sli was 'hack forced' . On my rig by over 50%.
You can see thats fixed just compare combined score results from the top gtx 480/580 at futuremarks web scores and the authors.

The article is fine, this is a known issue. I used the hack and the latest drivers, saw a few 100 points diference. You may have a system configuration issue or mis-applied the hack.

Doing comparisons to the top futuremark scores is irrelevant to this discussion. Take some time to look at systems using the same setup and the huge difference in scores you see. My system gets about 13.7k, I see the same setups get as low as 11k and nearly 15k.

This issue is with Nvidia's drivers having high cpu overhead which reduces their performance relative to AMD's drivers and cards. Not Futuemark benchmarking, you can replicate the issue using other games as well.

 

[notty22]

From the OP article.

For the record, current nVidia drivers do not support SLI for 3DMark 11, and nor does the standard procedure of creating an application-specific SLI profile work.

However, it is straightforward to get SLI in 3DMark 11 to work if you follow these instructions using nVidia-Inspector posted by LethalRise750 on Overclock.net and it does work like a charm. This is not really the core part of this article but at the same time it is necessary for reproducing some of the results so we’ll briefly cover the procedure. Moreover, we have some interesting findings about

how the choice of graphics card can hugely impact the Physics benchmark results - which supposedly is a CPU test.

The graphics card which is now running with a hacked driver.

 

[BFG10K]

Perhaps you could share a few links to these "many" studies.

Here's one: http://www.rage3d.com/articles/catalyst_2011_image_quality_investigation/index.php?p=3

 

[nitromullet]

Here's one: http://www.rage3d.com/articles/catalyst_2011_image_quality_investigation/index.php?p=3

I wouldn't consider this exactly conclusive...

AMD commissioned a study on driver performance.

I would be equally distrusting of an NVIDIA funded test.

Even rage3d isn't too solid on making any quick judgements based on this data:

So what does this tell us? Mostly that AMD performs strenuous WHQL certification testing, and can pass the Microsoft tests, which was kind of implied by the WHQL certification. Nvidia's failures all occurred on the GTS 250 card, which failed one of the twelve passes using the 258.96 tests (test failed to complete), and two of the twelve passes using the 197.45 driver (once with a BSOD).

We are lacking some data to form a conclusive opinion about the results - were all the cards retail box samples? Were they provided by AMD or did the testing company purchase them themselves? Was the test system enclosed with normal system cooling or open on a bench? Was the power supply a high quality unit, and connected to a line conditioner?

I would have hoped they tested more than one card from each to determine overall driver stability. If they did use multiple cards and the same card failed, perhaps it's a hardware issue. Also, isn't the fact that drivers are WHQL certified imply they can pass these tests?

Honestly, I'd trust you and apoppin more if ABT ran stability tests on the ever growing stable of cards apoppin has than I do some AMD funded test. Although for stability testing he might want to break out a can of air duster on some of those cards.

 

[Wreckage]

Its strange how Nvidia gpu's are all part of the
3DMark 11 Top 20 (Performance preset)


AMD gpu's / cpu's are nowhere to be seen. Seems like AMD needs to do some work .

Don't use facts to shoot holes in their fallacy, it just angers them more.

 

[llee]

http://www.guru3d.com/news/nvidia-grabs-71-of-discrete-graphics-card-market/

NVIDIA grabs 71% of discrete graphics card market

Hmmm... it does not take a physics professor to do the math on that one. :hmm:

+1

 

[apoppin]

Honestly, I'd trust you and apoppin more if ABT ran stability tests on the ever growing stable of cards apoppin has than I do some AMD funded test. Although for stability testing he might want to break out a can of air duster on some of those cards.

What kind of "stability tests" do you want? It isn't like i am "sitting" on my cards. i am continually running benches (except for this weekend, i am just writing as my MB malfunctioned).
--As to dust, i live in the desert; they get the dust blown out at least once-a-month (you should have seen my CPU Cooler, it had been several months).

i just finished my evaluation of SLI vs CF with Cat 11.1a vs GF 266.58 - 32 benches. For that review, i tested GTX580/570/560/460 & GTS 450 plus GTX 560 SLI, 460 SLI and GTS 450 SLI plus HD 6970, 6950, 6870, 5870 and HD 6870 and 5870 CF). You'll see low/mid-range GPUs in Multi-GPU vs. powerful single cards and we will be able to see impressive SLI and CF scaling

There was going to be more testing this weekend but my X58 MB literally melted the PSU's 12V connectors during SLI'd GTX 560Ti testing and it went back to Gigabyte on Friday for RMA

The CF vs SLI Excel charts are done and i have to make the images and write the review; as soon as my SSD vs HDD eval is done tonight, i will start working on it. Expect it this week as soon as i finish evaluating my new Diamond USB sound card for a later review next weekend. And i am proofing two other editor's articles for publication tonight. Then there is Part two and three of the CF vs SLI evaluation and i am certain there are even more video cards coming
^_^

 

[Concillian]

Isn't it a good thing for drivers to use more CPU if they can improve GPU performance by doing so, given how non-CPU limited games tend to be at graphics settings people actually use?

I mean, in the games I play (on my AMD card) I can turn off my overclock and not notice any difference in framerates on games I actually play at the settings I actually use. I believe ABT demonstrated something similar not too long ago. For most people at the settings they use to play games, they aren't very CPU limited.

Now whether the drivers actually do improve GPU performance is another issue that I don't think is something we can necessarily assume, but the point is that one piece of data that says CPU usage is higher or lower for a given brand of video card doesn't really mean anything in and of itself.

 

[nitromullet]

What kind of "stability tests" do you want?

That is a very good question. I'm not exactly sure how one would test overall driver stability. It probably really can't be done in a holistic sense. Maybe you could look at a few apps on a few cards on a few driver releases, but what would that really tell you? Of course, if one of the drivers were just awful you'd probably notice, but if that was the case there probably wouldn't even be debate.

I was just responding to someone making the claim that "AMD drivers are better, many studies have proven this without a shadow of doubt." BFG piped up linking me to an article making reference to a study paid for by AMD that made it look like NVIDIA drivers were only 85% WHQL certified.

I trust you guys more than that source, but that being said I don't really want you to conduct any tests. I don't have any issues with my card, but I also really don't care about 3DMark either.

The bottom line is the claim is BS, and anyone with any experience with both AMD and NVIDIA cards knows it. Both drivers are mostly good, and both have issues from time to time.

 

[RussianSensation]

Ahh. I see now, you are saying that article has no real world relevance.

Precisely. The title of the thread implies that NV's cards actually perform worse [in games] than comparable AMD cards. However, not a single piece of evidence in that article supports this view. In other words, we can only conclude that AMD's driver seems to be more efficient -- but so what? How does that impact your videocard purchasing decision? We can't deduce any useful information unless actual games are tested with various CPUs/clock speeds, etc!! (see Xbitlabs' article).

For example, if you have a Q9550 @ 3.4ghz, is it better to get an HD6970 or a GTX570? We don't know. If you have a Core 2 Duo E8500, is it better to buy an HD6850 or a GTX460? We don't know. Therefore, we can only make an educated guess based on how these cards performed relative to each other when they were tested with top-of-the-line Core i7 systems. It's just a shame they didn't investigate this further.

 

[OCGuy]

Too bad I have to be scared to upgrade my drivers...

I never had a single problem with my 280 SLI setup, I cannot say the same for my 5850s. (Although it is only one game, and I am otherwise happy with the performance)


This is a sensationalist thread...still trying to figure out why the OP has an axe to grind....



Edit: Also, if I was a mod, I would make OP edit title, as it is misleading and completely inaccurate.

 

[badb0y]

http://www.guru3d.com/news/nvidia-grabs-71-of-discrete-graphics-card-market/

NVIDIA grabs 71% of discrete graphics card market

Hmmm... it does not take a physics professor to do the math on that one. :hmm:




Also add "in one test in 3DMark11"

You've been told this before but Intel holds more than 50% of the graphics market yet they were at 8.8 percent reported crashes. I think you need to go ask a statistician for help there buddy .

 

[Pantalaimon]

+1

Isn't that article from 2008?

 

[shangshang]

This is a sensationalist thread...still trying to figure out why the OP has an axe to grind....

+1. Me me me.

 

[Ben90]

Well I can't test tonight because I managed to dislodge my CPU heatsink. My main beef with the article is well, its not setup to make this kind of measurement. 3DMark 11's physics test cannot give us any useful results since faster rendering puts more overhead stress on the CPU that has nothing to do with physics.

If someone totally badass wanted to test this, I would try something more along the lines of this:

-Set system cores to 1.

-Run repeatable GPU benchmark WITH some sort of framerate limiter on. Set priority to realtime. Framerate cannot dip below limit or the test is invalid. Log framerate to check for validity.

-Run repeatable CPU benchmark while GPU benchmark is running. Hotkey CPU and GPU benchmark so their start times are repeatable. Scores from CPU determine how efficient the driver is.

 

[MrK6]

lol @ this thread. They're video cards kids, go outside.

Has it been shown in any other applications with any significance?

 

[cusideabelincoln]

That's not the same as saying NV's graphics cards are slower than AMD's because their driver is less efficient.

The article itself is very weak in trying to portray this idea. Xbitlabs produced a far more comprehensive article on this topic many months ago.

Thanks. It would be good to know if a 6850 would be better for a system with maybe a core2duo over a 460. Seeing as most review sites use heavily clocked i7s so you cant see the drivers would make a real difference.

http://alienbabeltech.com/main/?p=22167&page=13

This review is just as comprehensive. Far Cry 2 is a shining example of what can happen with this bottleneck we're investigating. With a dual core the GTX 480s performance drops below the performance level of the 5870; with a quad core the 480 is faster. Also it appears Crossfire also follows the similar trend as Nvidia drivers and appears to be even more influenced by processor speed. Apoppin didn't test SLI but I think it's safe to assume SLI would follow the same trend as well.

World In Conflict is another good example. Here the AMD card paired with the Phenom X4 is faster than the 480 paired with the Phenom. The tide changes when switching to Core i7.

This doesn't happen in every game.

The article shows that NV's driver has a higher overhead compared to AMD's in 1 test. They specifically look at "Physics" 3dMark11 test to highlight this difference. Ironically, NV graphics cards utterly dominate AMD in any game that actually uses PhysX effects (i.e., some attempt at realistic physics effects). So do you really care how good a CPU is in running physics when hardly any modern games run realistic physics effects on the CPU? In other words, if they wanted to look at CPU dependence by brand, they should have compared various CPUs in actual games.

PhysX, as is it's currently implemented, isn't even relevant to this discussion and is not physics. But that isn't the point of what the article was highlighting; you're fabricating an argument out of nothing. As currently implemented in games I'd take something like Havok over PhysX.

It would have been better if they tested different CPUs, but that doesn't make their one test invalid. The Alienbabel and your own Xbits articles verify what the OP's article is showing: When CPU resources become limited then Fermi's performance drops like a rock, and that can possibly result in lower performance than a competing AMD card. You do see a trend in the ABT article - the GTX 480 is very susceptible to CPU speed, even if the end result in some games doesn't drop its performance below that of the competition.

Precisely. The title of the thread implies that NV's cards actually perform worse [in games] than comparable AMD cards. However, not a single piece of evidence in that article supports this view. In other words, we can only conclude that AMD's driver seems to be more efficient -- but so what? How does that impact your videocard purchasing decision? We can't deduce any useful information unless actual games are tested with various CPUs/clock speeds, etc!! (see Xbitlabs' article).

You can deduce what can happen under situations. Even the Xbit and ABT articles don't show what will happen under every condition - and for that you simply have to make your best educated guess. Just consider the OPs article a data point among many, highlighting the theoretical worst case scenario. And it does exactly that.

http://www.guru3d.com/news/nvidia-gr...s-card-market/

NVIDIA grabs 71% of discrete graphics card market

Hmmm... it does not take a physics professor to do the math on that one.

May not take a physics professor to do what you are showing, but it does take someone with common sense and thoughtfulness to not post an irrelevant stat. What does the discrete graphics market have to do with that chart? Your link doesn't mean anything, because there are (were) also integrated graphics which contribute to those crashes.

But even if I do play along, look at that chart. If we isolate AMD and Nvidia, then Nvidia has 75% of the total crashes between the two. Hmm... 75% of the crashes for "71% of the discrete market" is not a favorable ratio here... it doesn't take a math professor to...

Don't use facts to shoot holes in their fallacy, it just angers them more.

Don't use red herrings to deflect away from the true nature of this discussion. Because notty's link and your approval are not relevant to the discussion, which is and always has been in this thread about how Nvidia's drivers react under CPU-resource limitations. The top 20 list does nothing to show the correlation because the users are not running their tests under processor resource limitations (they are running the fastest processors possible).

 

[Obsoleet]

I've been on AT since 2004, have been a video card enthusiast the entire time, and I do not recall there ever having been a single study that has definitely answered the question of which drivers were better.

Perhaps you could share a few links to these "many" studies.

I wouldn't consider this exactly conclusive...

You apparently want to believe that NV drivers are equal or superior, and times have changed. Some people won't ever be convinced.

1. BFG linked the main one I had in mind. http://www.rage3d.com/articles/catalyst_2011_image_quality_investigation/index.php?p=3

2. But it really kicked off in '07 in the XP to .Net transition with Vista. http://arstechnica.com/hardware/new...it-paints-picture-of-buggy-nvidia-drivers.ars

3. Nvidia has the honor of owning the most serious driver bug in the history of the entire industry, which destroyed cards. Very serious and far worse than AMD or ATI ever did as far as 2nd rate quality assurance. http://www.incgamers.com/News/21293/nvidia-19675-kills-video-cards My anecdotes are negative experiences with both AMD and Nvidia, but this certainly is the king of the hill if there was any doubt left on NV's driver support failing.
As an aside, NV does not allow customers who buy their cards to use them for Physx processors alongside Radeons. Nvidia has stated this is due to quality assurance issues, which I believe entirely. Nvidia's miserable Quality Assurance was confirmed with the 196.75 driver release linked above.

I'm sure there's more info out there regarding NV's systemic driver problems. Yet, it is a fact that the problem started trending with the launch of Vista and has continued to this day as I've illustrated with links above. In 2011, the industry leading, gold standard drivers are AMD's, as tough of a pill as it may be to swallow for many.

 

[badb0y]

In 2011, the industry leading, gold standard drivers are AMD's, as tough of a pill as it may be to swallow for many.

Lol, you better hide.

 

[Obsoleet]

It takes a while for people to break out of old mindsets and habits.. and there's definitely a financial interest out there in these ideas remaining in the minds of consumers.. but I backed up my case thoroughly.

 

[nitromullet]

In 2011, the industry leading, gold standard drivers are AMD's

MSI 5870 on 10.6

Why aren't you using the latest (from 2011) drivers yourself then, or is there something special about last June's drivers?

 

[zerocool84]

http://www.guru3d.com/news/nvidia-grabs-71-of-discrete-graphics-card-market/

NVIDIA grabs 71% of discrete graphics card market

Hmmm... it does not take a physics professor to do the math on that one. :hmm:




Also add "in one test in 3DMark11"

Well I'd say it didn't help that back then Nvidia had their fairly common Display Driver crash.