Fermi's lead over Cypress shrinks w/ new drivers

[Tempered81]

Those tests results look pretty reasonable to me.
So current DX11 video card standings:
1.The Overdog HD5970
2.Fermi GTX480
3.Cypress HD5870
4.AMD HD5850
5.Fermi GTX470
6.ATi HD5830
7.NVidia GTX465
8.AMD 5770

How's that look?^_^

looks good, but i would put the 470 above the 5850, and add the 285 below the 5850. disregarding dx11, then the 5830/465/280/275/4890 are more or less performance equals (not including the new technology). The 5970 is above everything (295/4870x2/etc), including dual card rigs up to the 5850 CF and 285 SLI.

Then you have the 4GB 5970s from 3 or 4 manufacturers that are the cream of the crop.

 

[Puffnstuff]

I don't regret buying my gtx480's one bit. Their performance increases with each new driver release and the 257.15 beta is no exception.

 

[HurleyBird]

Yeah puff, you have a sweet setup. The hefty water-cooling setup undoubtedly helps alleviate some of the unpleasantness of running 480's in SLI.

 

[tincart]

Are you serious? If we did a poll right now out of all the members of the Anandtech forums who have the 5870 and 480, I bet you'd be quite surprised to see just how many have more than one. I don't think you've cruised around forums long enough to realize what "enthusiasts" do. I don't think you know what the word "enthusiast" means when it comes to PC hardware.

I gave you an opportunity to show some proof for your grandiose claim, and you suggest that we do a poll on Anandtech - this is a clear admission that you have no evidence whatsoever beyond your opinion on the matter. You have provided absolutely no information of any relevance to back up your claim and an informal poll on one tech website would not do anything to validate it.

Further, you seem to be taking "enthusiast" to mean "someone who buys two graphics cards" in order to argue by definition. That is an extremely loose term. Many tech 'enthusiasts' don't care about graphics cards at all... We are a subset of a subset of a subset.

 

[Tempered81]

Yah, Puffnstuff. 480 SLI on watercooling is the way to go. I bet that rig screams in any game.

 

[happy medium]

BCBC2 results from 60 seconds of fraps beginning of the same level

Frames, Time (ms), Min, Max, Avg
6050, 60000, 81, 125, 100.833

FPS, 95 98 95 89 94 97 103 108 104 103 99 100 95 90 90 102 107 103 107 105 105 111 119 87 95 98 90 79 83 87 97 96 96 96 94 93 93 94 95 93 93 94 96 94 88 76 72 81 79 85 85 84 85 83 83 84 80 81 80 83

Frames, Time (ms), Min, Max, Avg
5972, 60000, 72, 124, 92.867

FPS, 95 96 100 92 91 98 97 107 105 104 102 91 89 82 110 91 90 90 91 91 89 91 90 91 89 90 89 90 103 105 104 111 107 104 103 109 103 104 106 107 105 106 104 104 104 105 104 107 111 104 104 99 115 123 122 113 101 106 109 106

Look at the end of the bench mark 80's vs 100's.

Just to let you know this was with the same 5750 card @ 1280x1024 high settings in the same computer.
I'm sure I can run the test all day and get different results using this method.

 

[HumblePie]

Errr... I don't think some people here have any idea what multiple GPU scaling is.

If a card produces 50 frames per second consistently for a given benchmark, but produces 100 frames per second in a crossfire or SLI setup then the scaling was 100% perfect. However if it produced anything less than 100 frames per second it lost some performance which may or may not be related to scaling problems. Guru3D had an article about scaling with the GTX480 and bottlenecks from the rest of the system. http://www.guru3d.com/article/geforce-gtx-480-3way-sli-review/1

Basically, anything less than max 24x15 resolution without maxed AA/AF and quality settings could have significant system bottlenecks in practically every 3D application. This also holds true for the 5870 card as well as the 480.

Also, percentage of scaling doesn't change like evolution2 was stating. Where he learned that really bad elementary math I have no clue. When talking about percentages, they remain constant with good scaling and not double.

Here's an example. A gtx480 has at a hypothetical benchmark 55 fps and the 5870 has 50 fps. This is a 10% lead for the gtx480. If the xfire/sli config had numbers reflecting a 100% scaling for both, the numbers would be 110 fps versus 100 fps. That is still a 10% lead. Now if either xfire or sli setup ran into a bottleneck or bad scaling issue then the percentage of difference would change.

From reviews I've read on multiple GPU scaling with both the 5870 and gtx480 it has been shown the scalability of both is damn near identical. The problem is not all reviews do high resolution, quality, and AA/AF benchmarks of both company's cards for comparisons. This leads to people picking numbers from different reviews that really are no where near comparable. This is because the systems are probably not the same from both reviews and as the Guru3D review shows, there is a SIGNIFICANT system bottleneck with benchmarks of either card setup in xfire or sli.


So a very valid test comparison is between single GPU setups with these cards. The test comparisons of sli versus xfire is nothing of significance. This is because both cards scale very well in most games.

 

[lifeblood]

I'm not sure any test is truly valid. Anadtech uses the exact same hardware to test the two different cards. But he doesn't use the same hardware as me, so for me and my rig the test is invalid. When I say invalid I mean that if I put the two cards in my rig, both 480 and 5870 will perform equally bad because they are both CPU limited by my rig. I won't get anywhere near the fps the review gets. The only way to get a truly valid test is to get the cards in question and test them on MY hardware. However, Anands reviews, as well as everyone else, does give me an idea of comparative performance at the settings and hardware used.

The test using actual gameplay is harder to replicate, but if the tester uses multiple runs per card and averages he should reduce the margin of error enough to claim the results are valid. If enough different reviewers/sites do similar tests and get similar results then it adds to the validity. As always, variation is expected based on different settings and hardware used.

I think the main thing for a test such as this is that you need multiple runs then averaging them to insure your getting apples to apples comparisons. Given that caveat I think actual gameplay comparisons to be more useful then timed demos. Canned demos can be coded for and both companies have in the past shown a willingness to use such dirty tricks.

BTW, I performed an unscientific poll and called the two guys I know who own 5870's and they only own 1 each. No crossfire for either of them. When I asked why they didn't get a second one they said it wasn't worth it to them. Maybe everyone else in the world who has a 5870 or 480 has two of them, but the two I know don't.

 

[GaiaHunter]

BCBC2 results from 60 seconds of fraps beginning of the same level

Frames, Time (ms), Min, Max, Avg
6050, 60000, 81, 125, 100.833

FPS, 95 98 95 89 94 97 103 108 104 103 99 100 95 90 90 102 107 103 107 105 105 111 119 87 95 98 90 79 83 87 97 96 96 96 94 93 93 94 95 93 93 94 96 94 88 76 72 81 79 85 85 84 85 83 83 84 80 81 80 83

Frames, Time (ms), Min, Max, Avg
5972, 60000, 72, 124, 92.867

FPS, 95 96 100 92 91 98 97 107 105 104 102 91 89 82 110 91 90 90 91 91 89 91 90 91 89 90 89 90 103 105 104 111 107 104 103 109 103 104 106 107 105 106 104 104 104 105 104 107 111 104 104 99 115 123 122 113 101 106 109 106

Look at the end of the bench mark 80's vs 100's.

Just to let you know this was with the same 5750 card @ 1280x1024 high settings in the same computer.
I'm sure I can run the test all day and get different results using this method.

Because the FarCry 2 canned benchmark doesn't give different results each run for example?

But lets average you run numbers:

Min 76.5, Avg 96.85, Max 124.5

So we are talking of a margin of error of +-6% (tops).

Seems reasonable.

What I don't understand is how the minimum is 81 in the first run when I see a 72. Or how the max is 125 when the max I see is 119.

 

[Puffnstuff]

Well adding the 2nd 480 to my system brought fantastic scaling with it in demanding games. In games that are not that demanding both cards scale back but when you're already running over 200fps it really doesn't matter.

 

[HumblePie]

I'm not sure any test is truly valid. Anadtech uses the exact same hardware to test the two different cards. But he doesn't use the same hardware as me, so for me and my rig the test is invalid. When I say invalid I mean that if I put the two cards in my rig, both 480 and 5870 will perform equally bad because they are both CPU limited by my rig. I won't get anywhere near the fps the review gets. The only way to get a truly valid test is to get the cards in question and test them on MY hardware. However, Anands reviews, as well as everyone else, does give me an idea of comparative performance at the settings and hardware used.

Umm, assumption does not equal validation. AT's tests are valid because they usually use a repeatable prescribed methodology for tests that try to place as much of stress on the GPU instead of the system. This way the tests reflect the performance of the video card and not the rest of the system used. Also, tests do show a wide range of resolution and usually at least 2 different quality settings. What this does show is that at lower resolutions on most games this is similar as, but not the same, as analogous to a lower end system comparison. Any video card no longer stressed by the 3D app will exhibit roughly the same performance. So if you have a lesser system and/or run at lower graphic settings then the best performing cards may not do much for you. As such, you get a very VALID result from the tests in relation to any system setup.

 

[luv2increase]

The hefty water-cooling setup undoubtedly helps alleviate some of the unpleasantness of running 480's in SLI.

It is official. You are a troll. You keep sarcastically talking about the 480 and heat. You've never had 480s in SLI. How could you possibly know anything about is being or not being unpleasant. Seriously, you need help.

 

[MrK6]

It is official. You are a troll. You keep sarcastically talking about the 480 and heat. You've never had 480s in SLI. How could you possibly know anything about is being or not being unpleasant. Seriously, you need help.

Butthurt much?

Anyway, interesting review and it just adds more credit to what I've been saying all along - get a 5850 because the high-end parts don't do enough to differentiate themselves. Seriously, $275 gets you within 10% of anything else out there.

 

[happy medium]

Because the FarCry 2 canned benchmark doesn't give different results each run for example?

But lets average you run numbers:

Min 76.5, Avg 96.85, Max 124.5

So we are talking of a margin of error of +-6% (tops).

Seems reasonable.

What I don't understand is how the minimum is 81 in the first run when I see a 72. Or how the max is 125 when the max I see is 119.

Settings: Demo(Ranch Small), 1280x1024 (85Hz), D3D10, Fixed Time Step(No), Disable Artificial Intelligence(No), Full Screen, Anti-Aliasing(None), VSync(No), Overall Quality(Optimal), Vegetation(Very High), Shading(Very High), Terrain(Very High), Geometry(Very High), Post FX(High), Texture(Very High), Shadow(Very High), Ambient(High), Hdr(Yes), Bloom(Yes), Fire(Very High), Physics(Very High), RealTrees(Very High)

Loop 1
Total Frames: 4085, Total Time: 51.00s
Average Framerate: 80.10
Max. Framerate: 107.51 (Frame:618, 6.43s)
Min. Framerate: 64.13 (Frame:1749, 21.25s)

Loop 2
Total Frames: 4103, Total Time: 51.01s
Average Framerate: 80.44
Max. Framerate: 107.43 (Frame:625, 6.50s)
Min. Framerate: 64.16 (Frame:2700, 34.22s)


Loop 3
Total Frames: 4089, Total Time: 51.00s
Average Framerate: 80.17
Max. Framerate: 107.50 (Frame:614, 6.41s)
Min. Framerate: 65.06 (Frame:1737, 20.90s)


Average Results
Average Framerate: 80.24
Max. Framerate: 107.11
Min. Framerate: 64.93

Well its seems very consistent to me. What's the margin of error here?

This method of testing is accurate and consistant.
The Fraps way is for the most part inconsistant and inaccurate.

Why change a method thats been working for years and is consistant and accurate?
Thats simple, to put ATI's cards in a better light, because Nvidia's drivers are letting the gtx 4xx line pull away in terms of performance vs ATI and making the higher price seem worth it.

 

[luv2increase]

Butthurt much?

You are the one doing the kissing LMAO

 

[GaiaHunter]

Settings: Demo(Ranch Small), 1280x1024 (85Hz), D3D10, Fixed Time Step(No), Disable Artificial Intelligence(No), Full Screen, Anti-Aliasing(None), VSync(No), Overall Quality(Optimal), Vegetation(Very High), Shading(Very High), Terrain(Very High), Geometry(Very High), Post FX(High), Texture(Very High), Shadow(Very High), Ambient(High), Hdr(Yes), Bloom(Yes), Fire(Very High), Physics(Very High), RealTrees(Very High)

Loop 1
Total Frames: 4085, Total Time: 51.00s
Average Framerate: 80.10
Max. Framerate: 107.51 (Frame:618, 6.43s)
Min. Framerate: 64.13 (Frame:1749, 21.25s)

Loop 2
Total Frames: 4103, Total Time: 51.01s
Average Framerate: 80.44
Max. Framerate: 107.43 (Frame:625, 6.50s)
Min. Framerate: 64.16 (Frame:2700, 34.22s)


Loop 3
Total Frames: 4089, Total Time: 51.00s
Average Framerate: 80.17
Max. Framerate: 107.50 (Frame:614, 6.41s)
Min. Framerate: 65.06 (Frame:1737, 20.90s)


Average Results
Average Framerate: 80.24
Max. Framerate: 107.11
Min. Framerate: 64.93

Well its seems very consistent to me. What's the margin of error here?

This method of testing is accurate and consistant.
The Fraps way is for the most part inconsistant and inaccurate.

Why change a method thats been working for years and is consistant and accurate?
Thats simple, to put ATI's cards in a better light, because Nvidia's drivers are letting the gtx 4xx line pull away in terms of performance vs ATI and making the higher price seem worth it.

And do you average 80 fps while playing farcry 2 at exactly those settings?

Because interestingly the FRAPS one actually correspond to your game play.

As you can see it can vary by 10% in exactly the same area for 60 seconds.

 

[happy medium]

And do you average 80 fps while playing farcry 2 at exactly those settings?

Whats that have to do with a test being accurate or inaccurate?

I just proved thats its inaccurate. End of story.
I have 27 more MODERN games on my drive and I'm sure the results would be the same.

When measuring performance ,I dont want something that's off by 6,8,10%.
I want something consistant and fair and most important as accurate as possible.

 

[v8envy]

These benches don't prove anything because the bench runs themselves will vary from run to run on the same card. The results can't be compared to each other on the same card much less on a different card.

Think about what you said for a moment.

If two products are tested in a way which mimics real world usage and the differences between them fall in the test's margin of error then how could anyone rightly claim one is significantly better than the other? If the differences are so small as not to be noticeable from a fraps recording then why would I care about them?

Synthetic benchmarks and drivers optimized for them are nothing new. I don't care a bit about benchmarking and scores. So if I can reasonably expect the two products to offer indistinguishable gaming experiences I can focus on value add (features like physx, cuda and eyefinity), pricing, and ergonomics.

I think we all agree that purely for benchmark comparison e-peen length contests the 5890 is the king of the hill, followed by the 480. But that's not what I'm getting at here.

 

[GaiaHunter]

Whats that have to do with a test being accurate or inaccurate?

I just proved thats its inaccurate. End of story.
I have 27 more MODERN games on my drive and I'm sure the results would be the same.

When measuring performance ,I dont want something that's off by 6,8,10%.
I want something consistant and fair and most important as accurate as possible.

Now you need to prove the canned benches correspond to actual game play.

What is the point if it is exact when it doesn't matches game play?

Also I would still like to know why the numbers reported by fraps don't correspond to the sample data.

 

[MrK6]

Whats that have to do with a test being accurate or inaccurate?

I just proved thats its inaccurate. End of story.
I have 27 more MODERN games on my drive and I'm sure the results would be the same.

When measuring performance ,I dont want something that's off by 6,8,10%.
I want something consistant and fair and most important as accurate as possible.

So you would rather have benchmark runs to compare performance and assume that they're indicative of gameplay? You're already assuming that the site isn't doing everything possible to make sure each run is as consistent as possible, so I guess that wouldn't be a stretch for you then.

 

[lifeblood]

Umm, assumption does not validation. AT's tests are valid because they usually use a repeatable prescribed methodology for tests that try to place as much of stress on the GPU instead of the system. This way the tests reflect the performance of the video card and not the rest of the system used. Also, tests do show a wide range of resolution and usually at least 2 different quality settings. What this does show is that at lower resolutions on most games this is similar as, but not the same, as analogous to a lower end system comparison. Any video card no longer stressed by the 3D app will exhibit roughly the same performance. So if you have a lesser system and/or run at lower graphic settings then the best performing cards may not do much for you. As such, you get a very VALID result from the tests in relation to any system setup.

I'm not sure what you said is contradictory to what I said. My point is that no review is the ultimate answer because their are to many variables, ie, hardware and settings. I apologize if my example was confusing, I'm not sure I'm writing it clearly.

Let me try a different approach. I think my real point is that we have to look at real world gameplay as requiring a different methodology. In physics I can set up exact experiments which allow for someone else exactly replicating it. In medicine or psychology, I can't. Each patient is different and results will vary. So for performing medical/psychological experiments I must have a large number of patients in order to get an meaningful results. Using timed demos is like a physics experiment, using real world gameplay is like the medical/psychological experiment. It requires many more test runs (patients).

 

[Lonyo]

Settings: Demo(Ranch Small), 1280x1024 (85Hz), D3D10, Fixed Time Step(No), Disable Artificial Intelligence(No), Full Screen, Anti-Aliasing(None), VSync(No), Overall Quality(Optimal), Vegetation(Very High), Shading(Very High), Terrain(Very High), Geometry(Very High), Post FX(High), Texture(Very High), Shadow(Very High), Ambient(High), Hdr(Yes), Bloom(Yes), Fire(Very High), Physics(Very High), RealTrees(Very High)
[snip]

Well its seems very consistent to me. What's the margin of error here?

This method of testing is accurate and consistant.
The Fraps way is for the most part inconsistant and inaccurate.

Why change a method thats been working for years and is consistant and accurate?
Thats simple, to put ATI's cards in a better light, because Nvidia's drivers are letting the gtx 4xx line pull away in terms of performance vs ATI and making the higher price seem worth it.

BECAUSE IT ISN'T.
It is consistent. It is NOT accurate.

http://www.hardocp.com/article/2008/02/11/benchmarking_benchmarks/6

I think we have proven that timedemo benchmark results do not represent real world gameplay. We have also seen one card enjoy a “benchmark advantage” when comparing those benchmark numbers to actual gameplay framerates. If you are looking at other sites' framerate results and thinking it is showing you a framerate, resolution, or graphical setting you should expect in a particular game, you are likely being totally mislead.
[...]
Anandtech’s results in no way suggest to the reader what the video card might actually perform like in Crysis

And part of my post from earlier in the thread:

http://www.anandtech.com/show/2233/4

During actual game play the perception was the CrossFire setups were more fluid and "seemed" faster than the single card setups although the net timedemos said otherwise. We also ran our custom timedemo and the scores were just the opposite with the 975X scoring up to 12% better.
At present, we would have to say that our Q4 test is almost completely CPU limited, and this is often not the case during actual gaming.

Just one example of where built in demos vs custom runs show differences, and gameplay shows a difference to canned benchmarks.

http://www.anandtech.com/show/2673/3
Here's built in benchmarks vs a custom benchmark in Far Cry 2 (like you want). It goes from the GTX280 beating the HD48701GB and GTX260 beating HD4870 512MB to the complete reverse when the custom timedemo is run.

 

[HumblePie]

It is official. You are a troll. You keep sarcastically talking about the 480 and heat. You've never had 480s in SLI. How could you possibly know anything about is being or not being unpleasant. Seriously, you need help.

Yep. I will also state this. The law of Thermodynamics does not change for HurleyBird just because he feels that it does.

A piece of equipment that uses a specific amount of power and produces a specific amount of heat will continue to do so regardless of the cooling solution used. The cooling setup is just a method of transfering heat from one place to another. Regardless of using water cooling or air cooling the amount of heat put into a room from a video card is CONSTANT. The rate is different but the amount is the same. All water cooling does over air cooling is transfer more heat away from the video card chip to the ambient environment faster than an air cooling solution. This allows for the video card to run consistently cooler but NOT the room. Meaning if a given room is unbearably hot with a couple of GTX480s in sli on an air cooling solution then that means a water cooled solution would have the same affect ON THE ROOM. The heat (energy) doesn't magically disappear because a person wants it to. It has to go somewhere. All water cooling does is transfer more heat faster. This means a room would get hotter sooner.

 

[v8envy]

Yep. I will also state this. The law of Thermodynamics does not change for HurleyBird just because he feels that it does.

A piece of equipment that uses a specific amount of power and produces a specific amount of heat will continue to do so regardless of the cooling solution used. The cooling setup is just a method of transfering heat from one place to another. Regardless of using water cooling or air cooling the amount of heat put into a room from a video card is CONSTANT. The rate is different but the amount is the same. All water cooling does over air cooling is transfer more heat away from the video card chip to the ambient environment faster than an air cooling solution. This allows for the video card to run consistently cooler but NOT the room. Meaning if a given room is unbearably hot with a couple of GTX480s in sli on an air cooling solution then that means a water cooled solution would have the same affect ON THE ROOM. The heat (energy) doesn't magically disappear because a person wants it to. It has to go somewhere. All water cooling does is transfer more heat faster. This means a room would get hotter sooner.

There's also the extra power to run a WC setup. Remember that power supplies aren't 100% efficient, so yet more heat being dumped into the room.

Where a WC setup has a massive advantage is noise levels. 3x120mm thick, low speed fans are worlds quieter than a low profile 80mm fan. The tone is more mellow. WC would go a long way to eliminating some of the ergonomic problems of a stock, air cooled 480 gtx SLI setup. At least in my case.

The reason I mentioned WC is: without it it'd be trivial to identify which machine is running Fermi and which one is not. That could easily influence the users' perception re: expected performance. In fact, WCing both setups and placing a few small, noisy fans in the ATI box may lead testers to perceive it as being faster since we've all been conditioned to equate noise and harshness with better performance.

 

[MagickMan]

Oh noes, HumblePie using the logix. It's like bringing a gun to a knife fight, that's dirty pool!

Seriously guys, he's right. The heat goes somewhere, and that somewhere is at your desk. It's baking your legs on a hot summer's day.