Why isnt the R580 a lot faster in Shader intensive games?

[boltneck]

I was reading through the reviews and found this quote very interesting. Its from the Tech Report review.

The Radeon X1900 cards are significantly better performers than the Radeon X1800s that they replace. I am a little bit perplexed, though, by ATI's choices here. They've tied up roughly 60 million transistors and 50 square millimeters of die space on the R580 primarily in order to add pixel shading power, but even in ShaderMark, we didn't see anything approaching three times the performance of the R520. Would this chip have performed any differently in any of our benchmarks with just 32 pixel shader units onboard? Seems like it is limited, in its present form, by other factors, such as pixel fill rate, memory bandwidth, or perhaps even register space. Who knows? Perhaps the R580 will age well compared to its competitors as developers require additional shader power and use flow control more freely. I wouldn't be shocked to see future applications take better advantage of this GPU's pixel shading prowess, but no application that we tested was able to exploit it fully. For current usage models, NVIDIA's G70 architecture appears to be more efficient, clock for clock and transistor for transistor. Here's hoping that ATI's forward-looking design does indeed have a payoff down the road.

This is something i noticed as well. Where is the huge but kicking over the GTX512 in games and apps that are clearly shader intensive? in past generations you can clearly see the advantages in fill rate and bandwidth over the previous "current crop" of cards when benchmarks are run.

In this case however the 1900's are in most cases barely eeking out 10 FPS wins against the GTX in many cases where shaders are heavy. Sometimes it even basically ties.

Why doesn’t it get a downright huge gain in edmark05 and 06 in the shader 2 and 3 tests? What about Shadermark? Why doesn’t it get a huge gain is source? Or HDR rendering?

Any ideas?

 

[Razor1]

I was reading through the reviews and found this quote very interesting. Its from the Tech Report review.



This is something i noticed as well. Where is the huge but kicking over the GTX512 in games and apps that are clearly shader intensive? in past generations you can clearly see the advantages in fill rate and bandwidth over the previous "current crop" of cards when benchmarks are run.

In this case however the 1900's are in most cases barely eeking out 10 FPS wins against the GTX in many cases where shaders are heavy. Sometimes it even basically ties.

Why doesn’t it get a downright huge gain in edmark05 and 06 in the shader 2 and 3 tests? What about Shadermark? Why doesn’t it get a huge gain is source? Or HDR rendering?

Any ideas?

Think Dave summed it up in his review very well. It will all depend if developers make use of the r580 shader structure.

ATI's decisions for the R580 chip that powers the X1900 have been interesting, to say the least, even a little bold in some ways. Even though we've looked at the X1900 fairly thoroughly in this article we've possibly yet to see it pay off fully. As shown by the pure pixel shader tests, but not fully realised by the games, this really is an architecture that has more to come, but only if games are going to going to go in the way that ATI are predicting, and quickly. For now, though, its clear that Radeon X1900 XTX offers a performance increase over X1800 XT in all situations, but it also has the price premium to match.

This is what I was weary of last last week when having a discussion with Jawed.

 

[Dave Baumann]

Look at the shader test in our review. Rightmark3D shows shader scaling much as you would expect: ealier shader levels (hence, short shader code, with a high texture to math instruction ratio) have small performance gains, increase the shader level and the length of the shader and the performance gain scales up nearly to its peak. Many games are only at the lower end of that scale in term of overall texture to ALU usage, with only a few begining to poke their way up a little.

I also suspect that Shadermark isn't the best case for ATI right now. Going by the results of Splinter Cell in the X1800 and X1900 review I believe that ATI haven't significantly altered their shader compiler post PS2.b - they've display reticence in moving away from that and I believe that a certain person who post on these forums wrote it and has since left ATI.

 

[Geo]

. . .and I believe that a certain person who post on these forums wrote it and has since left ATI.

Somebody who posts here isn't an ATI employee anymore? Somebody send me one of those secret tips, please.

 

[boltneck]

Wow that’s got to be a blow.

I hope the ATi X1800XT users like myself don’t end up completely out in the cold as far as optimizations go. Since they improved on the Distributor/scheduler logic on the R580 and stopped production on the R520 after only 3 months.. My fear now is that the 2.0b is optimized and the second gen Sm3 will get optimized, but the 520 will be out of the loop.

I wonder how much optimization work on the memory controller and scheduler will benefit the R520 going forward.

 

[boltneck]

Oh, it sounds like ATi needs to throw some major cash at a headhunter to find them some hot talent to replace what they have lost.

 

[sireric]

Look at the shader test in our review. Rightmark3D shows shader scaling much as you would expect: ealier shader levels (hence, short shader code, with a high texture to math instruction ratio) have small performance gains, increase the shader level and the length of the shader and the performance gain scales up nearly to its peak. Many games are only at the lower end of that scale in term of overall texture to ALU usage, with only a few begining to poke their way up a little.

I also suspect that Shadermark isn't the best case for ATI right now. Going by the results of Splinter Cell in the X1800 and X1900 review I believe that ATI haven't significantly altered their shader compiler post PS2.b - they've display reticence in moving away from that and I believe that a certain person who post on these forums wrote it and has since left ATI.

Honestly, the shader compiler group, for UDX, has been around for 2 years and nobody's changed there, that I know of. It's not a trivial group size. The lead is still the same. As for the driver optimization, I can't think of anyone either.

As for changes, the shader compiler is continuously improving, and getting better, but that does come at the cost of some regressions.

 

[digitalwanderer]

As for changes, the shader compiler is continuously improving, and getting better, but that does come at the cost of some regressions.

I don't understand what "regressions" means used in this context, could you elaborate just a little bit please?

 

[3dcgi]

Wow that’s got to be a blow.

I hope the ATi X1800XT users like myself don’t end up completely out in the cold as far as optimizations go. Since they improved on the Distributor/scheduler logic on the R580 and stopped production on the R520 after only 3 months.. My fear now is that the 2.0b is optimized and the second gen Sm3 will get optimized, but the 520 will be out of the loop.

I wonder how much optimization work on the memory controller and scheduler will benefit the R520 going forward.

You worry too much. :smile:

I don't think the memory controller and scheduler changed much anyway.

 

[3dcgi]

I don't understand what "regressions" means used in this context, could you elaborate just a little bit please?

Maybe he means as some things get better other situations get worse so it takes time to find the right balance.

 

[Megadrive1988]

hopefully we will see R580 and R580 derivative based cards stretch their legs through 2006 and into 2007 as optimizations happen and newer SM3.0 intensive games get released.

even if R600 launches in late 2006, I don't see it or DirectX10 / SM4.0 / WGF2.0 being very big until late 2007.

 

[digitalwanderer]

even if R600 launches in late 2006, I don't see it or DirectX10 / SM4.0 / WGF2.0 being very big until late 2007.

I don't see DirectX10 / SM4.0 / WGF2.0 being around until 2007 with all the delays...

 

[Moloch]

Isn't far cry 2 supposed to ship Q4?

 

[sireric]

I don't understand what "regressions" means used in this context, could you elaborate just a little bit please?

I mean that as you change things to improve a given situation, you might, inadvertently, cause a bug in something that used to work. Or reduce performance in some other way. Unless you regress every single app (at all resolutions and modes) against every single change, problems do crop up. Good news is that we eventually catch them and either revert the change or fix the issue. This is true of all SW, really.

 

[Reverend]

I also suspect that Shadermark isn't the best case for ATI right now. Going by the results of Splinter Cell in the X1800 and X1900 review I believe that ATI haven't significantly altered their shader compiler post PS2.b - they've display reticence in moving away from that and I believe that a certain person who post on these forums wrote it and has since left ATI.

ShaderMark tells us the performance of whatever-video-card-is-being-used for that apps specific tests. Unless the use of ShaderMark in a review can be extrapolated by the reviewer, its results are meaningless especially when games are also used in the same review. I.e. "Shadermark tells us this about this chip/video card; this is proven by the games we benched in this same review". I haven't really seen this yet.

ShaderMark provides us with some really useful info, but as IHVs continue to give us newer hw and ISVs continue to learn about such hw's shortcomings, we need to know what those ShaderMark results mean. Provided its use by review sites tells us anything meaningful. Yeah, I know... it's a tough job .

IMO, Shadermark has proven to be far more useful to IHV engineers than the hw review-reading public, as evidenced by the explanations provided by hw review sites of their ShaderMark results.

More needs to be done by hw review sites that use synthetic benchmarks like ShaderMark that also uses games in their review suite (and stress that "Games are what it's all about").

IOW, using synthetic benches and games are good but are only useful if the two can be used to explain each other's results. Hard to do, yes, but unless this is done, reviewers are only wasting their time -- and the readers! -- by benching games and synthetic apps without explaining the results of the game benches.

I understand why Dave said ShaderMark isn't the "best case" for ATI right now. What I'm saying is why this should matter to those that read his reviews that feature games prominently

PS. Don't shoot me.

 

[Moloch]

Why is the 1800XT so slow in AOE 3 btw?
The 1900XT is twice the speed up even at 2048x1536 with 4x fsaa and 8xAF
edit- and why is the GTX 512MB slower than even the 1800XT despite having a ~7:1 alu:tex ratio with fsaa?
Does the GTX take a huge FSAA hit?

 

[digitalwanderer]

I mean that as you change things to improve a given situation, you might, inadvertently, cause a bug in something that used to work. Or reduce performance in some other way. Unless you regress every single app (at all resolutions and modes) against every single change, problems do crop up. Good news is that we eventually catch them and either revert the change or fix the issue. This is true of all SW, really.

Thanks for the clarification, I get ya now.

 

[Jawed]

I'm with Reverend - synthetic shader benchmarks have told me essentially nothing in all the reviews I've ever read. If you know the architecture, you'll know more...

For example, the recently revealed mini-ALU in R3xx, R4xx, R5xx that is capable of ADD? If the synthetics were worth a damn then it would have been possible to infer this - instead, years after the introduction of R3xx we find out directly from ATI.

Jawed

 

[NocturnDragon]

I also suspect that Shadermark isn't the best case for ATI right now. Going by the results of Splinter Cell in the X1800 and X1900 review I believe that ATI haven't significantly altered their shader compiler post PS2.b - they've display reticence in moving away from that and I believe that a certain person who post on these forums wrote it and has since left ATI.

Didn't ATI hire a whole team of compiler experts some years back? (IIRC the rumors said that they were going to work on the R400 shader compiler, I guess they then switched to the xenos project, but what about now?)
I tried to google for the news but I had no luck!

 

[Demirug]

ShaderMark tells us the performance of whatever-video-card-is-being-used for that apps specific tests.

And games tells us the performance of the used video card in the specific game at the selected place with the selected settings.

We can never be sure if a game that will be released next week will show the same results. We can even not sure about games that are already released but not used in any review.

Maybe we should start asking about the reasons why one card is in the tested situation faster than an other. And the IMHO even more interesting question what it hold back of being more faster. The little information the IHV gave us about the internals of their hardware may help us to understand it better. But do we understand the tests and benchmarks that are used to show how fast the hardware is?