I dont mean to spread fud..

[PeterAce]

Evidance has been posted here before about legitimate performance increases for ATI's shader compiler :

Using 'Marko Dolenc's Fillrate Tester' :

(Catalyst 4.1)

PS 2.0 - Per Pixel Lighting - 167.174103M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 167.174225M pixels/sec

(Catalyst 4.3)

PS 2.0 - Per Pixel Lighting - 188.016754M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 188.015060M pixels/sec

An improvement of ~11%.

http://www.beyond3d.com/forum/viewtopic.php?p=235782&highlight=catalyst#235782

And from ATI's release notes :

Performance Improvements
As with most CATALYSTâ„¢ releases performance has increased in various situations. The following performance gains are noticed:

The Pixel Shader performance of our DX9-class products has improved considerably with Catalyst 4.3.

Halo and Tomb Raider framerates are up 3-7%.

Shader-specific technology tests such as ShaderMark show gains in the region of 10-20% or more.

3DMark03's Mother Nature scene runs 2-5% faster across the entire DX9 product line.

Our 3DMark03 Pixel Shader 2.0 framerate improves 7-15%.
Aquamark 3 performance has also improved slightly, at higher resolutions.

http://www2.ati.com/drivers/Catalyst_45_Release_Notes.html

Dave hint's about new or improved shader compiler (from ATI's ex-DEC employees) :

http://www.beyond3d.com/forum/viewtopic.php?p=279840&highlight=compiler#279840

Dave about the Catalyst 4.5 beta's :

http://www.beyond3d.com/forum/viewtopic.php?p=289350&highlight=compiler#289350

 

[Bouncing Zabaglione Bros.]

Just Curious, Have there been any comparisons really made between the Beta 4.5 and the Normal 4.5s? It would be nice for 4.6 to hurry up and get here so we can see, If I had an X800 Pro I'd be doing all sorts of comparisons.

If they stick to their usual release schedule, we should see Cat 4.6 in 2-3 weeks. Not too long really.

 

[Pete]

IQ comparisons? Because you posted performance comparisons in your thread starter.

 

[ShePearl]

There is one thing which makes me curious.

If I'm not mistaken, the new generation graphic cards (nv40, R420) are bottlenecked by CPUs. (At least, this was the impression I've got from various benchmarks). If this is the case, how come there's an improvement in games like Far Cry with Catalyst 4.5 Beta ??

If these cards are bottlenecked by CPUs, no matter how much better new drivers are, we wouldn't be able to see the improvement because CPU is holding it back anyway ? But this doesn't seem to be the case with Catalyst 4.5 Beta.

Also, if they are indeed bottlenecked by current CPUs (be it Intel and/or AMD), we would not be able to see any performance boost in future drivers unless we change/upgrade CPUs. (I know this bottleneck issue greatly depends on different games/engines, visual settings such as (resolution, AF, AA etc...) But generally, speaking....

What am I missing ?

 

[ZenOps]

I wouldn't be surprised that there will be the first tangible increase in Catalyst peformance with the next release.

Although I don't like it when it happens - Nvidia did the same thing on several of ATi's launches... They held back all the speed tweaks until the release of the next gen chipset from a competitor (They could have put the speed tweaks in earlier revisions)

Yes, all current GPU's are CPU bound. Enchanced tweaks in the drivers usually improve performance by putting less strain on the CPU. Sort of like how DMA mode for a hardrive is faster than PIO mode even at the same Mbits/sec because it uses much less CPU resources to get the same job done.

 

[Pete]

Next-gen games like Far Cry are shader-heavy, thus they're also bottlenecked by the GPUs (which are, in turn, bottlenecked by immature video card drivers. We're not always dealing with one fixed bottleneck. Heck, even the games themselves are being constantly optimized to yield better performance on the same hardware.

But with Far Cry and other shader-heavy games, the GPU drivers and game code itself will continue to be "soft"/moving bottlenecks as large as the generally "hard" (MB drivers and code compilers tend to have less room for improvement) CPU+RAM bottleneck.

Edit: Zen, I'm not sure it's just a case of optimizing their driver code, but also of optimizing their shader compilers (which would seem to have a lot of catching up to do compared to CPU compilers).

 

[dan2097]

I wouldn't be surprised that there will be the first tangible increase in Catalyst peformance with the next release.

All seems very odd as in terms of dx9 performance boosts the 4.3s gave a nice boost and the 4.5s (official) gave another similar boost. Both drivers increased mother nature about 1fps although I think only the 4.3s increased the ps2.0 test. The 4.3s increased halo performance, havnt got around to testing halo with the 4.5s.

Thats on a 9700np @near pro speeds.

Seeming as Ive had more performance increases in the 4.3s-4.5s than say 3.7s to 4.2s it would be odd for there to be a sudden 0-10% performance increase but I woudnt be complaining

 

[jvd]

SIgh Chalnoth we've seen for years drivers from nvidia that apear out of no where and give all cards a huge performance increase . Why wouldn't a new card with new memory controllers and new feature not have alot of power left untapped ?

Why do you think only nvidia is capable of that ?

That's not what I'm saying. What I'm saying is that the X800 appears to be too similar to the R3xx architecture to achieve such significant performance increases from "regular" optimizations in new drivers.

Well mabye you should back and read reviews and see about the features it has on p.s 2.0b and of course other things that it has added plus bug fixes and other things they could have done inside the core .

 

[KimB]

If, and only if, we see a general performance increase on pixel shader benchmarks with these drivers will these performance increases be valid. If this happens, then, of course, it means that the R4xx architecture is a bit more different than I currently think.

 

[jvd]

If, and only if, we see a general performance increase on pixel shader benchmarks with these drivers will these performance increases be valid. If this happens, then, of course, it means that the R4xx architecture is a bit more different than I currently think.

I suspect we will see you raise this question everytime nvidia has performance increasing drivers from now on too ?

 

[KimB]

There are enough people that do that now that I don't have to.

Besides, it's not that I don't think ATI can increase performance with new drivers. It's that the R420 didn't change much at all from the R3xx, such that it seems to me that the drivers should already be mature. I don't believe there is a precedence for performance-increasing drivers nearly two years after initial release.

 

[jvd]

There are enough people that do that now that I don't have to.

Besides, it's not that I don't think ATI can increase performance with new drivers. It's that the R420 didn't change much at all from the R3xx, such that it seems to me that the drivers should already be mature. I don't believe there is a precedence for performance-increasing drivers nearly two years after initial release.

Well I disagree with you . Where does it say the r420 isn't much of a change from the r3x0 series . Do you think they spent the last 2 years scratching thier ass and all they came up was lets double the amount of pipelines of the r3x0 ?

 

[Dave Baumann]

Besides, it's not that I don't think ATI can increase performance with new drivers. It's that the R420 didn't change much at all from the R3xx, such that it seems to me that the drivers should already be mature. I don't believe there is a precedence for performance-increasing drivers nearly two years after initial release.

Just by altering the number of pipelines (and potentially altering the cache / temp sizes that go along with that) can make significant differences in the bubbles in the pipeline, so just preliminary bubble analysis can give R420 gains even if it didn't change all that much from R300. You are also making the assumption that R300 was maxed out from the start which might not be the case given the shader performance disparities from the last generation.

However, there are still specific alterations to each quad pipeline from R300, which will also gain performance once they are understood.

 

[KimB]

Just by altering the number of pipelines (and potentially altering the cache / temp sizes that go along with that) can make significant differences in the bubbles in the pipeline, so just preliminary bubble analysis can give R420 gains even if it didn't change all that much from R300.

I really don't think so. That problem should have been solved long ago, back when they started cutting pipelines out to make lower-end chips. Besides, we have seen benchmarks with the R420 underclocked to appropriate levels, and it was almost identical to the R3xx in performance. If there were any problems with simply having more pipelines, you'd think that those benchmarks would favor the R3xx.

Then, if you claim that the R420 had efficiency improvements that offset the problems in moving to a 16-pipe design, you have to explain how those efficiency improvements just happened to exactly offset the losses from increasing the number of pipelines.

No, I think that with a simple increase in cache sizes, provided with lessons learned on previous designs, there is no reason for the move to 16 pipelines to cause any performance problems (that might be solved later).

You are also making the assumption that R300 was maxed out from the start which might not be the case given the shader performance disparities from the last generation.

Not from the start, but is maxxed out now.

However, there are still specific alterations to each quad pipeline from R300, which will also gain performance once they are understood.

We'll see.

 

[Dave Baumann]

That problem should have been solved long ago, back when they started cutting pipelines out to make lower-end chips.

We saw considerable changes between the true 4 pipe parts and the cut back R300's in some tests, with RV350 having some largely different PS performances from 4 pipe R300's.

But anyway, you're sure (without knowing the architectural changes) that R300 was already maxxed out and that there would be no PS improvements at all to come for R420 or any other R300 based board. I suppose that the increase we see here are the work of cheats then?

 

[KimB]

Well I disagree with you . Where does it say the r420 isn't much of a change from the r3x0 series . Do you think they spent the last 2 years scratching thier ass and all they came up was lets double the amount of pipelines of the r3x0 ?

Yes, that isn't far from the picture they have painted. Consider: they've had all this time to put out a new architecture, and all they've managed to do is increase the instruction counts and increase the number of pipelines.

Obviously work has been done, but where has that work gone? From the rumors, the original R400 was cancelled, and this R420 put in its place. I do believe that not much work (in comparison to its complexity/preivous products) was put in the R420 design. It seems apparent that whatever serious work ATI has been doing over the past few years won't be seen in a real project for another year (in the R500).

 

[KimB]

But anyway, you're sure (without knowing the architectural changes) that R300 was already maxxed out and that there would be no PS improvements at all to come for R420 or any other R300 based board. I suppose that the increase we see here are the work of cheats then?

If there was any more performance to pull out of the R3xx architecture, then ATI has been damned slow on the driver optimization front. So I definitely expect no significant performance improvements on the R3xx.

I further suspect that the R420 is too close to the R3xx to show driver improvements, either. We'll see whether this is the case or not, of course, but yes, if these performance improvements are not across the board in shader benchmarks/games, then I suspect it's a cheat.

 

[Dave Baumann]

If there was any more performance to pull out of the R3xx architecture, then ATI has been damned slow on the driver optimization front. So I definitely expect no significant performance improvements on the R3xx.

Optimisation efforts are put into the areas where you are most required to optimise - from a competetive standpoint Shader performance wasn't one of those areas in the previous cycle, so why expand significant efforts there?

I further suspect that the R420 is too close to the R3xx to show driver improvements, either. We'll see whether this is the case or not, of course, but yes, if these performance improvements are not across the board in shader benchmarks/games, then I suspect it's a cheat.

These drivers were approaved by Futuremark, when generally means there are no 3DMark03 application specific elements in there, and considering that improvments are seen with all PS applications tested this would appear to be something generic - its also not the case that ATI can do something such as simply lower precisions.

This kinda undermines your entire argument - you say that R300 is a done deal and no further optimistations can be done, but the beta drivers, and the performances shown at the start of this thread, shows thats not the case.

 

[ChrisRay]

The question will be answered when Cat 4.6 Become available, By then I think some real comparison should be made between 4.5, Beta 4.5 and 4.6

 

[jvd]

Well I disagree with you . Where does it say the r420 isn't much of a change from the r3x0 series . Do you think they spent the last 2 years scratching thier ass and all they came up was lets double the amount of pipelines of the r3x0 ?

Yes, that isn't far from the picture they have painted. Consider: they've had all this time to put out a new architecture, and all they've managed to do is increase the instruction counts and increase the number of pipelines.

Obviously work has been done, but where has that work gone? From the rumors, the original R400 was cancelled, and this R420 put in its place. I do believe that not much work (in comparison to its complexity/preivous products) was put in the R420 design. It seems apparent that whatever serious work ATI has been doing over the past few years won't be seen in a real project for another year (in the R500).

And what has nvidia been able to do with that time ? What has nvidia been doing the last few years ? They have thrown out rehash after rehash. Each time finding ways to continue improving drivers .

Ati has done more than increase insturction counts (Which may lead to speed increases ) and the number of pipelines.
You say obviously work has been done . While the vast majority has been put into new projects neither of us know what has been done.

Unless you have proof and can post your proof that there are no changes that have been done on this hardware aside from the increased pipelines .

These increases apear to be valid . Dave also doesn't seem to think they are cheats . So untill someone can prove that it is a cheat i wont really believe it .

If there was any more performance to pull out of the R3xx architecture, then ATI has been damned slow on the driver optimization front. So I definitely expect no significant performance improvements on the R3xx

Why ? Ati has said time and time again they are working on getting all games working with no bugs than working on speed increases .

You assume to much and i think you just want to join in with the witch hunting