FarCry Performance Revisited: ATI Strikes Back with Sh 2b

[Skinner]

Strange enough its the other way arround in the Source VST with the GF6800U coming out on top in the heaviest part of the demo (with 6600 drives)

Are you 100% sure that the 6800U walks the DX9.0 path with these 66.00 beta's and not the 'mixed mode' path?

I don't know, it looks the same, but I guess someone have to investigate this to be sure, I wonder if they do the same path!

 

[PeterAce]

Does the "Mixed Mode" use any PS1.4 shaders in place of PS2.0 shaders used in the default DX9 path?

[Brian Jacobson] Not for the 5900. Yes for the 5600 and 5200.

Does the mixed mode use the precision hint on all available instructions, most instructions or are the majority still full precision?

[Brian Jacobson] The majority are partial precision; about 10%-20% are full precision.

http://www.beyond3d.com/forum/viewtopic.php?t=8389&highlight=mixed

Has the number of paths changed from this?

The second FP16/FP32 mixed mode has now been removed?

So the FX 59X0 boards now run the 8.1/PS1.4 path?

 

[vb]

I often wish ATI's next generation part R5xx would be out already. Frankly if there's one thing I can't stand reading anymore are those precision related topics.

You may excuse the interruption.

next spring we'll talk about how each hardware is rounding the 24th bit

 

[Demirug]

Does the "Mixed Mode" use any PS1.4 shaders in place of PS2.0 shaders used in the default DX9 path?

[Brian Jacobson] Not for the 5900. Yes for the 5600 and 5200.

Does the mixed mode use the precision hint on all available instructions, most instructions or are the majority still full precision?

[Brian Jacobson] The majority are partial precision; about 10%-20% are full precision.

http://www.beyond3d.com/forum/viewtopic.php?t=8389&highlight=mixed

Has the number of paths changed from this?

The second FP16/FP32 mixed mode has now been removed?

So the FX 59X0 boards now run the 8.1/PS1.4 path?

Yes, yes and yes.

 

[Unknown Soldier]

Ta for the reply .. Always thought FP16 was < DX9.

US

 

[Ailuros]

I often wish ATI's next generation part R5xx would be out already. Frankly if there's one thing I can't stand reading anymore are those precision related topics.

You may excuse the interruption.

next spring we'll talk about how each hardware is rounding the 24th bit

Frankly I think it was Kristof that had predicted on these forums that it'll turn into another 16/22/32bpp colour depth alike debacle (think Voodoo times), but I can fairly say that it has gotten even worse.

Mixed floating point precision is not a panacea; pure FP32 is in terms of bandwidth though and I suspect it'll remain so for quite some time.

 

[jvd]

It be interesting if the r420 sieres turns out to be faster than the nv40 series in 3dmark2005 .


Would it again be who cares we don't play 3dmark from the nvidia camp like when nvidia was cheating in it ? Or will nvidia cheat in it again and the nvidia camp claim that its okay to cheat .

 

[Bjorn]

It be interesting if the r420 sieres turns out to be faster than the nv40 series in 3dmark2005 .


Would it again be who cares we don't play 3dmark from the nvidia camp like when nvidia was cheating in it ? Or will nvidia cheat in it again and the nvidia camp claim that its okay to cheat .

Or if the NV40 series is faster.....

I doubt that we'll see any cheating from any of the IHV's though. But i'm guessing that we'll see a lot of excuses from the loosing side. If there is such a thing. 3D Mark 2003 seems to be pretty even.

 

[jvd]

It be interesting if the r420 sieres turns out to be faster than the nv40 series in 3dmark2005 .


Would it again be who cares we don't play 3dmark from the nvidia camp like when nvidia was cheating in it ? Or will nvidia cheat in it again and the nvidia camp claim that its okay to cheat .

Or if the NV40 series is faster.....

I doubt that we'll see any cheating from any of the IHV's though. But i'm guessing that we'll see a lot of excuses from the loosing side. If there is such a thing. 3D Mark 2003 seems to be pretty even.

it wont be a big about face as the nv40 is already faster in 3dmark 2003 (though it may be cheating after daves comments about a compiler that will beat any changes done by futuremark). Don't see any ati fans bashing it right now. So i don't excpect we will in the future .

 

[vb]

It be interesting if the r420 sieres turns out to be faster than the nv40 series in 3dmark2005 .


Would it again be who cares we don't play 3dmark from the nvidia camp like when nvidia was cheating in it ? Or will nvidia cheat in it again and the nvidia camp claim that its okay to cheat .

What I would like to know if future control center will feature a "optimize 3DMark" option.

 

[Bjorn]

it wont be a big about face as the nv40 is already faster in 3dmark 2003 (though it may be cheating after daves comments about a compiler that will beat any changes done by futuremark).

If the compiler can beat any changes that FM does then aren't we talking about a generic optimization ?

Don't see any ati fans bashing it right now. So i don't excpect we will in the future .

3D Mark 2003 is old news and there are games out now which are better, or at least more interesting benchmarks (Doom 3, CS stress test, Far Cry). Besides, the difference isn't anything to talk about either.

 

[PatrickL]

Next 3dmark could be a hot topic again if some rumors are true.

 

[Ichneumon]

it wont be a big about face as the nv40 is already faster in 3dmark 2003 (though it may be cheating after daves comments about a compiler that will beat any changes done by futuremark).

If the compiler can beat any changes that FM does then aren't we talking about a generic optimization ?

Not if they are ignoring changes to the shaders and using some other hook to execute their own code from the drivers. There's nothing generic about faking a benchmark.

 

[Bjorn]

Not if they are ignoring changes to the shaders and using some other hook to execute their own code from the drivers. There's nothing generic about faking a benchmark.

That should be really easy to detect though.

 

[jvd]

it wont be a big about face as the nv40 is already faster in 3dmark 2003 (though it may be cheating after daves comments about a compiler that will beat any changes done by futuremark).

If the compiler can beat any changes that FM does then aren't we talking about a generic optimization ?

Don't see any ati fans bashing it right now. So i don't excpect we will in the future .

3D Mark 2003 is old news and there are games out now which are better, or at least more interesting benchmarks (Doom 3, CS stress test, Far Cry). Besides, the difference isn't anything to talk about either.

Well

I also have text from an MSN converstaion with Derek Perez who openly stated that part of their compiler optimisers task is to "replace shaders where they feel they need to" (I wasn't going to get into a discussion with him about what a compiler actually is), and that they would keep replacing shaders in their drivers every time 3DMark issues another patch to defeat their detections. We also have the case where we found in the intial 5900 drivers merely renaming UT would offer full trilinear. ATI have also openly stated to me that they have some detections in R200 drivers - there are thing like vertex buffers that get reassigned in order to better suit how they handle the vertex shaders.

http://www.beyond3d.com/forum/viewtopic.php?p=361734#361734


Everytime we bust them they just get better at hiding stuff.

Personaly shader replacements are a good thing except when comparing two cards .

If card a is faster with out optimizations to a game and game b is just as fast with the optimizations but without is slower. The consumer should know that. Because not every title is going to have that work put into it and mods or any updates can break it.

 

[hovz]

id wager that the nv40 will be faster.

 

[Luminescent]

If clocked equally, NV40 would open up a can on R420.

 

[Ostsol]

Ta for the reply .. Always thought FP16 was < DX9.

US

Before DX9 there was simply no FPxx in the programmable pixel pipeline. Everything was some sort of fixed point precision. PS1.x cards had 8 bits of mantissa precision, plus the sign bit. ATI's implementation of PS1.4 allowed the same precision, but more range. Instead of a range of [-1,1], they had [-8,8].

EDIT: To clear up more on precision. . . FP24 is the minimum for full precision in PS2.0 (and probably PS2.0_extended). For PS3.0, full precision must be FP32. Partial precision is a minimum of FP16 for all DX9 shader models. HLSL also specifies a double precision (minimum of FP64), though like partial precision, there is no requirement to support it (personally, I think it'd be more useful in the vertex pipeline).

 

[Unknown Soldier]

To clear up more on precision. . . FP24 is the minimum for full precision in PS2.0 (and probably PS2.0_extended). For PS3.0, full precision must be FP32. Partial precision is a minimum of FP16 for all DX9 shader models.

Ye that's what I thought .. I guess I just forgot the partial percision thing.

US

 

[Bouncing Zabaglione Bros.]

If clocked equally, NV40 would open up a can on R420.

But they don't clock equally, do they? Probably one of the reasons why ATI left out SM3.0 was so that they would have that clock speed advantage.