NV40 vs R420 Extreme Pipelines

[radar1200gs]

DW: I wasn't referring to PS3.0 in my post.

Shader Model 3.0 breaks down into Vertex Shader 3.0 and Pixel Shader 3.0.

It's almost a given that ATi's pixel shaders will remain at PS2.0, however they might impliment VS 3.0 or part of it, specifically I think they might want to support displacement mapping and possibly beef Truform back up to what it was in R200.

 

[digitalwanderer]

DW: I wasn't referring to PS3.0 in my post.

Shader Model 3.0 breaks down into Vertex Shader 3.0 and Pixel Shader 3.0.

It's almost a given that ATi's pixel shaders will remain at PS2.0, however they might impliment VS 3.0 or part of it, specifically I think they might want to support displacement mapping and possibly beef Truform back up to what it was in R200.

Ah, I got VS & PS confuzled again...sorry.

I think it's already a given that ATi's next card will support PS3.0, but I think the bajillion dollar question right now (at least for me) is how well they'll support it.

They could support it in such a way that it is a useful feature that adds performance/eye-candy, or they could support it in a FX5200 way where it looks really good for the PR but don't mean squat for real-world useability.

Which it's going to be I honestly don't know yet, and I'm just as anxious to find out as anyone else.

BTW-Isn't the PS3.0 standard somehow tied to VS too? Or am I WAAAAY off-base? :|

EDITED BITS: Nevermind my "BTW", my brain finally digested the "Shader 3.0" bit you wrote...thanks for the explanation.

 

[radar1200gs]

They can't support Pixel Shader 3.0 because of their branching limitations. That leaves only the Vertex Shader to improve in terms of DX9 features (they can add more raw speed, but thats a fairly lame "extreme" pipeline and various folks are hinting ATi has added more than just speed).

 

[digitalwanderer]

They can't support Pixel Shader 3.0 because of their branching limitations.

I believe there is more to the PS 3.0 spec than branching, and I believe the R420 will support those features of PS 3.0....that's what I was trying to say.

 

[Heathen]

That leaves only the Vertex Shader to improve in terms of DX9 features

There's more to PS 3 than just the branching, Look at the PS profile chart in the 6800 review. There's lots ATi can add while sticking within the PS 2.0 boundaries.

Other things they could add
1) FP32
3) More Z-check units for faster MSAA
4) Fully programmable AA engine ALA 3dlabs P10
5) Second full ALU
6) their own version of Nvidia's Ultrashadow.
7) clock doubling of the PS ALUs
8) bi directional colour and Z-buffers

I'm guessing a few of the big brains may be able to comment on the practicality of some of these or even thrown in a few more possibilities

 

[Bjorn]

I think it's already a given that ATi's next card will support PS3.0, but I think the bajillion dollar question right now (at least for me) is how well they'll support it.
...
BTW-Isn't the PS3.0 standard somehow tied to VS too? Or am I WAAAAY off-base?

The thing with PS3.0 (if i'm not entirely mistaken) is that all features are required. Unlike PS2.0 where there's loads of optional features like MRT f.e.

And afaik, you can support (some people here looked that up if i remember it correctly) VS3.0 without supporting PS3.0.

 

[T2k]

It's almost a given that ATi's pixel shaders will remain at PS2.0, however they might impliment VS 3.0 or part of it, specifically I think they might want to support displacement mapping and possibly beef Truform back up to what it was in R200.

Actually it would be weird if they would REMOVE DM since it's already supported by the R3x0-series.

BTW AFAIK DM itself has nothing to with SM3.

 

[KimB]

From my limited understanding, ATi is going to be supporting the most important features of PS 3.0 that will be used in games first, but they won't be supporting the full standard.

nVidia's documents released to NV40 reviewers include the specs of "PS 2.0b." This is probably indicative of the PS support of the Radeon X800:

http://www.beyond3d.com/previews/nvidia/nv40/index.php?p=5

Notice the lack of Instruction Predication, Arbitrary Swizzling, and Gradient Instructions.

Instruction Predication is essentially a rudimentary form of branching. Given the lack of real branching, this is a disappointment.

Arbitrary swizzling is another drawback: some shaders will require extra instructions.

Gradient instructions are absolutely necessary to do any sort of antialiasing in the shader. Antialiasing is necessary in the shader, for example, with procedural textures (since procedural textures cannot be filtered).

Of all these, I think the most disappointing is the lack of gradient instructions. The other two will merely serve to reduce efficiency, but gradient instructions offer new functionality. It's really disappointing that those are not available.

So no, I don't think ATI is exposing the "most important features of PS 3.0." I think they're only exposing extra instructions, temporaries, and texture accesses. No additional instruction support seems apparent.[/img]

 

[Tim Murray]

And afaik, you can support (some people here looked that up if i remember it correctly) VS3.0 without supporting PS3.0.

I believe that this was determined to be a limitation of early DX9.0c betas, but as of the newest beta, this limitation was removed. You can support ps_2_whatever and vs_3_0 with no problems.

 

[digitalwanderer]

Gradient instructions are absolutely necessary to do any sort of antialiasing in the shader. Antialiasing is necessary in the shader, for example, with procedural textures (since procedural textures cannot be filtered).

I'm not quite sure exactly what you mean since it's a bit above my head, but is this why we have such headaches getting AA with all the dx9 games lately?

So no, I don't think ATI is exposing the "most important features of PS 3.0." I think they're only exposing extra instructions, temporaries, and texture accesses. No additional instruction support seems apparent.

I'm not gonna argue with you, at least not until:

1. We know what the specs are

2. I know a bit more of what I'm talking about

Which of those will come first is anyone's guess at this point....

 

[BRiT]

I thought PS2.0b would more directly map to the ATI Rv380/Rv370 or perhaps even 9x00XT series.

 

[digitalwanderer]

Is "PS2.0b" the same thing as "PS2.0 extended" that we were talking about a few weeks/months ago?

 

[KimB]

Gradient instructions are absolutely necessary to do any sort of antialiasing in the shader. Antialiasing is necessary in the shader, for example, with procedural textures (since procedural textures cannot be filtered).

I'm not quite sure exactly what you mean since it's a bit above my head, but is this why we have such headaches getting AA with all the dx9 games lately?

No. Even though anti-aliasing is the proper term, I guess a more understandable description of what I'm talking about is texture filtering in the shader. Essentially, if standard texture filtering doesn't work (it won't in a procedural texture: there is no texture to filter!), one has to do texture filtering in the shader to avoid aliasing. Without gradient instructions, this cannot be done adequately.

I'm not gonna argue with you, at least not until:

1. We know what the specs are

2. I know a bit more of what I'm talking about

Which of those will come first is anyone's guess at this point....

Well, I'd say it's 95% certain that the shader capabilities of the R420 are equivalent to "PS 2.0b."

 

[BRiT]

Is "PS2.0b" the same thing as "PS2.0 extended" that we were talking about a few weeks/months ago?

Yes, we already discussed PS2.0b a couple of weeks ago...

 

[digitalwanderer]

Well, I'd say it's 95% certain that the shader capabilities of the R420 are equivalent to "PS 2.0b."

I'd actually go a bit higher than that, but not quite all the way to 100%.

Either I'm mellowing or burnt out from the nV40 launch, but I'm just not up to arguing over it until it's out and I understand a bit better what the what is about it all.

 

[KimB]

I thought PS2.0b would more directly map to the ATI Rv380/Rv370 or perhaps even 9x00XT series.

I don't think so. There's nothing in ATI's XT product page about increased instruction limits in the pixel shader, and there was nothing about it in Wavey's B3D review.

 

[radar1200gs]

I was under the impression that Displacement mapping required vertex texture capabilities (a mono texture that is used to offset the vertex mesh).

Can someone provide a link to the PS2.0b specifications (microsoft specs, not vendor specs)?

IMO PS2.0b < PS3.0 feature wise, the only possible advantage is speed (and daves preview stated NV40 can swizzle all its PS registers, removing a lot of register overhead).

 

[T2k]

Depends what you look for...

http://msdn.microsoft.com/library/d...ncedtopics/vertexpipe/displacementmapping.asp

 

[radar1200gs]

It looks to me like R300 is doing displacement mapping in software, not hardware going by that diagram (if it truly is capable of displacement mapping). nVidia could do that too.

 

[Tim Murray]

Did we ever reach a consensus on whether the "FP32 requirement" for PS3.0 meant that the internal precision must be totally FP32 and that the final output can still have a minimum precision of FP24 or whatever, or is it FP32 everywhere (except with _pp)?