Nvidia Against 3D Mark 2003

Well then what of 1.1? I think you are overestimating the importance of PS 1.4 and underestimeting the importance of PS 1.1.

I think depth_test is quite objective and he says that 1.4 is more advanced than 1.1 but that 1.4 can't be told as the unique raison d'etre of PS 2.0 i.e. no more than 1.1, and PS2.0 doesn't look too good on GFFX actually! So what's the point?

And if he works at Nvidia, i hope he will come more often over here

For the rest well fanboyism

 

What *rest* is fanboyism.. posted by who..????

And no, i am not overstating the importance of PS 1.4 V.s PS 1.1 And if you knew anything about it, you would not be suggesting that its fanboi talk to state such a thing.

Ask youself, If they really were almost the same? and made nearly no difference? why would Nvidia even care???

A shader that accomplishes the exact same visual affect in PS 1.1 and PS 1.4 would only require 1/2 the work and only one pass on the PS 1.4 version. The same cannot be said about PS 1.1, 1.2, or 1.3

 

I agree, but it doesn't really weight anything in the overall score, and consider the utter disbelief by consumers who bought a fancy DX7-level card last spring/summer (ahmm, GF4MX) just see this report:
GT1- not supported
GT2- not supported
GT3- not supported
GT3- not supported

 

good.
The sooner the bottom end goes up, the sooner games with better graphics come.
I wish this had happened.

 

Well, sort of agree here (about the disbelief ). But, since it's only one out of 4 tests and then as you say, have a very low weight in the final result then i don't think that they're going to be that happy anyway.

 

Point was 1.1 an advance at the time, as was 1.4? Response yes
Was it a bigger step than 1.1-1.4? I would say yes, but i think you would say no . I tend to think first steps are more important than the next important but not critical (as PS2.0)

I remember that Carmark stated that it should be faster on 8500 but in fact was on a GF4 (or was he speaking of another thing? ), so as for performance...

Why does Nvidia care? Perhaps because none of their cards on the market use PS1.4? Perhaps because 99,9% of their market share in 3 month won't use 1.4?

Fanboyism? Well look at the way you talk of Nvidia and people working over there Should give you a response

 

You want to talk fanboyism get a mirror. I have had it with Nvidia and their tactics and rhetoric and twisting of the truth as of today. And im not pulling any punches anymore. Neither am i going to sit around while someone tries to disipate the truth of the issue with a bunch of muddled aguments over technicalities of terminology.

This has nothing to do with the speed or benefit of the pixel shader versions of each and everythign to do with some low level hardware limitations in the core of the 8500.

 

Irrelevant. I never said anything about 1.3's importance or quality at all.

Incoherent. What is your last sentence supposed to mean? What program are you talking about? As an aside, I never made any claim that 1.3 == 1.4 or that 1.3 added anything of real value over 1.1. It is you who are attempting to equate 1.4 and 2.0.

There is a difference between saying something is a stepping stone (as was PS1.0, and DX7) and saying something is based on, or an extension of something else. Clearly, 1.4 is not a subset of 2.0, and 2.0 is not simply an extension of 1.4. Did 1.4 influence 2.0? I'm sure it did, just like 1.0 did. I'm sure there were atleast 3 or 4 proposals floating around Microsoft for what 2.0 shaders should look like, and many of them probably drew inspiration from what IHVs were doing.


Anyway, how can you even put forth an argument about how important 1.4 was without being willing to disect the minute details of it?

I don't believe I ever argued that 1.4 wasn't more advanced than 1.1. I went into "disection" remember? I mentioned what it added over 1.1: In 1.1-1.3, you were allowed 4 texture instructions and 8 color instructions yielding a pixel shader max length of 12. 1.4 increased the length of the texture instructions to 6, yielding a max limitation of 14 instructions. The "big leap" if you want to call it that, was the ability to chain two of these programs together, thus yielding a program of max length 28, the main benefit of which is that it could perform 1st order dependent lookups.

However, everything is not all that rosy as it appears on the instruction counts. Since 1.4's texture shader instructions are more RISC, it takes more 1.4 shader ops to handle what used to be a single CISC op in 1.1, consuming more slots.

What 1.4 enables you to do over 1.1 is a subset of shading effects that require single-pass 1st order dependent texturing. Ok, cool, but it is not as earth shattering as you make it out.

Perhaps Humus can shed his feelings on 1.4 vs 2.0 if you feel I am being overly biased here in my assessment that 1.4 was not some kind of revolutionary, disruptive, sea-changing advancement.

Nice theory, but wrong. I don't work for any IHV. Did you just miss my negative comment about the GFFX's ability to exceed 3rd order dependent texturing?

I suggest you sell your 8500 and stop wishing for 1.4 to be relevant. It's legacy, and we need to dispense with 1.x in the market place and move forward to 2.0.

 

Yeah Hellbinder

Well, my thinking

http://www.gamedev.net/columns/hardcore/dxshader3/

 

AFAIU DX9 is either Shaders 2.0 or Shaders 3.0. There is no inbetween.

 

I think that the main point NVidia has against 3D Mark 2003 is that "bizarre" shadowing technique (and not pixel shader 1.4 support).
Here's why:
Second and third tests don't actually hit the same bottleneck on GeForce 4 and Radeon 8500, Radeon 9500. The thing is that Radeon 8500 is pixel shader limited in both situations, while GeForce 4 is definitely vertex shader limited in both situations (after all it has to transform 3x more vertices and still comes out on par).
I think it's damn obvious that any reduction of triangles would help GeForce 4 a great deal and that "bizarre" shadowing technique is probably a very good place to start reducing triangles.
Mark Kilgard said in his "Robust Stencil Shadow Volumes" presentation: "Carmack’s optimizations in these areas will make new Doom engine awesome".

 

Simon F: You have extended 2.0 in DX 9. And it comes quite a long way to ps_3_0...

 

There is PS2.0 Extended which adds the gradient instuctions, predicates, and dynamic/static flow control to 2.0

 

Oh typical. I suppose this is hidden away somewhere. I wish MS had one single reference rather than having to piece things together from docs/source code/help files/runtime ....

Edit: Ahh I think I was assuming VS and PS versions were linked together. I think I should go back to messing around with texture compression...

 

Actually 1.4 enables you to do up to 6 texture look-ups in a single pass (compared to "only" 4 with 1.1-1.3). For Doom3-style rendering (Game 2 and 3) this subtle difference has a tremendous impact, enabling you to render a single pass per light instead of two (using the frame buffer's destination alpha as temporary storage). As Doom3-style rendering renders the whole scene geometry each pass, 1.4 basically enables you to save on both vertex and pixel processing performance. HUGE difference, and perfectly valid to implement in a game.

 

No, it's placed on quite open place in the SDK: DirectX Graphics -> Reference -> Shader Reference...

 

I don't think so. I think, it's because game 4 is a mixture of DX9 and DX8.1 as pointed in all the reviews of 3DMark03.

http://www.hardocp.com/article.html?art=NDI4LDM=

 

Hee hee... I can just picture all the effort that would go into the HLSL compilers to make them recognise specific shaders and instantly choose some hand-optimised object code.

 

Well I think it has some pixels without any shaders applied, so I guess that makes it DX7.

"Partially DX9." Am I missing something, or is this a completely incoherent argument?

 

Actually you can do up to 12 texture lookups in a single pass using PS1.4 but only from 6 different textures. With PS1.1->1.3 you can do 4 lookups from 4 textures. This is actually quite an important capability of PS1.4 because there are circumstances where you may want to look up data from some texture maps more than once per pass (for example normalisation cube maps), and PS1.4 has the resources to enable you to do this.