My response to the latest HardOCP editorial on benchmarks...

[Fred]

Bottom line. A benchmark is only as good as the people who interpret the meaning. For that reason, I don't care what FutureMark does or does not do.

Yes the average consumer might be duped by the scores, but thats what makes him *average*.

 

[Morris Ital]

"Approximately 250,000 polygons are rendered per frame using 1.1 pixel shaders. With 1.4 pixel shaders this number is lowered to approximately 150,000 due to reduced number of rendering passes." (emphasis mine)

Considering most (but apparently not all in game2) polys undergo two passes with PS1.4 and four with PS1.1, these numbers (along with the 560,000/280,000 numbers for game3) are entirely expected.

If it was 2 and 4 then it would be 15000 000 and 300 000 for 1.4 and 1.1 respectively ie 2 to 1.

No matter what the white paper says 250 000 and 150 000 will always have a 1.66667 ratio so it is not to do with rendering passes unless you can have fractions of rendering passes.

 

[Ichneumon]

"Approximately 250,000 polygons are rendered per frame using 1.1 pixel shaders. With 1.4 pixel shaders this number is lowered to approximately 150,000 due to reduced number of rendering passes." (emphasis mine)

Considering most (but apparently not all in game2) polys undergo two passes with PS1.4 and four with PS1.1, these numbers (along with the 560,000/280,000 numbers for game3) are entirely expected.

If it was 2 and 4 then it would be 15000 000 and 300 000 for 1.4 and 1.1 respectively ie 2 to 1.

No matter what the white paper says 250 000 and 150 000 will always have a 1.66667 ratio so it is not to do with rendering passes unless you can have fractions of rendering passes.

Yes, but as noted by Maverick, with 3 passes for 1.4 vs 5 passes for 1.1 the numbers fit quite nicely.


I don't know enough about how the additional passes work, but could the other reason for it be because the additional passes are only for lighting, and not all geometry needs to be resent for those passes?

 

[Joe DeFuria]

You are all assuming that EVERY POLYGON requires multiple passes. This is probably not the case at all.

Edit...except Ich!

 

[Morris Ital]

Am I reading this wrong ?

Yes.

I'm just making these numbers up, but they do fit:

The scene has 50,000 polys.
Using 1.4 shaders, 3 passes are required to achieve the required result. 3 * 50,000 = 150,000.
Using 1.1 shaders, 5 passes are required. 5 * 50,000 = 250,000.

See?

Acording to www.rage3d.com ( and ignoring Kyle to make Doomtrooper happy )

"With 1.1 pixel shaders, objects need one rendering pass for depth buffer initialization and three passes (stencil pass, light fall-off to alpha buffer, and diffuse and specular reflection) for each light source. When 1.4 pixel shader support is available only one pass for each light source is needed"

So thats 4 pases v 1 pass

Not 3 passes as the above mentions for 1.4. I do not know where 5 passes came from, if it is 3 for 1.4 it should be 12 for 1.1 and not 5 ?

 

[Morris Ital]

If the hardware doesn't support PS 1.4 there isn't much you can do about it is there. Enough of the PS 1.4 whining, If futuremark inlcuded PS 1.3 then the Geforce 3 folks would be ranting and raving as Pixel Shader 1.1 isn't supported...can't please everyone.

you miss the point entirely.

I am not whining about 3DM03, I actually paid for the Pro version. What I am asking for is a reason for extra polygons for 1.1 shaders. At present I have not heard a valid reason based on number of passes.

Regadrs

Andy

 

[Dave H]

You are all assuming that EVERY POLYGON requires multiple passes. This is probably not the case at all.

Exactly. OTOH, the 3 vs. 1 extra passes are per-light, so the total number varies depending on the total number of lights in the scene. These two effects--not all polys probably get the multi-pass shaders, but with multiple lights the PS1.1 card does more than double the number of passes--balance to get the result that total polys rendered is about double for PS1.1 vs. PS1.4.

 

[Luminescent]

It seems the 3Dmark result number scales proportionally (almost 1:1) with the amount of gpu transitors dedicated to shading (to an extent). The R300 sports about 110 million, and the NV25, about 66 million, on the same process. Accoridng to the diagrams in the following web links, by observation, about 60-70 percent of those transistors account for the shading on the R300 while 25-30 account on the NV25. This would be about 110*.65 and 66*.275, which yields 71.5 and 18.15 respectively. 71.5 corresponds to approximately 4 times the amount of 18.15. This seems to correspond with the 3DMark03 scores. The same may even go with Geforce FX. This benchmark might just scale with complexity, a term which implies productivity in the technological arena; not bad . This is just a theory, comment if you wish.

R300:
http://www.anandtech.com/video/showdoc.html?i=1656&p=2

NV25:
http://www.extremetech.com/article2/0,3973,36877,00.asp

 

[Nagorak]

If the hardware doesn't support PS 1.4 there isn't much you can do about it is there. Enough of the PS 1.4 whining, If futuremark inlcuded PS 1.3 then the Geforce 3 folks would be ranting and raving as Pixel Shader 1.1 isn't supported...can't please everyone.

you miss the point entirely.

I am not whining about 3DM03, I actually paid for the Pro version. What I am asking for is a reason for extra polygons for 1.1 shaders. At present I have not heard a valid reason based on number of passes.

Regadrs

Andy

What do you want them to do, create a scene with 1/4 the polygons just for a card that has 1.1 pixel shaders?

 

[Morris Ital]

"Yes, but as noted by Maverick, with 3 passes for 1.4 vs 5 passes for 1.1 the numbers fit quite nicely. "

I BLOODY GIVE UP



It might do but then if it is 3 passes for 1.4 and 5 passes for 1.1 then I'd expect roughly 3/5 the performance for a 1.1 GF4 card compared to a 1.4 Radeon and actually I get 1/3-1/4.

To me it still looks as if the 1.1 card does mutliple passes for every one the 1.4 does and it also looks, from reports, that it does more polygons. t

Nothing anyone has said on this thread has convinced me otherwise.

 

[Ichneumon]

I don't understand what your issue is Morris.
it DOES do more polygons on a 1.1 card... but it isn't like 3dmark goes AHAH! a 1.1 supporting card, we'll make the scene use extra polygons now!

1.1 and 1.4 cards render each scene with the same number of polygons. However, on a 1.1 shader card, some of those polygons must be rendered multiple times per frame because of the additional lighting passes.

Edit: You can see what happens with PS 1.1 and 1.4 on a 9700 from my tests in this thread

 

[Morris Ital]

What do you want them to do, create a scene with 1/4 the polygons just for a card that has 1.1 pixel shaders?

Is that a trick question ?

I've stated that I want to know what is the reason why people on hardware sites keep quoting more polygons for 1.1 v 1.4. I thought I'd get a simple answer but no,

Nobody expects the Beyond3d Inquisition

I'd like equal polygons for all and more passes for cards not supporting a more efficient shader, ie 1.4

Is that unreasonable ?

I need a stiff 1.4 drink that I will consume in one pass

 

[Morris Ital]

I don't understand what your issue is Morris.
it DOES do more polygons on a 1.1 card... but it isn't like 3dmark goes AHAH! a 1.1 supporting card, we'll make the scene use extra polygons now!

1.1 and 1.4 cards render each scene with the same number of polygons. However, on a 1.1 shader card, some of those polygons must be rendered multiple times per frame because of the additional lighting passes.

Edit: You can see what happens with PS 1.1 and 1.4 on a 9700 from my tests in this thread

I agree but the figures given by two sites are 250 000 and 150 000 and that means a fractional numebr of passes. If it was 200 000 and 100 000 then I would agree as it would be two.

Even the sites are confused, if it is down to the number of passes, because they state it as if it was passes AND extra polygons ( read it yourself ! ) .

They state polygons and passes and I am trying to clear it up because if you take them to mean the same thing then the numbers they quote adds up to 4 passes for 1.1 and 1 pass for 1.4 and 250 000 polygons for 1.1 and 150 000 polygons for 1.4 and that does not add up.

Look, I am not trying to defend anyone of say someone is better than another, I am just trying to understand the maths and nobody is coming up with the right mathmatical answer.. yet

 

[jvd]

morris.... Mabye the sites rounded or just made up a number out of thin air... or it could be an act of god

 

[Dave H]

The reason the PS1.1 rendered poly count is not a simple multiple of the PS1.4 count is because different shaders are applied to different polygons. In the most common case--poly lit by 1 light according to the unified lighting algorithm--it takes 2 passes with PS1.4 and 4 with PS1.1 But different cases will be different.

 

[OpenGL guy]

I've stated that I want to know what is the reason why people on hardware sites keep quoting more polygons for 1.1 v 1.4. I thought I'd get a simple answer but no,

Nobody expects the Beyond3d Inquisition

I'd like equal polygons for all and more passes for cards not supporting a more efficient shader, ie 1.4

That's exactly what is happening. However, when computing the total number of polygons rendered you must count how many times each polygon was rendered. Thus, doing more passes on the same polygons will result in a larger number of total polygons rendered.

This is what everyone has been saying all along.

As I stated earlier, if you have to draw the same polygon five times, then that counts as five polygons.

It might do but then if it is 3 passes for 1.4 and 5 passes for 1.1 then I'd expect roughly 3/5 the performance for a 1.1 GF4 card compared to a 1.4 Radeon and actually I get 1/3-1/4.

You'd be correct if the two cards rendered everything at exactly the same rate, which they don't.

 

[Hellbinder]

I am still so pissed off about this that I am really having to be careful what i write. I just cant believe how many people are falling for and supporting Nvidia's out misinformation. It's sickening.

The whole point is that 3dmark03 is future looking not past looking. Every single solitary comment nvidia made in their PR releases are complete and utter rubbish. I will not apologize for speaking the truth, or try to tone down the message. There is no balance, there is no middle ground in this case. The ONLY thing that anyone could say is no more synthetic tests and leave it at that.

The bottom line is that over the next 3 years we are going to see exactly what 3dmark is offering. Hell They have dozens and dozens of developers and hardware companies in the beta program all entering input. Only Nvidia withdrew.

1. Future games are going to use PS 1.1, PS 1.4, and PS 2.0
2. Future games are going to use a lot of basic Vertex shader 1.1
3. Future games are going to be single texture with PS and VS
4. Future games will primarily use a lot of DX8 features to their fullest and incorperate partial DX9 support. meaningthe occasional PS 2.0 or Vertex shader 2.0 functionality.

I dont see how anyone can debate that wile treating the future with an honest eye. it follows the eact pattern of the last 3 years and its technology of T&L etc. Do we really need to go over again how PS 1.3 support would not have affected the outcome of any scores for Nvidia? it does not have that functionality. Further their origional GF3 series does not support PS 1.3 thus 3dmark HAD to use PS 1.1 for the greatest Nvidia compatability.

Look at the pattern its exactly the same as the last 3dmark, it heavily used the first generation of DX7, and incorperated features of DX8 here and there. The same will be true with the next 3dmark it will have a little DX8, heavy use of DX9, and begin to incorperate dX10.

The truth is so obvious it is beyond me how it seems that so many are falling for nothing but cheap PR misinformation tactics.

 

[Maverick]

I agree but the figures given by two sites are 250 000 and 150 000 and that means a fractional numebr of passes. If it was 200 000 and 100 000 then I would agree as it would be two.

Even the sites are confused, if it is down to the number of passes, because they state it as if it was passes AND extra polygons ( read it yourself ! ) .

They state polygons and passes and I am trying to clear it up because if you take them to mean the same thing then the numbers they quote adds up to 4 passes for 1.1 and 1 pass for 1.4 and 250 000 polygons for 1.1 and 150 000 polygons for 1.4 and that does not add up.

Look, I am not trying to defend anyone of say someone is better than another, I am just trying to understand the maths and nobody is coming up with the right mathmatical answer.. yet

But we can't really give the right mathematical answer, as we don't have the right numbers to begin with. The 150,000 and 250,000 are just approximate numbers, and will vary depending on whereabouts in the scene you are, both because the amount of geometry in the scene will vary, and the number of lights affecting that geometry will change too.

One marine chappie from game 2 might be standing someplace where there is only one light shining on him, so in 1.1 it'll take 4 passes, and with 1.4 it'll only need 2. A second marine might be standing futher down the corridor, where there are 3 lights, so 1.1 needs 10 passes, and 1.4 takes 4, and so on. Over the whole of the scene, it apparently averages out to approx 250,000 and 150,000 polys, which happens to be a ratio of 5:3. So on average 1.1 is doing 5 passes for every 3 that 1.4 does.

 

[Maverick]

]3. Future games are going to be single texture with PS and VS

I seriously doubt that future games will use single texturing. They may use one texture, do some calulation, then get another texture, and so on, but that's a whole different kettle of fish to single texturing.

I don't really think that game one is really representative of any recent game, let alone any future game. Heck, it doesn't even compare to IL-2, which is going on for 2 years old now...

 

[Hellbinder]

Maverick,

It has beed discussed countless times at this forum that the next generation of games are moving away from the DX7 Multi-textured approach.

Game one is closer to just about any Intro sequence seen in games today. Especially those using the actual engine for the intro. You could have dolphins jumping in the ocean and it would be the same damn thing.