How many Polys can the New Consoles Push?

[Fafalada]

But you can NOT claim that "1 triangle can = 1 vertex, effectively."

Actually, in an ideal mesh topology, 1vertex = 2triangles.

 

[Wunderchu]

AFAIK, that 1.1B number isn't from Sony, rather it is an extrapolation. The G70 is capable of setting up 2 triangles per clock or 860 million. Ramp the clockrate up to 550mhz and you get 1.1B

That is my understanding, atleast.

ya, AFAIK, Sony never stated that particular spec.

 

[hasanahmad]

if one takes that pic. I posted as a 3D shape, then there are 4 triangles there

to draw a triangle you draw 1 triangle with 3 vertices, then a line (1) more vertex


if you take 2 squares and cut them into 4 triangles, will have 6 vertices, every square will have 4 vertices, every triangle will have 3,

 

[Titanio]

That is true, but is it fair to simply not account for the increased complexity of those vertices?

Something smells fishy to me.

For every other console in the past, they release how manys poly per second it can push in the spec sheet, this is the first time I have seen vertices/second rather than polys/second. Which leads me to wonder why.

The answer I come up with is that M$ released their numbers first and then Sony's PR team decided using vertices/second rather than polys ( and most people don't even know the difference ) looks much better on paper.

I want to see how many polys the PS3 an push OR how many vertices the Xbox 360 can push. Comparing polys to vertices seems designed purely to misslead.

This is silly. Read the links posted above by aaaaa00.

You'll often hear companies claim their peak vertex rate as a peak triangle rate. For that reason I actually think quoting vertex figures is more accurate - there's less ambiguity about what a vertex is. But I've no doubt that when you see a company quote x triangles per sec, that means x vertices per sec. Why would they quote x triangles per sec, if it's actually x vertices per sec, when they could quote so many more triangles per sec with the same number?

(Hint: they're not looking to avoid misleading you, and companies quoting x triangles per sec when that's actually x vertices per sec also are not misleading you)

hasanahmad - read the links posted above for more, perhaps better examples. Or for example this:

As the number of triangles increases, the closer you get to 1:1.

For example - from http://www.mpi-sb.mpg.de/units/nwg3/ws0405/cg/exercises/sheet03-solution.pdf:

Imagine you have a triangle mesh with 70000 triangles, which you render with 700 triangle strips.How many vertices do the strips have altogether? Suppose you optimize the number of triangle strips to 150, how many vertices do you then have? What does this mean for the rendering speed?

Solution: As already said, each triangle requires one vertex, and each triangle strip requires two additional initial vertices. So in the rst case, we have to stream 70:000+700 2 = 71,400 vertices, in the second case 70,000 + 150 2 = 70,300 vertices to the graphics card.

As you can see, the ratio is virtually 1:1

 

[Wunderchu]

to draw a triangle you draw 1 triangle with 3 vertices, then a line (1) more vertex


if you take 2 squares and cut them into 4 triangles, will have 6 vertices, every square will have 4 vertices, every triangle will have 3,

yes, true

 

[aaaaa00]

That is true, but is it fair to simply not account for the increased complexity of those vertices?

The number for Xenos is purely a triangle setup limit. Going all out on vertex shading, Xenos has enough ALU power to run dozens of instructions per vertex and still be maxxing the triangle setup rate.

If it wasn't setup limited you could crank 6 billion vertices per second through the ALUs if you limited it to a simple 4 instruction transform that does absolutely nothing else.

BUT the point is that it is expected that devs will use vertex shaders much more complicated than just basic transform, which means the setup limit choosen is high enough that more setup hardware would just be wasted silicon.

In any case, I don't really expect either chip to be setup limited in all but the most ridiculous manufactured benchmark scenario, so comparing 500 million triangle setups/s to whatever RSX can do is not particularly useful.

 

[Bohdy]

Actually, in an ideal mesh topology, 1vertex = 2triangles.

Can you demonstrate such a mesh, faf? I'm having some trouble picturing it.

 

[hasanahmad]

if you look at PS2, its max poly/s was 75 million but the most it achieve was with Jax i believe which was around 10-15 million. Does this mean that if ATI are telling truth that they can being efficiency up to 90%, we can see polygons pushed up to 450 million triangles per second in the game?

 

[Shifty Geezer]

That is true, but is it fair to simply not account for the increased complexity of those vertices?

Yes. That's the way GPU's work. You provide a string of vertex data and it assembles them into triangles, using three vertices at a time, sharing previously used vertices if needs be. The number of edges attached to a vertex is irrelevant - setup is not influenced by vertex 'complexity', basically because there's no such thing as complexity. Every vertex is just a point in space.

 

[Urian]

I don´t Know, but since Xbox can do 116.5 milion vertices with 2 Vertex Shaders running at 233Mhz I Believe that PS3 can do 2200 milion vertices per second.

 

[hasanahmad]

Yes. That's the way GPU's work. You provide a string of vertex data and it assembles them into triangles, using three vertices at a time, sharing previously used vertices if needs be. The number of edges attached to a vertex is irrelevant - setup is not influenced by vertex 'complexity', basically because there's no such thing as complexity. Every vertex is just a point in space.

yes but is this how a unified architecture GPU works? with proposed 90% efficiency? and completely different from any GPU made? comparitive to RSX which will work like the 7800 in terms of architecture

 

[hasanahmad]

I don´t Know, but since Xbox can do 116.5 milion vertices with 2 Vertex Shaders running at 233Mhz I Believe that PS3 can do 2200 milion vertices per second.

the max i think xbox got was 50-60 million polygons in real game performance


also in game graphics i dont think quadrilaterals apply, triangles do.

 

[Bohdy]

the max i think xbox got was 50-60 million polygons in real game performance

Yeah right.

 

[Shifty Geezer]

yes but is this how a unified architecture GPU works? with proposed 90% efficiency? and completely different from any GPU made? comparitive to RSX which will work like the 7800 in terms of architecture

The way it works is still exactly the same in terms of triangle setup. The difference is the portioning of ALUs to process the data. On Xenos same as every other GPU meshes will be constructed from vertex data, where in an optimized mesh you are at around 1 vertex for each triangle.

You need to appreciate that US is not a marked difference in rendering technology to what's already out there. What Xenos adds is tile based rendering and flexible ALUs. The way it works, the way it textures and shades polys and renders pixels, is still the same as any other GPU.

 

[Wunderchu]

Yeah right.

heheh .. I also think the real world number is less


AFAIK, most ppl say that in actual Xbox games, the max. has been about ~20 to ~30 million polygons/sec.

 

[hasanahmad]

The way it works is still exactly the same in terms of triangle setup. The difference is the portioning of ALUs to process the data. On Xenos same as every other GPU meshes will be constructed from vertex data, where in an optimized mesh you are at around 1 vertex for each triangle.

You need to appreciate that US is not a marked difference in rendering technology to what's already out there. What Xenos adds is tile based rendering and flexible ALUs. The way it works, the way it textures and shades polys and renders pixels, is still the same as any other GPU.

i know the way the polys will be made is the same, im asking the way they are output, with more efficiency, Huddy and the Beyond3d article both say 90% in real performance

 

[SanGreal]

i know the way the polys will be made is the same, im asking the way they are output, with more efficiency, Huddy and the Beyond3d article both say 90% in real performance

The shaders are more efficient in that the ALUs don't waste as much time idle.

It has nothing to do with the way they are output.

 

[Fafalada]

Can you demonstrate such a mesh, faf? I'm having some trouble picturing it.

I guess the simplest example would be a sphere.

Basically, so long as you have a closed solid that is not dominated by discontinuities, you will get more then one polygon per vertex.

 

[hasanahmad]

The shaders are more efficient in that the ALUs don't waste as much time idle.

It has nothing to do with the way they are output.

the efficiency described in the article pointed to performance rather than time

 

[SanGreal]

the efficiency described in the article pointed to performance rather than time

I'm really not sure what you're trying to say