How many Polys can the New Consoles Push?

[Shifty Geezer]

Why don't you agree with him? You were so QUICK to agree with him when you thought that 1.1 B vertices would = 2.2 B tri's.....hmmmmm....

There has been talk in this thread about a 1:1 ratio for polys to vertices but those have only been under rare circumstances such as rendering a sphere or pyrimid.

I didn't agree with him. I was just pointing out where you were wrong in your calculation based on dotproduct's assertion. You divided by two, instead of multiply by two, which wasn't what he was saying. Notice I did say 'if dotproduct's assertion is right'.

The 1:1 ratio is NOT rare and considering the number of replies to this it almost seems as though you're unwilling to accept the point. It's a matter of fact that in an optimised mesh, every time you add a triangle you only need add one vertex to those already in place. That's all 3D models. The only time you don't achieve that is when you have a triangle connected only on one edge, or no edges, which only really appear in quads. I'd expect say a tree made of several textured plane to be nearer 2 verts by triangle. I don't know how many people have to explain this before you will accept it. :?

 

[rendezvous]

As for V-E+F=2
V is the number of vertices,
E is the number of edges,
F is the number of faces.

In a triangle-mesh a vertex is shared by about 6 faces.
Each face contains 3 vertices which gives about 2 faces/vertex.

 

[DotProduct]

You have to differentiate between the vertex triangle ratio on the actual mesh and the ratio on the triangle strips that actually go through the rendering pipeline.

As I stated the mesh has a ratio of 2:1; faces:vertices.

But when you send the polygons down the pipeline you get a ratio of 1:1 (triangle strips).

Basicly every vertex is transformed twice and you had to cache every vertex to circumvenit this.

 

[BenQ]

I didn't agree with him. I was just pointing out where you were wrong in your calculation based on dotproduct's assertion. You divided by two, instead of multiply by two, which wasn't what he was saying. Notice I did say 'if dotproduct's assertion is right'.

The 1:1 ratio is NOT rare and considering the number of replies to this it almost seems as though you're unwilling to accept the point. It's a matter of fact that in an optimised mesh, every time you add a triangle you only need add one vertex to those already in place. That's all 3D models. The only time you don't achieve that is when you have a triangle connected only on one edge, or no edges, which only really appear in quads. I'd expect say a tree made of several textured plane to be nearer 2 verts by triangle. I don't know how many people have to explain this before you will accept it. :?

No, my calculation wasn't wrong. Yours was. Unless you seriously believe that the PS3 can push 2.2 billion polys/second

My calculation shows that the PS3 has a very slight advantage in poly pushing, your calculation shows the PS3 being able to push more than 4X's as many polys as the Xbox 360...... not bloody likely.

 

[robofunk]

Show me a game that does more then 1/3 the theoretical max.

 

[London Geezer]

Show me a game that does more then 1/3 the theoretical max.

Oh jesus it's like talking to kids.

After i show you, what r u gonna do? count how many polygons and how many verts are in the picture?

You have lots of people, among which developers, telling you like it is, so if you and Benq don't get it, it's your problem.

Only a game with every single polygon being detached from every other polygon will have a 3:1 verts to polys ratio.

 

[BenQ]

Oh jesus it's like talking to kids.

After i show you, what r u gonna do? count how many polygons and how many verts are in the picture?

You have lots of people, among which developers, telling you like it is, so if you and Benq don't get it, it's your problem.

Only a game with every single polygon being detached from every other polygon will have a 3:1 verts to polys ratio.

I never claimed that the ratio was 3:1. But I don't think it's 1:1. I agree with Dot Product's 2:1 numbers.

 

[DotProduct]

Only a game with every single polygon being detached from every other polygon will have a 3:1 verts to polys ratio.

And that is why Euler is only correct for mainfold meshes.

The typical mesh of any game should be mainfold, though - with the exception of particle systems, tree leaves, grass == Billboards.

Actually you can load up any mesh in a 3D viewer and check the ratio.

 

[robofunk]

I just asking for a real life game that pushes more polygons then 1/3rd the max (even 3dmark 2001 always has less the 1/3rd). I'm just saying that 1/3rd is much closer to what we can expect even if it's not what is mathematically provable. I'm not saying 2:1 isn't true in a sense, 3:1 is what I expect.

 

[Jabbah]

I never claimed that the ratio was 3:1. But I don't think it's 1:1. I agree with Dot Product's 2:1 numbers.

But you are getting mixed up with way round of the ratio. DotProduct was saying 2 Polys to 1 Vertex.

2 Vertices : 1 Poly is just as improbable as the 3 vertices : 1 poly as it will only happen when you have 2 connected triangles.

In most cases it will tend towards 1:1 as in the case of triangle strips where the number of polys = number of vertices - 2, for large strips this tends to 1:1

The 2:1 (Polys:Vertices) is a theoretical maximum that can be achieved when storing a mesh.

 

[Shifty Geezer]

Re-look at dotproduct's statement. You'll find you've read it wrong.

Actually it is quite easy to derive that the number of polygons is about twice the number of vertices for every manifold 3D mesh.

The number of polys is about twice (2x) the number of vertices.
Now your number...

So the Ps3 can push 1.1 billion vertices

According to dotproduct, polys = 2x vertice. Therefore PS3, if it can shift 1.1 billion vertices, can, by this equation, push 2x as many triangles, which is 2.2 billion.

That's the bit you've got wrong. You divided by two, but 2 triangles a vertex means you need to multiply by two for a given number of vertices. And no, I don't think PS3 manages 2.2 billion vertices. Dotproduct explained this just now as, where the mathematical state of 2:1 tri's:verts is true in the model, when passed to a GPU as tringle data it comes out at 1 triangle for each vertex. Which is what everyone's said. And I'll leave it at that.

In summary, triangles per second and vertices per second are interchangeable when talking about GPU performance.

 

[Gollum]

BenQ, I get the impression you're confusing the numbers. DotProduct is talking about 2:1 face/vert, while you're arguing 2:1 vert/face! You actually don't need much experience in 3D to ralize you're arguing a lost case. Do as he suggests, get a 3D Object Viewer and download some sample meshes. Almsot all should be somewhere between 1:1 and 2:1 fece/vert...

A 2:1 face/vert ratio, as Faf pointed out, is a best case scenario. While something that should always be a goal and not exactly being unrealistic, anything more complex than a tesselated sphere tends to loose some efficiency though.

A complex character mesh with attached clothing, gadgets and weapons can only dream of achieving that 2:1 ratio. The more "extras" and extrusions you add to an otherwise optimal mesh (base character model), the closer you get to a 1:1 ratio in my experience. Non-organic, mechanical and environment meshes are often highly irregular, not closed and thus often inefficient too. This generation game meshes will become a whole lot more complex, so the 2:1 poly/vertex ratio might end up being little more than an aesthetic.

Thanks for clarifying the hardware setup limitations involved DotProduct. So even if we have an optimal mesh (tesselated sphere), we'd still only get a maximum throughput of 1:1 face/vert due to the way graphics pipelines handle triangle strips?

 

[BenQ]

Re-look at dotproduct's statement. You'll find you've read it wrong.
The number of polys is about twice (2x) the number of vertices.
Now your number...

According to dotproduct, polys = 2x vertice. Therefore PS3, if it can shift 1.1 billion vertices, can, by this equation, push 2x as many triangles, which is 2.2 billion.

That's the bit you've got wrong. You divided by two, but 2 triangles a vertex means you need to multiply by two for a given number of vertices. And no, I don't think PS3 manages 2.2 billion vertices. Dotproduct explained this just now as, where the mathematical state of 2:1 tri's:verts is true in the model, when passed to a GPU as tringle data it comes out at 1 triangle for each vertex. Which is what everyone's said. And I'll leave it at that.

In summary, triangles per second and vertices per second are interchangeable when talking about GPU performance.

I don't think either of us are right.

 

[DotProduct]

I'm not saying 2:1 isn't true in a sense, 3:1 is what I expect.

The numbers have nothing to do with each other.

2:1 is the face vertex ratio for a mesh and your 3:1 is a supposed efficiency ratio (and a very optimistic one).

I do not think that vertex operations will ever be a bottleneck with XBox 360/PS3. The bottlenecks will probably be pixel shaders, the old rasterization/fillrate/z-buffer test bottleneck and of course memory/texture/framebuffer bandwidth.

Thanks for clarifying the hardware setup limitations involved DotProduct. So even if we have an optimal mesh (tesselated sphere), we'd still only get a maximum throughput of 1:1 face/vert due to the way graphics pipelines handle triangle strips?

Yes...

But the 2:1 rule is true (~) for every mainfold mesh and not only for regular ones. A regular mesh has a valence of 6 for every vertex and is usually seen as: "good, but not worth the trouble".

 

[BenQ]

So, after all this, anyone care to take a stab at how many polys/second each next gen console can push.... that was the point after all.

 

[Gerry]

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

Probably the easiest example to visualise wouldn't be something like a ten-sided dice (like the ones used in D&D). They'd be modelled with only 7 vertices, but consist of 10 triangles. This is of course just a very simplistic sphere if you think about it.

 

[version]

max visible polycount on 1280*720P is 1mill /frame, 60 mill/sec
max invisible polygons 1 billion/frame 60 billion/sec on ps3
xenos ? idont know

 

[Jabbah]

So, after all this, anyone care to take a stab at how many polys/second each next gen console can push.... that was the point after all.

I would estimate a realtistic ratio would be 90-95p:100v

So if you have 1billion vertices/sec then you will have a "realistic" theoretical maximum of 900-950 million polygons/sec

 

[DotProduct]

(Maximum) Polycount (as always) depends on the case/goal.

Usually you do not need that much polygons for architecture and stuff like that.

Polygons you usually need for your characters and those need to be skinned, making character polygons extra expensive.

Afaik in 720p you have (~) 1million pixels. This means 1m polygons per frame + no overdraw =>~ 1 pixel=1 polygon.

Why do we have textures and pixel shaders then?

In reality we have to deal with overdraw and the polygons won´t be even distributed, but 1m polygons (in the frustrum) are still a huge number.

An intersing number would be how many 100pixels triangles the GPU can rasterize or z-buffer check only.

 

[Bohdy]

As for V-E+F=2
V is the number of vertices,
E is the number of edges,
F is the number of faces.

In a triangle-mesh a vertex is shared by about 6 faces.
Each face contains 3 vertices which gives about 2 faces/vertex.

Thanks, that (in tandem with a picture of a mesh I googled) helped me see how an average of 2:1 faces/vertices was possible. Basically every vertex becomes the hub of 6 different faces.

And BenQ: no need to be difficult about getting the numbers mixed up, just move on. And this isn't about proving that one console will perform 2X another or whatever, but merely finding the uppermost peak limits of the GPU's. Not very useful, but interesting enough.