How many triangles can draw 1 VMX unit in Xenon ?

[Love_In_Rio]

It is said that Xenos will be manage up to 500 million polys with effects, but these polys must be created in the CPU. Could be enough only one of the VMX units to feed Xenos 500 million polys capacity ? If so, Xenon CPU will have allways at least two VMX units free for other tasks such as phisics ?
I am also thinking about the ability of a Xenon VMX to process near 3,2 billion dot products per second.If Xenos has a limit of 500 mill. polys wouldn´t it be also enough ,for example, 1 VMX for phisics or lightning ( i am supposing a dot product help in these tasks ) ?

 

[nAo]

It is said that Xenos will be manage up to 500 million polys with effects, but these polys must be created in the CPU.

I don't know where you get this idea but Xenos don't need CPU intervention at all and CPU doesn't not have to create anything.

 

[Love_In_Rio]

But is not said that Xenos can´t create vertices ?

 

[nAo]

Xenos works as any GPU out there, it loads vertices from system memory.
These vertices can be prebacked or CPU generated, most of the time they're prebacked.
Usually CPU handles special cases such as procedural geometry.

 

[London Geezer]

But is not said that Xenos can´t create vertices ?

Maybe you should read articles on how graphics architectures work. Start from the basics.

 

[Love_In_Rio]

Ok, thanks nAo. And what about the dot products question ?.


P.D: lelo-boy ignore me. With your comments everybody should start from the basics everyday. Try to be constructive sometime please

 

[London Geezer]

Ok, thanks nAo. And what about the dot products question ?.


P.D: lelo-boy ignore me. With your comments everybody should start from the basics everyday. Try to be constructive sometime please

Err i was actually trying to be constructive. Reading articles, from the basics, on the subject will help you much more than posting here. Obviously the message was lost in translation.

If i offended you, sorry it wasn't my intention. You should tone the arrogance down though.

 

[Acert93]

L-B is right though. T&L and Vertex Shaders have been a part of GPUs since the advent of GPUs (GF 256). Basically triangle setup and vertex processing has been a task of GPUs since 1998/1999.

What Xenon offers is the oppurtunity to do some neat new stuff. e.g. It can take a Higher Order Surface--which is smaller than a prebaked triangle mesh--and use the CPU to do dynamic LOD and push that data to the GPU. This would save memory space AND bandwidth. That is one of the benefits of streaming. Basically it can be offloading a lot of the vertex work to the CPU, which in theory would allow more of the GPU to spend time on pixel shading.

But as Xenos has unified shaders and even a hardware tesselator (I think Dave said it can do 250M triangles) the GPU can take a HOS directly, tesselate it, and do all the processing also leaving the CPUs to do their own thing.

Basically it is a really flexible design that should meet different developer needs and goals. A game like PGR3 may be really graphics oriented and thus use the CPUs for a lot of proceedural synthesis, where a Forza may cut back on the CPU-graphics load and instead use the CPUs for more AI, physics, etc... and have the Xenos do most of the heavy vertex lifting.

We just need to wait and see how 2nd and 3rd generation games begin to use the system and see what options are realistic for game development. Everything sounds great in theory, but reality often has a bigger bite. Some of these new techniques may be underwhelming, so it is wait and see.

 

[Love_In_Rio]

l-b, the only person in this forum that makes my arrogance to spread out is you, and your "ironic" comments. I think many people avoid writing in this forum by you, as you behave as a barking dog whenever you have the opportunity. Many people agree with me, no doubt about it. So, from now:

MODE lelo-boy IGNORING ON.

 

[London Geezer]

Err sure kid. Obviously even a

If i offended you, sorry it wasn't my intention.

wasn't enough.

 

[AlNom]

But as Xenos has unified shaders and even a hardware tesselator (I think Dave said it can do 250M triangles) the GPU can take a HOS directly, tesselate it, and do all the processing also leaving the CPUs to do their own thing.

I thought you tesselate a lower order surface to create a higher order surface.



Maybe I just have my definitions mixed up.

 

[Simon F]

I thought you tesselate a lower order surface to create a higher order surface.



Maybe I just have my definitions mixed up.

You do have your definitions mixed up.

 

[pc999]

To Love in Rio: there is soe articles that can help you somehere in the front page (I hope I am not being arrogant )

Anyway, I also dont know how those 9.6B of dot products translate on what we see on the screnn, I I am not wrong it is some thing good to calculate directions in 3D, if so it should be good for physics, lightnig/shadows, refletions etc... (raytracing , jk), but I would like for some clarafication/correction also, if possible.Thanks.

 

[Love_In_Rio]

No, jeje, ,thanks.

 

[AlNom]

You do have your definitions mixed up.

Ah right.

Thanks for the "corrections"

 

[Edge]

A simple question seems to generating a lot of nonsense in this thread.

The CPU is not needed to transform polygons, as the X360 GPU has Arithmetic Logic Units (ALU's) that can perform transformation and lighting, and also pixel rendering, due to the dual nature of the ALU's.

The X360 GPU is rated at 500 million polygons per second, and I can only assume that's if all 48 ALU's are all doing polygon transformation? Hmmm, seems a bit low for that many ALU's running at 500 MHz.

I believe the upcoming Nvidia 7800GTX will be rated higher with only 8 vertex engines?

 

[ban25]

A simple question seems to generating a lot of nonsense in this thread.

The CPU is not needed to transform polygons, as the X360 GPU has Arithmetic Logic Units (ALU's) that can perform transformation and lighting, and also pixel rendering, due to the dual nature of the ALU's.

The X360 GPU is rated at 500 million polygons per second, and I can only assume that's if all 48 ALU's are all doing polygon transformation? Hmmm, seems a bit low for that many ALU's running at 500 MHz.

I believe the upcoming Nvidia 7800GTX will be rated higher with only 8 vertex engines?

It can set up 1 triangle per clock cycle and it cycles 500 million times per second.

 

[Rockster]

You're confusing the setup limit with vertex processing power. Vertex processing usually involves transforming the vertex data from world coordinates to camera relative coodinate space and typically also assigns light values to each vertex, along with potentially many other things. Triangle setup involves taking these vertices that are in camera space and map them to 2d screen locations. Setup limit is 500million polygons/sec and the hardware responsible for this is not the vertex engines.

 

[Edge]

> "It can set up 1 triangle per clock cycle and it cycles 500 million times per second."

OK, so that's 48 x 500 million = 24 billion calculations per second, and I assume 4 input multiply (4 x 32-bit values, with the 4 other values in the register) to generate a vertice per second? Assuming 2.5 vertices per triangle on average, that's 24 billion divide by 2.5 = 9.6 calculations per second for 500 million triangles per second.

Takes almost 20 cycles to generate one triangle per ALU?

> "Setup limit is 500million polygons/sec and the hardware responsible for this is not the vertex engines."

That's what I thought, but they are called transformation and lighting engines. Sounds like setup engines only? By hardware, I assume you mean the CPU. So how many vertices or polygons per second on one VMX unit as per the first post of this thread?

 

[ERP]

> "It can set up 1 triangle per clock cycle and it cycles 500 million times per second."

OK, so that's 48 x 500 million = 24 billion calculations per second, and I assume 4 input multiply (4 x 32-bit values, with the 4 other values in the register) to generate a vertice per second? Assuming 2.5 vertices per triangle on average, that's 24 billion divide by 2.5 = 9.6 calculations per second for 500 million triangles per second.

Takes almost 20 cycles to generate one triangle per ALU?

Your calculations are way off and the whole argument is moot anyway. You're just not going to be able to get an apples to apples comparison.

As far as your math goes.

Your going to need about 4 ops for any sort of useful vertex shader and most meshes average closer to 1 vert per tri than 2.5. You can actually approach 0.5 verts/tri with the post transform cache on closed primitives (notably a torus).

So that would either be roughly 6 billion transformed verts/second which could be as many as 12 billion tri's if you want to stretch the numbers.

But these are copmpletly meaningless numbers......

Even if this wasn't transform limited you couldn't do any pixel shader ops at peak vertex throughput, so what's the point....