Shaders performance and fillrates.

[patroclus02]

Well, after searching about this, I can't find the answer..

Vysez told me not to compare two arquitectures, But I need to compare 2 arquitectures in order to explain myself. But the question itself is not about Xenos or anything else, but shader's performance. I hope this time my post is correct.

Imagine Xenos is dedicating 8 shaders for vertes and 40 for pixels.
While, RSX, using 8 vertex shaders and 24 pixel shaders.

Xeon
8 x 2 = 16 vertex shader ops
40x2 = 80 pixel shader ops

RSX
8 x 2 = 16 vertex shader ops
24x4 = 96 pixel shader ops (no texturing, 4 ops per pipeline, not counting the fith operation, which is a normalization op I think).
24x2 = 48 pixel shader op (texturing, 2 ops per pipeline, plus a texture addressing op).

But, getting into fillrates.
Xeon
40*500 = 20 Gpixels/s
RSX
24*500 = 12 Gpixels/s

When NO texturing, RSX is capable of more pixel ops, but, at the same time, delivers less fillrate. How should I see this? Xenos is capable of working on more pixels, but applying less operations to them? If so, then, fillrate is not a objective performance indicator for pipelines.
I’ve seen for years that all graphic’s card review talk about fillrates, and not shader operations. This is something I started to see in recent reviews. Why is this?

This is, more or less, what I wanted to get clear… Thanks

 

[MasterDisaster]

Well, after searching about this, I can't find the answer..

Vysez told me not to compare two arquitectures, But I need to compare 2 arquitectures in order to explain myself.

search better, shaders detail are well discussed

and cracking an admin Closed Thread, opening another is not a good thing.

PS. you are wrong with the numbers, you are wrong making xenos a fixed PS-VS (40/8), you are wrong doing the math thinking that there're no textures.

 

[PeterT]

Your "fillrate" calculations are completely wrong. In modern GPUs, actually writing to the framebuffer is decoupled from the fragment "pipelines" you are referring to. The amount of pixels that can be written is primarily bounded by bandwidth, but also by the number of ROPs (raster operation units). Both RSX and Xenos (not 100% sure about the latter) actually have "only" 8 of those. But that matters little when taking real-world bandwidth constraints into account - though those should basically be non-existant on Xenos. Anyway, it's all slightly more complex in practice than just multiplying some numbers

[edit]And I believe a thread like this is totally okay generally as it poses a serious technical question - though this one is suspect because it asks something that could be answered by some diligent searching and reading.

 

[patroclus02]

I'm sorry, I really wanted to be correct. I'll try not to do again.
I don't write this kind of post for laziness...

Sure, I searched and searched for hours. I really learnt all I know this way. But no one ever explained anything to me, and that's why I need to do this kind of questions, just to make sure I know what I think I know. If it is wrong, I apologize.

If you guys know any place I can read all of this, please tell me. I'll pleased to read anything even if it is long, and then come back with real and interesting questions.
I read "Xenos Demistified" article, but I still don't get what it's being discussed here. Maybe any review of any new graphic card is a good "theory pool" ?? I'll try to search more, but if you guys have at hand any good resources, then I'll be thankful.

 

[Shifty Geezer]

I think a question you need ot answer before you get more help is 'why do you need to know?' The reason being, shader-ops don't in themselves mean anything concrete. What exactly is being counted as a shader op? So even if you get exact ShOP figures, that information doesn't tell you a great deal.

It's also worth mentioning that Xeno's shader work in groups of 16, so you won't get an 8/40 split as you suggested. You'd get either a 0/48 or a 16/32 split if there isn't much vertex work to do.

The shader ops also have nothing to do with fillrate. You could burn all the shader ops on a single pixel if you wanted to be crazy! Shader ops are how many operations you can apply per second to vertex/pixel data, which is independent of fillrate which is how many pixels you can render.