PS3 / 7 times faster than PCI Express.

[Geo]

2. Do you honestly belive that nAo from ninja theory, who is now more or less a first party developer for Sony is allowed to say that the Xenos is better?

I'm very puzzled about this comments relevance to the part that was quoted, and thus assumed to be addressing. No, you're not going to get exclusive ISV's to make that kind of generic statement that you suggest from either side. However, that's not what was going on there either. It was a specific piece being addressed (framebuffer effects), and there are a multiplicity of ways to answer such a question without stepping on any political landmines if he really did perceive Xenos to have an edge there. The time-honored traditional one would be to point somewhere else entirely different where you have an advantage and note that both parts have different strengths, yada yada. The answer given shows he didn't feel there was any need to do so --i.e. that he feels PS3 can go toe to toe there without any need to engage in the typical rhetorical techniques for sidestepping a competitors advantage in a given area.

Sure, "reading between the lines" is a very necessary skill at times in the community/industry in which we swim in --but sometimes "a cigar is just a cigar" too, and that looked like a cigar to me. . .

 

[dskneo]

i really cant see where x360 has BW advantage. this could only happen if i twisted my mind to think that the Edram could do a miracle or two, instead of what she was made to do in the first place (already mentioned in this thread)

so how can a x360, with 22gb/s splitted between CPU and GPU be less bandwidth limited than a Ps3 that has 25gb/s dedicated to cpu and 22gb/s dedicated to GPU... must be the edram miracle many believers are playing like an Ace card.


and yes, lost planet has the most impressive special effects i've seen to date, i grant x360 that!

 

[Jawed]

The time-honored traditional one would be to point somewhere else entirely different where you have an advantage and note that both parts have different strengths, yada yada.

If you dig up the thread then you will find that Marco did indeed do this

Jawed

 

[Dave Baumann]

RSX could potentially boast similar bandwidth figures by counting in frame and Z buffer compression when doing AA instead, or even early Z-rejection and hidden surface removal stuff - something xenos does not feature at all btw

There's a Heirarchical-Z buffer just post triangle setup that can reject multiple pixels prior to them even being shaded.

BTW - something else to consider is MSAA fillrate. With Xenos the fillrate isn't altered when MSAA is enabled, and it maintains its double Z rate as well, enabling 64 Z samples per clock. G7x can do full colour write speed at 2x AA, but loosing its double Z advantage, and halve the fillrate at 4x.

 

[Zeross]

Yep, those features are there but it seems early rejection granularity it's not very good

Latest figures I've seen were 16 z-samples per block in the hierarchical Z-Buffer, this granularity seems pretty good to me (one quad with 4x AA, two quads with 2x AA or four quads with no AA).

 

[nAo]

Latest figures I've seen were 16 z-samples per block in the hierarchical Z-Buffer, this granularity seems pretty good to me (one quad with 4x AA, two quads with 2x AA or four quads with no AA).

I've heard completely different numbers (16 X ). but who knows, maybe you're right

 

[Zeross]

I've heard completely different numbers (16 X ). but who knows, maybe you're right

256 samples per block ? I've already seen this number cited here by Fran : http://www.beyond3d.com/forum/showpost.php?p=786740&postcount=66

On the other hand in his article on the Xenos Dave gave the same figure than me :

. In Xenos's case the Hierarchical Z Buffer stores down to 16 sample groups, which equates to 2x2 pixel groupings with 4x FSAA enabled.

So indeed who knows ?

 

[Fafalada]

So indeed who knows ?

Well the docs are a bit contradicting on some parts in question, so without testing it on actual hw I'd hesitate to claim one way or another.

Although looking at how the stated H-Z memory size is pretty gargantous - I'd sooner believe that granularity HAS to be fine.

 

[Fran]

Yep, those features are there but it seems early rejection granularity it's not very good

Indeed. There's nothing like Early-Z on a per pixel level, which gives some problems on high frequency alpha-tested geometry. Hi-Z granularity is 64 fragments. Hi-Stencil granularity is also 64 fragments (and not 256 as I once wrote).

 

[Jawed]

I think ATI's PC GPUs have one-quad (four fragment) granularity. That comes from R420 documentation, so it may have changed with R5xx...

Does any GPU do per-fragment Z-rejection? I don't see how this can make sense, because the pixel pipeline is forced to work on quads. Per-pixel rejection can only happen in the ROPs.

The reason why Xenos's hierarchical Z is so coarse is that the SIMD arrays are 16-wide - 4x the normal size of a PC GPU. Plus, with the batch size being 64 fragments, there's no point in making hierarchical-Z any more fine-grained.

Jawed

 

[Zeross]

Indeed. There's nothing like Early-Z on a per pixel level, which gives some problems on high frequency alpha-tested geometry. Hi-Z granularity is 64 fragments. Hi-Stencil granularity is also 64 fragments (and not 256 as I once wrote).

As far as I know, the Xenos can test and reject up to 64 fragments per cycle (an 8 x 8 fragment block and this is the same with all GPU from ATI since a while), but the resolution of the Hierarchical Z-Buffer is more precise.

Anyway Hi-Z granularity is neither better nor worse on the Xenos than other GPU (at least from ATI, I don't know the granularity of Hi-Z on NVIDIA GPU) )