xbox360 gpu explained...or so

[Jawed]

Clearly a great example of vertex fetch limited rendering is a stencil shadow rendering pass

Lay off the drugs, Please.

In fact the shader unit is vertex limited as there shouldn't be anything to be done for fragment shading. However the GPU pipeline as a whole isn't likely to be (vertex) shader limited but fillrate (z and stencil) limited .

Precisely.

Though it would be sort of entertaining to discover that Xenos is actually vertex fetch limited, even including fill-rate, in this scenario. 32 gigazixels per second?...

Jawed

 

[nAo]

Clearly a great example of vertex fetch limited rendering is a stencil shadow rendering pass. Xenos should be very good at those...

Stencil shadows rendering is vertex fetch limited? are you serious!? You must be kidding..

The R500 should be a monster at this. 64Z per clock and all 48 ALUs allocated to vertices. (is it possible?)

I hope it is! Anyway Jawed wrote about a "vertex fetch limited" case,
even if R500 could fill 1024 zixels per clock cycle I doubt it would be vertex fetch limited
Obviously when the GPU is rendering stencil shadows one would assign all ALUs to vertex shading.

 

[Jawed]

If the CPU can't generate the vertices fast enough the GPU is vertex fetch limited, is that correct?

If not, what's the correct term?

Jawed

 

[fek]

If the CPU can't generate the vertices fast enough the GPU is vertex fetch limited, is that correct?

That's partially correct. The pipelines will be vertex fetch limited. The rest of the GPU might be limited by other factors depending on the situation.

If not, what's the correct term?

You spoke of "vertex fetch limited rendering" which is not entirely correct. But we are all being a bit sloppy (it's a forum); pointing your imprecision out is just trolling imho, the concet behind it was correct.

 

[Rockster]

Has ATI given any indication of what kind of efficiency to expect from the architecture under typical workloads?

 

[Fafalada]

If the CPU can't generate the vertices fast enough the GPU is vertex fetch limited, is that correct?

Well if CPU is the only thing generating vertices, that would be correct, but rendering stencilshadows it's virtually impossible to get into a situation when that happens.
Unless your CPU royally sucks really.

 

[fek]

Has ATI given any indication of what kind of efficiency to expect from the architecture under typical workloads?

No, AFAIK.

 

[Jawed]

Well I was referring to the z-only pass for stencil shadowing, when the viewport of the render coincides with the light's illumination frustum.

Jawed

 

[nAo]

Well I was referring to the z-only pass for stencil shadowing, when the viewport of the render coincides with the light's illumination frustum.

Are you talking about shadow mapping or stencil shadows here?
Stencil shadows are not rendered from some light space.

 

[DemoCoder]

The Xenos has a tesselation unit that can generate 250Mtris/s plus a geometry shader to assist, plus a tri-core 3.2Ghz CPU tied to the GPU via L2 cache, so I can't see any situations where the GPU will be starved for new primitives.


On another note, fek/nAo, you guys should chill a little bit. I think there's a big ego argument going on over sloppy language. I think you guys are mostly on the same page, you just don't see it.

 

[Jawed]

I'm referring to algorithms like D3's shadow calculation pass, which does a z/stencil pass based on each light, as I understand it.

Jawed

 

[DemoCoder]

I doubt the calculation of the volumes will be CPU limited on XB360, moreover, the geometry shader might be able to assist. Hell, the XB360 could use 3 threads (or 6 if you want) to calculate the geometry for 3 lights simultaneously.

 

[Jawed]

To be honest, I doubt it too.

But still Xenos appears to have 8x the vertex shading power of X800XTPE (in terms of pipelines) in this scenario and when you throw multiple lights (5, 10?) into the mix, it should get quite interesting.

Damn, why hasn't ATI got some sweet videos of some techy stuff like this on their webby?!!!

Jawed

 

[nelg]

I am beginning to think that ATI should have named the R500 the Enigma.

 

[DemoCoder]

Yep, I was thoroughly disappointed with the MTV-style presentation Microsoft used vs the "industry showcase" style presentation Sony used, with technical demos and technical powerpoints. XB360 would have faired a little better had they showed some tech demos of what could be done (in terms of extremes and peaks, like they do with CELL) instead of having J Allard get a hipster makeover, sit in cool 'indian style' on the floor, showing off all these game titles many of which looked like they were Xbox1 ports.

The person who had this "lets dress everyone up with like MTV kids because this is how you 'take it to the customer directly' should be fired" I look forward to J Allard's next big demo where he shows off the result of his classes in Krump dance.

 

[Geo]

Don't repress, DC, let it flow.

Can't imagine why "DemoCoder" is miffed at abandoning the industry-insider model for the consumer-friendly one.

But you might be right about the impact.

 

[Rockster]

Jawed, what do you mean by interesting? I don't see how multiple lights should create any issues what so ever.

 

[DemoCoder]

There's a difference between consumer-friendly, targeting educated individuals from 18-35 and targeting MTV teens with brain-rot.

 

[Love_In_Rio]

Yep, I was thoroughly disappointed with the MTV-style presentation Microsoft used vs the "industry showcase" style presentation Sony used, with technical demos and technical powerpoints. XB360 would have faired a little better had they showed some tech demos of what could be done (in terms of extremes and peaks, like they do with CELL) instead of having J Allard get a hipster makeover, sit in cool 'indian style' on the floor, showing off all these game titles many of which looked like they were Xbox1 ports.

The person who had this "lets dress everyone up with like MTV kids because this is how you 'take it to the customer directly' should be fired" I look forward to J Allard's next big demo where he shows off the result of his classes in Krump dance.

They will do it in the X05 in september. No doubt about it.

By the way, how would this stuff manage the unreal engine 3 ? better than a conventional graphics chip ? And doom 3 engine ?. I think that in the end the differences will be settled in the way RSX and R500 behave above all with unreal engine 3.

 

[psurge]

Guys, check out this thesis paper - it deals with reducing the penalty of conditional branches in data parallel architectures (stream processors) :
ftp://cva.stanford.edu/pub/publications/kapasi_thesis.pdf
[/url]