The LAST R600 Rumours & Speculation Thread

[Per B]

they mention that RV630 has 64 shaders, are these the romoured vec-4 shaders or are these ones scalar?

Wouldn't it be strange if the R6xx architecure was NOT largely based on Xenon? So in that case, it should have the Vec4. Not that I fully understand the benefit of it though... anyone care to point me to a quick summary of the difference of one shader in G80 and one in Xenon, with their + and -.

Thanks

 

[Per B]

On the other hand, 128 vec4 shaders would totally kick *rse, to such an extent that I find it difficult to believe they could have pulled that off. (Or rather, difficult to believe that only one ou of ATI and Nvidia could have).

Doesn't Xenon have 48 vec4 shaders? So why couldn't the R600 (with 700 million trannies) have 128?!?

Or am I missing something about the Xenon shaders?

 

[INKster]

Doesn't Xenon have 48 vec4 shaders? So why couldn't the R600 (with 700 million trannies) have 128?!?

Or am I missing something about the Xenon shaders?

They don't need to be DX10 (or OpenGL, for that matter) compliant, therefore they are much less complex than R600's ?

 

[3dilettante]

Doesn't Xenon have 48 vec4 shaders? So why couldn't the R600 (with 700 million trannies) have 128?!?

Or am I missing something about the Xenon shaders?

Xenon is the Xbox 360 CPU.

Xenos is about 330 million transistors, but I don't think that's counting the daughter die, which has hardware that would be integrated on R600.

It might be rather cramped fitting all that in with more than double the shader count.

 

[Sound_Card]

Doesn't Xenon have 48 vec4 shaders? So why couldn't the R600 (with 700 million trannies) have 128?!?

Or am I missing something about the Xenon shaders?

1) It's Xenos..hehe...

2) I believe Xenos has vec3, not vec4.

3) I'm not buying into the R600 vec rumors at all.

 

[stevem]

Rumours? What rumours...?

 

[Sound_Card]

Xenon is the Xbox 360 CPU.

Xenos is about 330 million transistors, but I don't think that's counting the daughter die, which has hardware that would be integrated on R600.

It might be rather cramped fitting all that in with more than double the shader count.

Actully Xenos has 250 million transistors(rounding). That's without the doughter die which is said to be roughly around 150 million transistors. Were looking at about 400 million or so transistors if it was all combined into one package.

 

[jamis]

Xenos is about 330 million transistors, but I don't think that's counting the daughter die, which has hardware that would be integrated on R600.

Yes it is. The daughter is about 100M of which 80% is the ram. The actual GPU with ROP's and all is about 250M.
(The "mother" die is around 232M I think)

 

[nelg]

Would it be such a scheduling nightmare if all the shaders were not homogeneous?

What about 64vec4 + 1 Sp and 64 scalar.

 

[Sound_Card]

Yes it is. The daughter is about 100M of which 80% is the ram. The actual GPU with ROP's and all is about 250M.
(The "mother" die is around 232M I think)

Reading though B3D's Xenos artical(Which I have read I don't know how many times now...) states their is no said amount of transistors given to the doughter die. However, B3D speculates in a educated manner that 150 million seems about right. Their was a leak of a 150 million transistors before E305 for Xenos, but that proved later to be wrong as Xenos was 250m. But many beleive that leak of 150 million was actully the doughter die.

It is said that 80 million transistors is for the ram, which would put roughly about over 50% of the die.

 

[Per B]

Sorry bout the Xenon/Xenos mix-up...

Anyway, on http://www.beyond3d.com/articles/xenos/index.php?p=07 it says that "Each one of the 48 ALU's can co-issue a vector (Vec4) and a scalar instruction simultaneously, essentially allowing a "5D" operation per cycle.". What does that mean when comparing on Xenos ALU (and possibly R6xx) to one from G80?

Would a R600 with 128 Xenos-like shaders "trash" the G80?

 

[Dave Baumann]

The daughter die is 105M transistors, 80M of which is EDRAM.

 

[Sound_Card]

The daughter die is 105M transistors, 80M of which is EDRAM.

shucks, I don't see that in your Xenos artical. Must not have been updated in a while..

I stand corrected then. That's 80%.

 

[Anarchist4000]

7.5 - May

The new MSAA adaptive anti-aliasing is scheduled to be put into the May Catalyst release. This feature will bring their AA options on-par or beyond what NVIDIA currently offers on the G80 graphics cards. This feature timeline is a target, and not yet a commitment from AMD at this time.

http://www.pcper.com/article.php?aid=367

Found this bit a little interesting.

Would a R600 with 128 Xenos-like shaders "trash" the G80?

With 128 of them it would be an absolute slaughter. You likely won't be seeing nearly that many however.

 

[ERK]

shucks, I don't see that in your Xenos artical. Must not have been updated in a while..

I stand corrected then. That's 80%.

[/Tommy Chong voice] Hey, man! Dave's not here, man! [/]

 

[pjbliverpool]

With 128 of them it would be an absolute slaughter. You likely won't be seeing nearly that many however.

It would depend on clock speed aswell. If those Xenos like shaders were also running at Xenos like speeds then on paper even 128 of them wouldn't provide much more theoretical shader power than G80 (including the MUL). And it would be less efficient power at that. Of course we expect R600 to be clocked much hgher than 500Mhz.

 

[Russell]

If these shaders are scalar, perhaps this is where that 2GHz rumor from awhile back comes in. Sound_Card's clock domains/scalar shader theory certainly has some merit.

 

[Sound_Card]

If these shaders are scalar, perhaps this is where that 2GHz rumor from awhile back comes in. Sound_Card's clock domains/scalar shader theory certainly has some merit.

Theroy? hehe, I would say it's almost certain.. JMHO. At least the scalar part that is. Clcok domains could be anyones guess, but, personally I think having a clock domain is a must when going scalar.

 

[Farhan]

If these shaders are scalar, perhaps this is where that 2GHz rumor from awhile back comes in. Sound_Card's clock domains/scalar shader theory certainly has some merit.

If they are scalar, it still doesn't necessarily mean that they have multi ghz clocks as a design target... It could be vec4 and still be clocked very high if that is what they wanted in the design.

 

[KimB]

personally I think having a clock domain is a must when going scalar.

I don't see why.