G80 rumours

You do realize texturing takes a significant part of bandwidth too, right?

Uttar

 

Xenos has "just" 8 ROPs (capable of 32c/64z) and 16 (bilinear) TMUs...

 

I certainly haven't seen any evidence that we need more than 16 ROPs at the high end. Now, more capable ROPs. . .there you'd have my interest.

What kind of situation would more than 16 make sense for?

 

Oh, yeah, definitely. I'd love to see stuff like that. For example, ROPs that are capable of more novel FSAA methods, or ROPs that are more efficient at memory bandwidth usage with FSAA enabled. An addition of programmable blending could be very cool.

 

Well, I'll take the opportunity to beat my little drum one more time. . .I want 4xAA single-loop ROPs. Time to get past 2x as the minimum, as it is too limiting on the maximum side. . .

 

Great for the "bigger is better" bunch

 

And I want (more) smart ROPs that don't take 8x more time when they end up writing the exact same thing, bit-for-bit, due to Z/Color compression under 8x MSAA *shrugs*

Uttar

 

If it does 8x more of one thing, then its likely to do 8x less somwhere else.

 

Unless it does 2 or more things 8x at once.

 

"more than "2" times G71 Pixel Shader throughout"
4:1 is for sure
700MHz(just sample clock speed......)
Die Size of G80 is smaller than that of R580(not 80nm Low-K R580+)

 

Could someone spin off the R580/R580+ rumors into their own thread?

 

So let it be done: http://www.beyond3d.com/forum/showthread.php?t=32103&page=5

 

Doesn't make any sense...

G80 is supposedly 500M+ transistors on TSMC's Low-K 90nm, right ?
If R580 is made on the same process and has 384M, how can a much larger design occupy a smaller space ?

I can't figure it out.

 

NV40 and R420 were about the same size but NV40 had ~ 38% more transistors. 500M is only 30% more than 384 so it might just be true due to different transistor counting methodologies between the IHV's.

 

The point I was trying to make is that you don't reach that kind of theoretical GFLOP throughtput on a hypothetical "8-quad/G7x-alike" design.

 

I already signified more than 48 G8X PS ALUs. not G7X PS Configuration-alike.

 

Depends how you count ALUs in the end. G7x depending on different theoretical math scenarios has either 24 or 48 ALUs.

 

Given that I would take with a pinch of salt the comments on G80(1) from trumphsiao saying it is 700Mhz 4:1, 90nm and smaller than r580 .....

I doubt nvidia will go 4:1, and I still am not sure if it is 80nm or 90nnm even though trumphsiao and Uttar think it is 90nm. True, it is more likely the mid range parts will be 80nm first but if 80nm is fine then I think that nvidia will grab it as fast they can. We shall see.

 

There are two versions of G80 at least. one is bigger and the other is smaller..

 

2 chips with different pipeline configs? Then both won't be called g80's though, unless both are the same chip and one chip has the pipelines disabled, nor would they be both called g80's if the process of both chips are different.