AMD: R7xx Speculation

So are they saying that AMD/ATI went from 64 5D shaders --> 160 5D shaders?!

That is a 250% increase in ALUs..

 

I was playing around on my computer and tried to make little art-work of Radeon 4870X2
http://rapidshare.com/files/120694581/R-700.jpg

 

I wonder about TMU's and ROP's - if their is any surprise

 

So how many Z per clock do we all think RV770 will be doing? still 2?

 

Didn't someone from ATI mention in an interview somewhere that the ALU's on the R600 were relatively "cheap" in terms of trannies?

While I certainly scratch my head and frown in disbelief at an 800 SP count, it does seem to make the most sense for a "Terascale engine" with the current rumored clocks.

Although if we accept 800 SPs...where do you get room to beef up the RBEs either in count or in capability?

Regards,
SB

 

Are you telling me they packed over 3x G200 ALUs in less than half the a G200 die size?

edit: Unless they are now using their programmable units to carry on tasks that were processed by fixed function units in the previous generation(s). Still sounds quite unbelievable.

 

No, those are different ALUs than G(T)200's.

 

No, the RV770 has no G200 ALUs, but RV770 ALUs. U cant compare them.

 

800SPs and >1TFLOPs are nice numbers, which I hope AMD is able to market.
But efficiency seems not to be strength of RV770, if you connect this numbers with performance rumors.

So lets see how they are organized...

 

I'm not directly comparing them, just assuming that the G200 and R700 are not THAT much different from their predecessors.

 

Assuming the 800 figure to be true, and considering the direction of increasing flexible and programable resources as opposed to fixed function units, and then thinking of what might be most suitable for a Fusion co-processor, just how much ROP/TMU functionality could be handled into the shader arrays?

If a design took that path, would it possible be more scalable as smaller nodes became available, and quicker to design and debug?

 

G80/G92: 16/8 processors/clusters
RV670: 4 processors/clusters
G200: 30/10 processors/clusters
RV770: probably 5 processors/clusters

This seems to be the reason, why 1 TFLOP HD4850 is only on 9800 GTX (~430GFLOPs) level and ~ 2TFLOPs R700XT is just ~20% percent over 933GFLOPs GTX 280 - based on Vantage rumors from CJ.

 

It will be far too beefy for that. If anything I wonder if one of these might almost be in next gen Xbox/PS3.


There aren't a lot of die shrinks left so..

 

It depends on what efficiency you are looking at. In games, the efficiency "appears" bad because many (if not all) titles end up being "texture bound" (for lack of a better term) on R(v)6xx and possibly Rv770 also.

If you go into the scientific/professional arena, efficiency starts to take off as they push more math to the card and texturing ability becomes a non-factor.

Regards,
SB

 

Lets hope so, such a 110W HD4850, with its 1 TFLOPs would be a nice co-numbercruncher for PhysX, which hopefully AMD will adopt, or other tasks, while GTX 200 is doing graphics.

 

Well, so we will have a Teraflops on a 199$ vga!
This is could hurt Nvidia Tesla/CUDA: they will sold the same processing power for thrice price and over twice power consumption...
AMD makes happy small lab and University: high processing power for everyone!

 

Well I guess thats why ATI kept the SP count so secret until recently... I mean the GT200 had its SP count leaked at Computex while AIB partners didn't post ATI SP count at all...

 

Indeed. I´m going to enjoy this a lot, since that´s exactly what I had been discussing with Ailuros countless times (since about the time when the R420->R480 transition took place). Seems that this particular topic can finally be put to a good rest.

 

I don't believe ATi will adopt CUDA and PhysX, I don't even believe what Fudzilla says about nVidia offering those technologies to ATi. A few weeks ago I spoke to an nVidia PR guy and he told me CUDA and PhysX are two things their cards will have and the competition won't, giving nVidia a clear advantage. Now there's the question - would it be better for them if they risked keeping it for themselves, or will they play safe and let the competition support it as well so that no developers will be afraid of using it?

 

Then again, you still don't know if a G8x/G9x SP is as powerful as a G200 SP or not, so that raw count is meaningless...