AMD RV770 refresh -> RV790

If the rumour of "troubles with 40nm" is true, will we even see it this year? R600 all over again? The Most Definitely Maybe Final RV870 Rumours and Speculation thread. Shall we start it now?

GT212 is shaping up to be G92 all over again, arguably. AMD won't be able to compete with it for 9 months and then for the second 9 months NVidia will look silly as it'll be utterly outclassed by RV870.

So, the real question now is, how late will RV870 be? Maybe we should make that the thread title (yep, I know, there'll be some thinking, will there ever be an RV870).

Jawed

 

So, this RV790 is some kind of what R360 (9800XT) was.
On the other side: RV790@40 nm is just the next R650.

 

The first samples were run at 750 MHz core clock frequency and 900-975 MHz memory clock frequency.
So it looks like that AMD has got added new SIMD units or uses an separate ALU domain.

 

Or maybe the samples are running on lower clocks.

 

Since everyone knows about these RV790 and RV740 samples being out and about, another source has said that AMD has two 40nm samples. I assumed originally that RV790 was 55nm, but upon hearing that was dreaming of a 1200SP 40nm RV790 which this info disproves.

So what are the chances of the other 40nm chip not being lowend/mainstream but RV870? If so, Q3 shouldn't be that hard and no need to worry about '10.

PS- with this news of another RV740 respin, March seems to be just out of reach.

 

Who says, "RV790" is a 40 nm chip or a chip with more than 800 SPs? HWI? And who else?

 

http://theovalich.wordpress.com/2009/01/26/amd-preparing-worlds-most-advanced-chips/

 

Coolowski's post went AWOL (spam-trapped I guess) - it linked:

http://theovalich.wordpress.com/2009/01/26/amd-preparing-worlds-most-advanced-chips/

Jawed

 

Have to wonder about that article. It seems to source the hardware-infos site plus some info he has got somewhere about RV740 and RV790 been released soon as the 47XX and 49XX. The author claims as the clock is 750Mhz he is willing to wager on the Rv790 being 1200SPs. Anyone want to take this guys money?

The original leak was that for the RV790 99% of the viable chips on the wafer would run at 4870 clock speeds without any additional voltage. ie 750Mhz. Same sources said it was 55GT. Should be obvious that there will be chips on the wafer that will run in excess of 750Mhz for the high end users.

The number of shaders is still not 100%, but path of least resistance is just to use the RV770 design. 50% more shaders on 55GT sounds improbable, the die size and power usage would surely head too far upwards also contradicting the original leak claiming the same voltage and current as the RV770.

The 55GT process is supposedly more expensive than the 55GP they are currently using for the RV770.

 

Same no. of simd engines on 40 nm would make them pad bound, wouldn't it? after all, it was said that rv 770 on 55nm was pad limited.

 

this guy is definitely wrong about ati switching over to the foundry company. They are so far behind in process shrinks, 45 nm is only beginning to ramp up.

 

Mainly ati-forum.de. Noone mentioned chip configurations for 40nm, but they will have MUCH more than 800 SPs to avoid being pad limited.
I think a 55GT version, slightly redesigned and more expensive for additional 10-20% seems pretty unlikely, especially in AMD's case, and especially when we consider the sample clocks (which comes from the "40 nm people")

 

From Theo
"When ATI/AMD GPG moves to The Foundry Company, Nvidia will remain sole proprietor of TSMC’s most advanced lines and is set to continue GPU lead over CPU manufacturing." (my bold)

This guy honestly thinks, the whole world is just GPUs?

 

If ATI truly increased the cluster amount, then it sounds like 1280SPs or better 256 ALUs / 16 clusters. 60% more ALUs sounds a tad too much but you never know. Other possibilities would include 14 or even 12 clusters.

 

Perfect scaling from 58.5 nm to 40 nm suggests, near doubling, so 60% may not be so off the mark. And as KonKort says, may be they increased the clocks, but yes, 10 simd's will definitely be pad limited and hence a waste of die.

 

It looks like AMD plans a dual gpu of RV790, too.

HD 4995 X2
HD 4970
HD 4950

Source: ATI-Forum

 

10 SIMDs may be pad limited, but it's also very likely they wouldn't double the SIMD count, if the SIMDs are much like the current ones.

The potential would exist to double SIMD count in terms of physical space.
In terms of power draw, it is trouble since the GPU's peak power with high utilization is already pretty high with 10 SIMDs.
Transistor per mm2 is about double.
Watts per mm2 is not half.

There's going to be die space to spend on something else, so long as pad limitations persist.

 

so this means that they will keep the additional pcie connection between the die?

i've read on this forum that on the r770 it don't worked well, exactly what was broken?
and what can be the gain of an ipotetical corrected r770?

 

As far as I can remember, the small performance increases it gives weren't worth the extra powerdraw keeping it enabled would give, but AMD has the possiblity to enable it via drivers if the want to.

 

i thinked that was broken, instead it is useless?
what an over hyped feature!

and do you think that an enanced version (bandwidth/latency/other) can be usefull, or they will keep or drop depending on the power budget?

and why they don't use old fast simple cheap HT3?