AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

I'm not really sure if 2560 is really a RAM problem since most of those games are old, but yeah I tend to agree overall with the RAM thing.

GTX 285 was already a compute focused architecture, was it not? When they went to 1D ALU's was when I considered it. 285 to 480 was compute>>compute imo. I am sure there were plenty of changes and optimizations along the way but so where there from evergreen to caymen, or from any GPU to next. All these GPU's had HUGE dies for Nvidia and discussion at the time if Nvidia had focused too much on compute creating a lot of bloat for gaming performance.

The comparison wasn't unfair at all, as his assertion of a compute penalty being paid was proven unlikely by comparing a compute to a non compute arch, where the scaling was fine. He compared 7970 to a different arch, but the variable clearly wasnt necessarily compute.

As for the last part, huh? if anything the 285>480 transition would be more apt, since it was a refresh>new arch, like caymen>SI.

If that is the case it throws the whole "poor scaling" thing even more into question. I was operating on the incorrect assumption that Tahiti=2x transistor of caymen. I was wrong and it's actually 1.63X, which makes trinibwoy's case only look worse. Basically removes it nearly entirely. The scaling according to ever useful tech power up charts is tahiti=1.39X caymen (2GB) performance at 2560. You could probably chalk the other 24 percent up to use of lots of old, non-tesselated games benchmarks and unoptimized tahiti drivers, or close enough. Knowing that in heavy tess games like Crysis 2 Tahiti is 70% faster.

I like big silicon too, usually when AMD goes with small dies and performs well, I tend to think "gosh, think what they could have done with a big die!". It seems like missed opportunities. Overall it has worked okay for them though pretty consistently, including seemingly yet again with tahiti. It clearly can help with dual GPU card performance leadership too, for what that's worth, though I think single GPU performance crown is more valuable. Also, it seems to help them get product out faster, which is pretty big. Every day we go without an Nvidia answer to SI is a big win itself. Whether that's because of the small dies or completely unrelated though I dont know.

 

Fudzilla

Hmmm, 7850 would be my price range. A little high in the range really. It better provide a nice punch and improvement on current 249 cards.

 

According to Muropaketti, Shamino (nor anyone else outside AMD perhaps) has no idea how high the GPU would go - the 1650 was highest the current tools allowed to try

 

Lol, sure if cherry picking is your thing. No need to fabricate numbers, they're right there plain to see.

 

Fudzilla is refering to these fake numbers AFAIK: LenzFire

896 SPs didn't make sense, because the CUs are most probably a factor of 4 (8/12/16/...) because of SI's design.

12 CUs/768 SPs @ 1000 MHz are quite reasonable if it should be around 6850 (960SPs @ ~800 MHz). But interesting that the narrower memory interface doesn't have a bigger impact (even if clocked higher)

I guess this means 20 CUs max for the 7870 (which would reach 6970 levels @1GHz)

 

it doesnt have to be factor of 4.. i think Demers cleared that at Rage3D review.. additionaly 12CUs vs 8CUs wouldnt make any sense %50 difference between XT and PRO

 

Maybe it doesn't have to but IMHO it doesn't make sense to design a chip outside this structure. Maybe the 7750 will have only 10 CUs (2 deactivated) or it will just be clocked lower.

 

Dave himself as well, in this very thread.

http://forum.beyond3d.com/showpost.php?p=1608701&postcount=2080

 

Has it been confirmed anywhere wether Tahiti uses the rumored HPL process instead of HP process?

 

Bulldozer vs. Sandy Bridge vs. Nehalem using the 7970
I'm not surprised but wonder if all were oc to 3.60GHz would have made that much of a difference...

 

Unlikely, with the clocks it's able to reach.

Does anyone know the default voltage for 7970?

In the massive overclock, it's set to 1.17V.

The same screenshot shows 0.85V as an option. That corresponds to the standard voltage of HP. Of course, that doesn't mean it's the default voltage of 7970.

HPL has 1.0V as standard voltage: http://www.tsmc.com/download/brochures/2011_28 Nanometer Process Technology.pdf

 

AFAIK 0,85/1,17 volts are (one of?) the default configuration(s) for HD 7970 - at least that's what the tools say.
HT4U did some multimeter action on their sample: http://ht4u.net/reviews/2011/amd_radeon_hd_7900_southern_island_test/index17.php

 

Now *that's* the card AMD should have released. Stick a new age vapour chamber peltier thing on the card, something like their closed loop water cooling thing with Bulldozer. Ramp up the core and mem and crush the 6990 with it. Sell it for £600. I'd rather pay £600 for something with that spec, than £450 for what we actually got.

 

You do realize this would probably draw way over 600W, right?

 

http://www.chiphell.com/thread-336984-1-1.html

 

1335MHz factory overclocked card ?

 

wow
36CUs.

 

That card most likely would've easily matched and beat the HD6990/GTX590. I want.

 

There are no 36 CU designs.

 

over 44% factory OC'd, indeed

Also, does Tahiti secretly have 36 CU's after all? 2304 shaders on the "cancelled products" would suggest such