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

well %20 faster than 580 isnt bad for immature drivers.. it is still %45-50 faster than 6970..

Spoiler

http://www.computerbase.de/artikel/...60-ti-448-core/5/#abschnitt_leistung_mit_aaaf

 

Ok, who has been banging the almost as fast as 6990 drum? Should have said 5970!

Seems wrong, I don't believe a 560(448) is faster than a 6970. (unless that one is an OCed model).



^ this one is stock.

Must be the choice of games...

 

Haha... nice one.

 

Then again, as this very chart shows, the 6990 is nowhere near twice as fast as the 6970.

 

And TPU has the widest selection of games, so their numbers should be a bit more accurate

 

Well, the whole review (http://www.techpowerup.com/reviews/ASUS/GTX_560_Ti_448_Cores_DirectCU_II/1.html) doesn't list indiviudual 6990 scores, so it's hard to tell what does the 6990 score consist of. Maybe older drivers, maybe there are some games that didn't scale well. But anyways, it's obvious it doesn't scale 100%, but at least 60%. And the 7970 appears to be barely 40% over a 6970.

 

It does - they always just list in the "separate game benches" the closest few cards, in summary they have every single card in their current test-lineup (aka current & previous gen) benched with the set specifed in the test setup page, which in 448 reviews is 11.8 for Radeons and 270.61 for GeForces, except for 448 which has 285.88

 

Alright then. I'mma going to assume that driver optimisations will allow reaching said "prophecied" levels of perf in some time.

 

I guess GTX 580 must have been equally underwhelming for you then consider it's only about 15-20% faster than 6970 and only about 20-25% faster than a 5870 on average except in excessive tesselation benchs which is offset by those cases where 6970 is faster. As well being less than 20% faster than GTX 480.

At least Tahiti is presumably going to have a far larger bump in speed over it's direct predecessor (Cayman) than either GTX 580 (compared to 480) or 6970 (compared to 5870).

If you think a 40-50% bump over Cayman is underwhelming. Then both GTX 580 and 6970 must have been huge failures in your eyes.

Regards,
SB

 

580 to 480 and 580 to 6970 were jumps on the same process generation. Tahiti is using a new process. Thus 580+20% is really underwhelming, yes. If I've had 480 I wouldn't switch to 6970 or 580. If I have 580 now and rumoured +20% are true then I wouldn't switch to Tahiti either. Switch from Cayman is more interesting but then again I'd wait for GK104 performance figures first.

 

Well we do not know the system, nor the settings the cards were running. 3dm does mean nothing especially when comparing highend videocards in P-mode...

 

What kind of expectations you have? Do you really think the manufacturing challenges you face when switching to more and more smaller structures are comparable? I would doubt that. The 28nm TSMC process (maybe apart the 22nm Intel process) is top of current scientific/manufacturing knowledge. Do you really think you can have always 100% more perf, perf/W, perf/mm2 from each full-node progress? That is just ignorant considering the efforts and investments it takes to move to future process technologies. And someday there will be a limit for this kind of manufacturing semiconductor chips.

 

"Up to 768K" ?

 

43% more transistors than GF110
65% than Cayman

 

7950 will have less mem channels then.

 

During Demers' keynote at AFDS it was evident that the L2 is partitioned into 64KB slices -- presumably for each GDDR5 controller. That would also mean 384KB for the whole chip.

Also curious they kept the GDS, too.

 

32 ROPs are surprisingly low. For a 50% bw increase, if they can get ~40% more performance, then I'd say a job very well done.

 

Other slides indicated 64-128K per channel.

You would need 512K just to cover the L1 from all CUs. (16K x 32 CUs).