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

[raghunathan]

HD 7950 is a much better card than most people think it is. If you do some research you will know clock for clock HD 7950 is around 5% slower than HD 7970. Buying a factory overclocked HD 7950 at 900 Mhz and pushing it to 1050 Mhz in AMD CCC puts the card exactly in between ref HD 7970 (925 Mhz) and HD 7970 Ghz (1050 Mhz) performance wise.

The new batches of HD 7950 could have had a better branding strategy. Even a HD 7960 might have been sensible. AMD's product marketing is really stupid in keeping the same naming.

The boost voltage on the HD 7950 boost edition is very high at 1.25v. That is probably the reason that the ref HD 7950 cards which are flashed with the HD 7950 Boost BIOS aren't hitting 925 Mhz speeds consistently. They are mostly hitting their TDP restrictions easily and thus are only running at 850 Mhz. I hope AMD partners come out with HD 7950 Boost cards that use chips that are binned appropriately and hit the maximum boost speeds consistently.

As always manual overclocking is always the best option. Using AMD CCC just max out those power sliders to +20% and you should be able to hit 1100 - 1150 Mhz with 1.25v.

 

[neliz]

ALL 7950 Owners!! The article clearly shows that is not the case.

In other exiting news, American car companies say that fueling your car with the wrong fuel isn't covered by warranty and feeding your baby rat poison also is NOT advised.

A new report from Kentucky shows that leaving the gas on your stove on for 5 minutes before lighting it might not cause the desired effect.

God I'm happy I work for a company that supports overclocking and we actually release software and BIOSes that let you do so.

 

[lanek]

There shouldn't be any variability from card to card, because PowerTune is deterministic.

The benchmarks themselves could be different, though.

There's different Asic quality and for have seen some really bad overclockers 7950 reported on different forums ( a minority, and in all case without voltage increase ) The PT tune could well cut differently the turbo clock speed ( on some card the max clock speed vary less of the other )... I can imagine cards manufacturers or AMD choose the chips who will be used for thoses 7950"boost", when ofc this dont say the cards used for be flashed by consumers/reviewers, meet the min requirement.

Its what report Anandtech, basically their the Boost seems fluctuate widly, something who was not happend with 7970.. As i believe they have too just flash the bios to a standard 7950, this could explain it. Need been tested with +10 +20% PT limit. Or who know the driver could be maybe the reason behind this fluctuation ?

http://www.anandtech.com/show/6152/amd-announces-new-radeon-hd-7950-with-boost/3

 

[Dave Baumann]

Alexko is correct. If a card supports boost unde the current PowerTune rules then all cards will boost to the same level.

 

[Kaotik]

Alexko is correct. If a card supports boost unde the current PowerTune rules then all cards will boost to the same level.

I think the point was wether the cards are supposed to do that "zigzag" to boost clock and (under!) default clock, or stay at boostclocks like 7970GE's do

 

[Gipsel]

That depends on the position of the Powertune slider. At each specific setting, all HD7950s behave the same. That's actually true for all SI cards (when run with the same BIOS and driver).

And only the current monitoring tools are showing a "zig zag pattern". In reality, the granularity of the clock speed changes is about 4 MHz, iirc.

 

[lanek]

Alexko is correct. If a card supports boost unde the current PowerTune rules then all cards will boost to the same level.

The question was just because the test are now made on flashed cards and not "choosen "chips and we see some variation on results from the few " test/review/preview" ..

 

[Gipsel]

So i was believe thoses can hit more frequently the power limit, and so get more variation on clock speed.

As the power is not measured within the power regulator circuitry but estimated from the workload, all cards hit the exact same limits at the exact same spots, irrespective of any possible variations. This can only be influenced by the BIOS and driver version (which can change power states and also the powertune algorithm itself).

 

[Dave Baumann]

No. Flashing a "PRO" ASIC board with a PRO2 BIOS makes the ASIC run the same as a PRO2.

The Boost to non-Boost pattern is simply because there was less headroom on the board for PRO than XT/XT2. Boost is about maximimising the board design specifications as much as possible so you will see a varying degree of behaviours for different SKU's dependant on their thermal/electrical design parameters.

 

[lanek]

Okaaay... Now i get a better figure of how act the " Boost pattern".

( sorry i see i had edit my post inbetween ( no auto refresh ), )

 

[raghunathan]

No. Flashing a "PRO" ASIC board with a PRO2 BIOS makes the ASIC run the same as a PRO2.

The Boost to non-Boost pattern is simply because there was less headroom on the board for PRO than XT/XT2. Boost is about maximimising the board design specifications as much as possible so you will see a varying degree of behaviours for different SKU's dependant on their thermal/electrical design parameters.

Dave
The issue I see is why is the boost voltage so high at 1.25v. Thats just too much. The reviews at various sites are showing that with your HD 7950 Boost BIOS the performance is only similar to HD 7950 (850 Mhz) and not HD 7950 (925 Mhz) as we would expect. Definitely the clocks are being throttled. Users of HD 7950 cards were able to hit upto 1 Ghz without increasing voltage.
Also sorry to say but AMD's product marketing is very poor. AMD could have had new partner boards with properly binned chips that boost to 925 Mhz and provide the performance of a HD 7950 (925 Mhz) ready to be reviewed in time for the GTX 660 Ti launch. The entire AMD marketing team has failed. The HD 7950 is an excellent product and imo the best price perf product in the market. clock for clock HD 7950 is 3 -5% slower than HD 7970 and the best value for the consumer's money.
Excellent AMD partner cards like Sapphire HD 7950 950 Mhz edition with a HD 7970 PCB are selling for USD 350. AMD could have sent one of these cards to tech reviewers to include in their reviews .
imo manually overclocking a HD 7950 either in AMD CCC or using third party software like Sapphire trixx or MSI Afterburner is still the best option.

 

[jimbo75]

I tend to agree with most of that. I'm sure just about every reviewer got higher than 925 MHz at stock volts. I think this 7950 boost has been ill conceived and poorly executed, and the general feeling from the reviewers is that the small increase in performance isn't nearly worth the increase in power draw.

 

[Dave Baumann]

With regards to voltages - if you are taking that from the Anand review then bear in mind that he is taking a sample of one. As has been explained, this is a little meaninless as different boards will operate at different voltages.

 

[raghunathan]

Dave Baumann
The lack of sufficient number of reviews of the AMD partner boards of Radeon HD 7970 Ghz edition is quite baffling. Two weeks after they have been available cards like Sapphire HD 7970 Ghz Vapor-X , Gigabyte HD 7970 Ghz, XFX HD 7970 Ghz have almost no reviews. The only HD 7970 Ghz card that has been recently reviewed is Sapphire HD 7970 Vapor-X by tweaktown

http://www.tweaktown.com/reviews/48...3gb_overclocked_video_card_review/index1.html

The Sapphire HD 7970 TOXIC 6 GB was reviewed by many sites. But that card was priced at USD 700. These other cards are priced reasonably as USD 450 - 470 at newegg. AMD and its partners need to do a much better job with marketing and promotion. Nvidia really knows how to market its products. AMD is suffering a lot from self inflicted wounds. AMD has the best products but they mess it up with their bad marketing. The lack of HD 7950 Boost edition partner boards for the GTX 660 Ti launch being a good example. Another example is the launching of the HD 7970 Ghz launch. The partner products arrived more than a month after the initial reviews. For initial reviews the AMD reference cooler based design was sent out to reviewers and the card was criticised for its poor acoustics, noise under load and also the increased temps. What was worse is the partner cards were all going to be using their own cooling designs. Inspite of those noise and temps being inconsequential to partner versions the initial damage was done. AMD underestimates the effect of perception. Truth and perception are very different and its very well illustrated in the case of AMD products. AMD products are much better than the common perception and a lot of the blame for that lies on AMD itself.

This is not meant to criticise but to improve AMD as a company. I completely believe we need a strong and aggressive AMD to keep Intel and Nvidia honest and on their toes. Its in the best interests of the industry , the consumers and everybody that AMD do much better than what its doing as of today.

 

[gongo]

Hi Dave, you once said AMD made a marketing disicion to high balled GCN prices to build a brand...imo it has failed and backfired, i gone through several reviews of GCN vs Kepler, yes Nvidia has you beat on perf/w but even this is closer than previous gens, and when GTX680 beat 7970 'convincingly', those were Nvida TWIMTP games...even this $299, 660 Ti, imo objectively is not as impressive as what i can get of 7950 3GB...but why are reviewers calling out on AMD and not Nvidia?

If AMD had started GCN srp more competitively and focus more on helping game developers update their engines to GCN support, then you are building the brand... slowly. But i will forgive AMD this round if they start releasing unlockable 7950....

 

[jimbo75]

The problem AMD has had is Tahiti is too heavily saddled with compute to be an amazing gaming chip, and Pitcairn - though clearly very capable and competitive - is too small to be really stunning to people who were looking for a big performance increase.

Had AMD simply gone with a ~350mm2 pure gaming chip it would still be commanding $500 and Nvidia would have been forced to ditch their 680 plan.

If AMD would just take Nvidia head-on instead of trying to be faster to market with smaller chips, they'd win a lot of mindshare.

 

[lanek]

Hi Dave, you once said AMD made a marketing disicion to high balled GCN prices to build a brand...imo it has failed and backfired, i gone through several reviews of GCN vs Kepler, yes Nvidia has you beat on perf/w but even this is closer than previous gens, and when GTX680 beat 7970 'convincingly', those were Nvida TWIMTP games...even this $299, 660 Ti, imo objectively is not as impressive as what i can get of 7950 3GB...but why are reviewers calling out on AMD and not Nvidia?

If AMD had started GCN srp more competitively and focus more on helping game developers update their engines to GCN support, then you are building the brand... slowly. But i will forgive AMD this round if they start releasing unlockable 7950....

Reviews of GCN was extremely good, Kepler have been a good surprise too, and the reviewers tell it. Well some are maybe less objective when it come to Nvidia ( i will take the example of the price complain about 7900series, when finally 680 are way priced higher ) .

GCN is better in some aspect, kepler is better on some other. AMD is faster in some games, Nvidia is faster on some other. He, the 7970 is out since 8months, the 7870 since 6 months. You dont expect Nvidia to come 6months later with a middle range gpu who is priced higher of the AMD one's and is slower, they dont want destroy their place lol. they could even have release the 660TI with 920mhz turbo and let the memory controller as it is was.


The 660TI do his job, it is so close of a 670 spec, it could not have been different anyway, specially the OC models who run at 1111+mhz with 170W (gigabyte, Evga, MSI )
( its more a 670light of a 660 anyway, and the 660 will be a lot different ). Thoses cards, performance wise are more or less equal to the 670 for 100$ less.

Nvidia set a real low MSRP on the reference 660TI, but if i take the example of guru3D, even if he have listed the 660TI on the benchmark, he have absolutely not test any reference cards today: the 3 cards reviewed are custom OC 660TI ( MSI, Gigabyte and Palit ). ( A bit strange to see him have the benchmark fps of the 660TI "reference" when have not write a word about the reference card .)



About "GCN support", i dont know, there's many games today out where AMD have working closely with the developpers ( including ofc some as BF3 where they are a bit slower, working with developpers dont mean forcibly favor himself ), and if you look the list of future games where AMD HD3D, Eyefinity is even on the settings panel of the games, or where they are involved with the developpers, i dont think its so bad. ( Even the really new technology demonstrate on "Leo Demo", ( Forward + ) is used on Dirt Showdown and if im right, some future games too.

 

[no-X]

The problem AMD has had is Tahiti is too heavily saddled with compute…

I don't think so. Look at the dieshot, remove 1/3 of the memory interface/controller and the rest will be quite close to 300 mm². I think the main reason is the 384bit memory interface/controller, ton compute abilities. Anyway, 365 mm² isn't a big GPU. It's as big as GF104/114 and smaller than Cayman.

 

[raghunathan]

I don't think so. Look at the dieshot, remove 1/3 of the memory interface/controller and the rest will be quite close to 300 mm². I think the main reason is the 384bit memory interface/controller, ton compute abilities. Anyway, 365 mm² isn't a big GPU. It's as big as GF104/114 and smaller than Cayman.

The problem with Tahiti is its unbalanced. The lack of front end resources has been its biggest problem. Pitcairn and Tahiti have the same front end resources. 2 ACE, 2 rasterizer engines, 2 tesselators. But Tahiti has 60% more shaders. The perf scaling from Pitcairn to Tahiti is far less than optimal. In quite a few the scaling is as low as 25% at the same clocks. Also having a 384 bit memory controller with 32 ROPs is not a good choice. Ideally it should be 48. 8 ROPs per 64 bit memory controller is always a good balanced design.

The most important correction which AMD needs to make with HD 8970 is add more front end resources. So even if its just 2304 SPs but 3 ACE, 3 rasterizer engines,3 tesselators with 48 ROPs that would improve performance quite a bit over the HD 7970 Ghz. Each ACE, rasterizer, tesselator would handle 768 cores and would be well fed. Cayman had dual rasterizer with 1536 cores and thus the same ratio.

 

[no-X]

ROPs are quite big units. Would additional 16 ROPs boost perf/mm²? I don't think so. Tahiti doesn't suffer in high resolutions / high MSAA scenarios. In fact these are its strongest point.