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

[Blazkowicz]

Tahiti is the first > 256bit GPU from AMD, barring the failed R600.
so it ain't cheap, though 500 euros is usual for a brand new high end card. then don't buy it

GPU generations have slowed down that's all. me I'm waiting that Barts, Juniper etc. are phased out so the new generation gets a bit cheaper.
I would like to do the "VGA passthru" hack where a windows VM gets to use a real card.
so my set up would be a 8400GS for linux, a sound card, a radeon 7770 for windows gaming only, all running on a 350W power supply in a 12 year old case.

 

[mczak]

Still no news on Pitcairn - even though it should be here in a week, and the rumors start getting old?
There are some rumors floating around about increased clocks for the 7850 (900Mhz now). I certainly could believe that but only if the the 7870 clocks move up too as there'd be almost no difference otherwise to the latest rumors for 7870 (which was said to have 950Mhz clock). Though I still think the 7870 will have at least 1Ghz clock no matter what anyway .
No new rumors about pricing neither, and nothing new on the ROP front (don't buy 24 ROPs neither not for the 7870 at least). Boring .

 

[swaaye]

Probably it has something to do with this:
Whenever you create a new marketing paradigm, chances are it'll backfire on you as soon as you don't adhere to it anymore.

Hmmmm. Yeah I suppose. X850XT PE was $550, X1800XT $550 and then X1950XTX appeared at the low, low $450. That is a downward slide I guess.

I think I remember reading that ATI lost a lot of the R300-glory-days marketshare due to their weak midrange chips and exorbitant high-end prices during the R400 and R500 years. Then R600 forced them to be "the good guys" at $300. Next came the nasty R700/GT200 price war, probably caused by DAMMIT trying to regain marketshare from generations of overpriced and/or disappointing hardware.

Then, as mczak brought up I think, Cypress was priced against the price-warring GTX 285. Cayman wasn't strong enough to be high-priced (sort of R600-like). But, today, we have AMD well ahead in speed for a change, and NVIDIA running high prices, and so I think the return to high pricing is reasonable from their perspective...

 

[Silent_Buddha]

I think I remember reading that ATI lost a lot of the R300-glory-days marketshare due to their weak midrange chips and exorbitant high-end prices during the R400 and R500 years.

Which seems odd on the face of it as AMD and Nvidia were both leapfrogging and matching each other back in those days. But at the end of the day, both companies had basically similar performance, Nvidia was slightly faster while AMD had slightly better IQ. And price was basically the same in the enthusiast/performance range (only 3 market segments back then). So it's hard to say AMD suffered due to exorbitant high-end prices when Nvidia were doing the same.

On the mid-range front however, Nvidia definitely had a leg up in performance, and there AMD's better IQ couldn't make up for the performance gap. I think this is what really hurt AMD back then. After Radeon 9500 pro, it wasn't until the 4xxx line of cards that AMD again had a really compelling midrange solution. 9600 pro was already sometimes slower than 9500 pro and sometimes faster. And it just went downhill from there in successive generations.

Nvidia on the other hand came out with some fantastic midrange cards in the 6600 and 7600 series which made them quite popular with gamers on a budget. But then stumbled a bit with the 8600 and then nothing until Geforce 450 series (I picture 460 more performance than midrange).

Regards,
SB

 

[CarstenS]

8600 was crap(py slow), agreed. But 9600 GT was pretty decent for it's time IIRC.

 

[UniversalTruth]

...If "well developed, much cheaper process" is your target, then why did anyone leave 65nm? 90nm? 130nm? 1µm?!?...

Maybe because exactly of this- it was cheaper to produce chips on the new node instead of the old one. Now, the situation is a little bit different, people usually expect from a new node more performance for the same money or the same performance for less money. You see how CV brings 30% lower performance than Barts but at the same time it is as expensive as the old one. This gets ugly, no matter if the reason is greedy marketing department or expensive manufacturing process. Transitions to new nodes get more and more expensive with every new one.
Or maybe it is GCN as an architecture which we should blame. Like the colleague said:

It isn't about price, it's price/perf over prevous generations. 9800 GT/GTS 250 rebrands probably offer better price/perf increase over their original 8800 gt/9800 GTX+ counterparts

...All were cheaper and "more mature" than their successors...

Barts has always been a 40 nm product despite the fact that Cayman was backported to the old process. And the reason for this was exactly the price to manufacture it. See how things change?

...there's an obvious reason why we moved away from them -- just like the eventual move away from 28nm...

Yeah, like the never ending race for performance crowns.

...And as for stable? I'm sorry, I don't see any point you can make about instability. Please back this up with fact rather than your imagination...

Stable is just a word to express my opinion. Don't take it literally.


Special thanks to CarstenS for the posted picture.

 

[leoneazzurro]

Maybe because exactly of this- it was cheaper to produce chips on the new node instead of the old one.

It IS quite probably cheaper to produce chips with the same amount of transistors with the 28nm process compared to the 40 nm process. Chips with the same area will be, initially, more expensive on the 28 nm process compared to the 40 nm process.

Now, the situation is a little bit different, people usually expect from a new node more performance for the same money or the same performance for less money. You see how CV brings 30% lower performance than Barts but at the same time it is as expensive as the old one. This gets ugly, no matter if the reason is greedy marketing department or expensive manufacturing process. Transitions to new nodes get more and more expensive with every new one.

CV is the successor of Turks, not Barts. But, I think there is one thing that is mistaken, Reason for changing the process for a chip supplier is to make more money for a given amount of performance, not the customer paying less (which will also happen, in the long term). And I think we can all agree that performance/mm^2 of CV is unrivalled today. Moreover, I believe we cannot compare launch price for a new product directly with the EOL price of a mature product, especially in the case of (quite probable) inventory clearing. And, if the price will be considered high by the market, then there will be of course price cuts at the horizon.

Or maybe it is GCN as an architecture which we should blame. Like the colleague said:

I don't see the point. A new architecture is needed for many reasons (i.e. better scaling with frequencies, with shader count, modularity, GP performance, better programming, etc.). If we want to discuss about the price/performance ratio of the Geforce 8600 at launch, we could do so (and Nvidia sold tons of these parts).

 

[Blazkowicz]

8600 was crap(py slow), agreed. But 9600 GT was pretty decent for it's time IIRC.

yes, it was an excellent card though often only slightly cheaper than a 8800GT.

 

[Arty]

Probably it has something to do with this:

Whenever you create a new marketing paradigm, chances are it'll backfire on you as soon as you don't adhere to it anymore.

Worth pointing out that when X1950 was released, Nvidia had basically cratered the high end segment prices with their tiny G71. IIRC 7900GTX was about $400-450.

AMD has always been pricing their SKUs relative to the competition at the time. If AMD is gouging prices, blame lack of competition on the pricing front.

 

[mczak]

yes, it was an excellent card though often only slightly cheaper than a 8800GT.

I think that was its biggest problem - too close to g92 in just about all respects. Similar price, similar TDP (similar boards, similar cooling solutions, probably fairly similar manufacturing costs all around), similar performance even (and that it was "new" compared to 8800gt certainly meant absolutely nothing...).
And later probably due to price war 9800gt were fairly cheap too so there just wasn't much room for the 9600gt. I agree though in principle it was very nice.

 

[swaaye]

Considering 9600GT could challenge 3870, essentially AMD's best, NV probably had little reason to push prices down.

 

[Silent_Buddha]

8600 was crap(py slow), agreed. But 9600 GT was pretty decent for it's time IIRC.

Yes it was certainly decent. I think I mostly forgot about it because the 8800 GT covered such a wide variety of price points (from 199 to 279 depending on model when the 9600 GT launched at 169 to 189). It definitely was quite a step up from the anemic 8600 GT, however. So in all fairness I should have stuck it in there with the other decent midrange Nvidia products.

Regards,
SB

 

[UniversalTruth]

Radeon HD 7870 GHz Edition 2G GDDR5 OC Version -> $410;
Radeon HD 7870 GHz Edition 2G GDDR5 -> $400;
Radeon HD 7850 2G GDDR5 OC Version -> $300;
Radeon HD 7850 2G GDDR5 -> $280.

"Nice" prices if true

 

[McHuj]

Radeon HD 7870 GHz Edition 2G GDDR5 OC Version -> $410;
Radeon HD 7870 GHz Edition 2G GDDR5 -> $400;
Radeon HD 7850 2G GDDR5 OC Version -> $300;
Radeon HD 7850 2G GDDR5 -> $280.

"Nice" prices if true

I can't imagine that to be true. That would leave a pretty big hole between the 7770 ($160) and 7850 ($280). Granted a 7790 and 7830 could fill that gap, but I don't believe there are any plans to release those anytime soon. I would guess a large volume of GPU's is sold around $200 so it's hard to believe that AMD would give up that price point (maybe until Kepler is released)

 

[IbaneZ]

http://www.pcgameshardware.de/aid,8...deon-Grafikkarten-gesichtet/Grafikkarte/News/

 

[AnarchX]

Radeon HD 7870 GHz Edition 2G GDDR5 OC Version -> $410;
Radeon HD 7870 GHz Edition 2G GDDR5 -> $400;
Radeon HD 7850 2G GDDR5 OC Version -> $300;
Radeon HD 7850 2G GDDR5 -> $280.

"Nice" prices if true

Would match to AMDs Pitcairn between $199 - $399 slide: http://images.ht4u.net/reviews/2012/amd_radeon_hd_7700_test/product_pricing.jpg

Either Kepler is far away, priced in the same way or AMD is not interested in long-term customer relationships.

 

[mczak]

Would match to AMDs Pitcairn between $199 - $399 slide: http://images.ht4u.net/reviews/2012/amd_radeon_hd_7700_test/product_pricing.jpg

Yes but definitely on the high side of that slide.

Either Kepler is far away, priced in the same way or AMD is not interested in long-term customer relationships.

I dunno maybe someone just took the bogus old rumored prices, doubled them and called it a day? I guess it's more realistic than the old ones though. No 1GB editions?

 

[Silent_Buddha]

Radeon HD 7870 GHz Edition 2G GDDR5 OC Version -> $410;
Radeon HD 7870 GHz Edition 2G GDDR5 -> $400;
Radeon HD 7850 2G GDDR5 OC Version -> $300;
Radeon HD 7850 2G GDDR5 -> $280.

"Nice" prices if true

I'd say incredibly bogus if you compare non-OC versions.

7870 only 50 USD less than the 7950?

and then...

7850 is a whopping 120 USD less than 7870?

Whoever made up those bogus rumors isn't doing a very good job.

Personally, if I was to take a guess 7970 is going to be between 329-349 USD. Potentially as low as 299 USD, but that's fairly unlikely. And 7850 somewhere between 229-289 USD.

Pricing is likely to be governed by where GK104 ends up being priced as I'm expecting 78xx and GK104 to have similar performance with GK 104 being slightly higher but definitely lower than 79xx. BTW - that's just an educated guess and in no way represents actual knowledge of performance.

Regards,
SB

 

[Kaotik]

Well, if you think about it, 6950 is only around 13-14% faster than 6870, while 6870 is around ~20% faster than 6850

 

[mczak]

Hmm just saw this:
http://read2ch.com/r/poverty/1330742326/
It's got a nice pic in it with specs which finally look real (no more of that 24 ROP bs...).
http://www.gdm.or.jp/image/voices_image/201203/02a_800x445.jpg

32 ROPs
16/20 CUs for 7850/7870
860/1000 Mhz
1200Mhz gddr5 (for both)
130W/175W

That's really a 2xCV... (ok only 2.8 billion transistors that's not exactly 2xCV but close enough)
Even power consumption (which looks similar to Barts levels) would be 2xCV...