AMD: R9xx Speculation

AMD has recently postponed the launch schedule of its next-generation Radeon HD 6000 series GPUs (Southern Islands) from the original October 12 to November, according to sources from graphics card makers.

http://www.digitimes.com/news/a20100927PD228.html

 

Nature of the design and variation in the approach, but why sudden interest in tessellation?

A couple of pages back it will show the difference in tessellation approach in a set of benchmarks.

lol.. Barts launch was actually pulled forward and GT430's launch date has been set in stone for quite some time. With the coming price-cuts I can wholeheartedly agree, nv needs to do something.

 

Over did as in, they spent a lot of time and resources to have excellent performance with a feature that is hardly even used in current games. Now don't take that as a negative thing because I really would like to see tessellation become more mainstream but it is the case of the chicken or the egg. Nvidia delivered the card/architecture that performs very well with high levels of tessellation but where are the games to show it?

 

Speculation from Barts XT might be and conclude for Cayman XT

Cayman XT
Maybe 190 - 200 W TDP range
(480x4)1920sp or (512x4) 2048sp, i.e. 30 or 32 SIMD's.
385 - 390 sq.mm (40nm mature)
850MHz core GPU
EDIT: Total memory on-board 2048MB GDDR5 (202GB's bandwidth)
Performance wise between GF480 and HD5970

Upon first release MSRP $400 US dollars.

 

We don't even have hard numbers yet. All we can say so far is that Barts XT will generally perform somewhere between HD 5850 and HD 5870 (maybe nearer to the latter?), "probably" push DX11 performance a lot, and launch as a part of the new 68** family.

The most optimistic scenario I can think of is that it (a) actually performs even a little better than HD 5870 in some games and a lot better @ DX11 feature heavy tests; (b) is labelled as HD 6850, and (c) costs $229. Now that shouldn't sound too bad even to your ears.

The most pessimistic scenario I can think of is that it (a) performs only slightly faster than HD 5850 in general; (b) is labelled as HD 6870, and (c) costs $279 - and that would indeed be somewhat disappointing.

It's all about the price/performance ratio - and we know nothing really substantial in that context yet.

Again, we had GTX 470ish type of performance for one year @ $299 now. If they offer me a card with comparable or better performance for $50-70 less, I'll say "thank you very much" and buy that card

The "real" price/performance leaps as seen with RV770 > Juniper won't come before the 28nm shrink during the next year. Once a Bart-ish chip is shrunk to Juniper-ish die size, though, I expect Bart-ish performance for Juniper-ish price pretty much as a given.

 

I don't think, AMD under-delivered. AMD did a usual first-gen implementation which is good for developers to get aquainted with the technique and to produce some decent looking techdemos. That way it has been for years in the industry and it worked reasonably well.

Sure, Nvidia beats them by a large margin under heavy tessellation, but I honestly don't think that is going to influence gaming performance during the HD5ks lifetime that much.

That is not to say that Nvidias architecture seems to get their horsepower to the ground much better in some games: The lower the resolution, the better the Geforces look.

 

Even if egos were hurt, there's no way it could have affected Northern Islands. It taped out in April, just a few days after Fermi was released, according to Charlie (considering the October/November launch, it really couldn't have taped out much later). When AMD saw Fermi's tesselation performance, there was no time to do anything about it.

 

Look again.

On second thoughts, don't bother, the slide is fake.

 

If it's slow enough that developers think "nah, not worth the bother, we'll come back to it in 3 years for the second version of our D3D11 engine" then it's a problem.

Still, there's no decent analysis of performance though I reckon NVidia's performance is what we want, not ATI's.

But I also suspect Evergreen's tessellation performance is a victim of the 40nm fuck-up, with a kludgy fall-back solution implemented as good enough.

 

Ok, let's think in this direction. Very soon there will be no IGPs at all. They are going to be replaced by more powerful Fusion parts. How will you explain (Mianca, probably you have to) that the gap between low end and Barts (for example) becomes narrower, but the modelling scheme becomes wider? I think that Barts should be called even 6500 or 6600 series.

 

with the huge market with dx11 games currently out I guess its a huge problem for amd then to not have the extreme tesselation power today.
there is no ati, only amd.

software laggs behind the gpu evolution, basically dx11 will be maybe next years market due to new engines take time to devlop.
Bf3 comes to mind there.

the 5000 serie will be enough for the upcoming 5 years.
That is how slow software development is.
after all most play at 1680x1050 or 1920x1080.

 

The fusions bandwith cant match 128bit gddr5. Dont have high hopes. It will depend on AMD what gap they want to create betwen low end and first generation fusion.

 

(1) I don't see the gap between high-end and low-end getting any narrower. With Cayman, it's actually getting bigger. Compare Cayman and Caicos once both are out.

(2) I guess that - given the future impact of APUs - there bascially are two directions for current marketing directors in the GPU business to go: Drop the lower-number parts from your line-up (i.e., in the case of AMD, anything *4** and below) and add stuff at the upper end of the scale (i.e., in the case of AMD, bring back single-GPU cards to the *9** territory (last seen for R600). Or: Try to communicate to your customers than your *4** parts basically don't stand for "low end" anymore - but actually imply: "One step above APU level". I'd go with the first option.

(3) On second thought, forget about (1) and (2) - the whole NI line-up was probably planned for 32nm production. Barts arguably wasn't meant to be a 250mm2+ chip in the first place, it was more likely designed to be around the die size of Juniper - but that still was with 32nm production in mind. With the 32nm process canned and the fall-back to 40nm unavoidable, that same 32nm high-midrange design probably just turned out a lot bigger than originally planned when actually realized on 40nm - and the marketing guys had to somehow deal with the consequences.

I really don't think AMD actually planned the upcoming designs to slot into the market segments they're now positioned in. It probably was a very short-term marketing decision made on the basis of existing, hard-silicon realities and performance testing.

If they really clock Barts @ 900Mhz (i.e. an interestingly high core clock for AMD) - that actually looks a lot like a last-minute-attempt to push a chip actually designed for a somewhat lower clock rate to a performance-position where it's a better marketing-wise fit ...

 

http://bbs.expreview.com/viewthread.php?tid=37063&from=recommend_f

 

Good find Kaotik, first leak of official information.

 

Indeed.

 

Excitementtt!!!!!!!!!!!!!! http://www.youtube.com/watch?v=JoR31rQdaC4

 

Another leak from OBR: http://obrovsky.blogspot.com/2010/09/radeons-hd-6800-facts.html

 

Thanks goes to Sampsa Kurri though, as I picked it from his post at MuroBBS

 

So, what?!? Cayman will be released first in October? And Barts later?