AMD: R9xx Speculation

[Bouncing Zabaglione Bros.]

I'm disappointed there's been so little information on Cayman. C'mon AMD, some of us want to know about the top end cards! We're going to have to do another month of this excruciating speculation?

 

[no-X]

Yes. It's interesting, that early reviewers slides contained some HD6900-hints:

but it was removed in the final version:

 

[DavidGraham]

PCGH also does not take into account the high precision LOD and pure angle invaraint AF that is offered in all modes but not available in NVIDIA hardware.

We all know that when it comes to angle dependency , there was no difference even between HD 4000 and G80 , GT200 in that regard .

 

[Jawed]

Hmm, what's that, Gen 8 for Cayman, not Gen 7? A step at least as big as from 6 to 7?

With that earlier hint from the Anandtech article that 16 SIMDs with 16 ROPs is within 2% of the performance of 14 SIMDs and 32 ROPs, it seems there's a serious problem with ROP efficiency. Maybe that "scalability" problem will be fixed in Cayman.

And maybe that problem isn't just ROPs but rasterisation and thread generation.

 

[fellix]

Hm, are 16 extra ROPs cheaper than two SIMDs?

 

[Jawed]

Hm, are 16 extra ROPs cheaper than two SIMDs?

I was going to ponder the same question. But this whole subject is woolly because then you get into clocks/power/yield stuff.

In RV770, I guess that 16 ROPs are 11.5mm². The trouble with that guess is that the ROPs could be twice as big. One cluster (SIMD + LDS + redundancy + TMU) is 10.8mm² (pretty accurate).

 

[psolord]

Yes. It's interesting, that early reviewers slides contained some HD6900-hints:

http://extrahardware.cnews.cz/files...ijen/barts_radeon_hd_6800_arch/tess_small.png

but it was removed in the final version:

http://extrahardware.cnews.cz/files.../barts_radeon_hd_6800_arch/tess_new_small.png

@Whoever that knows / Can speculate

What's off chip buffering?

================================


Now as for the 68XX series, yes they are great cards, but I still cannot see why they were worthy of the 68XX name. They are basically at the same performance level as the 58XX cards, without introducing anything new performance wise really.

Yes it is a marvel by itself that a quite smaller chip can give the same performance, but I don't really care about that. AMD does.

TDP wise, there's nothing great there either. In order for the smaller chip to reach Cypress performance levels, a rise in clock was necessary, so at the end the performance to power draw ratio does not meet any benefits. Actually according to techpowerup, the 5850 is still king of the hill of the high end cards, in that regard.

Spoiler

I am not complaining really. The cards are solid and fairly priced. Could be a tad better though. Still, a substantial percentage of the money needed for 58XX performance levels, has been shaved off. So I can confidently say "Thank you AMD".

At the end of the day, the 68XX name is still beyond me though.

Here's hoping for some fair performance and prices for the Cayman cards.

 

[GZ007]

Hmm, what's that, Gen 8 for Cayman, not Gen 7? A step at least as big as from 6 to 7?

With that earlier hint from the Anandtech article that 16 SIMDs with 16 ROPs is within 2% of the performance of 14 SIMDs and 32 ROPs, it seems there's a serious problem with ROP efficiency. Maybe that "scalability" problem will be fixed in Cayman.

And maybe that problem isn't just ROPs but rasterisation and thread generation.

Strange is that the overclocked 6850 to 900/4300 is a litle bit faster than 6870. And thats 2 aditional simds for 6870.

 

[AnarchX]

Maybe they got one of the HD 6850 "1120SPs Review Edtions"?

 

[Jawed]

And 2.4% more bandwidth for the overclocked card.

 

[GZ007]

Strange , maybe if AMD makes a native 960 SP card with 16 ROP-s(256-bit) and clock it to 900 MHz it would be same .

 

[Kaotik]

Strange is that the overclocked 6850 to 900/4300 is a litle bit faster than 6870. And thats 2 aditional simds for 6870.

According to some site the cards are gaining a lot of performance for memory OC, not so much for core, so having that extra 100MHz effective memory clock is doing that trick I suppose

 

[ferro]

So, european users, which country has the cheapest cards at the moment, and do they ship to other EU countries?
At least here in Finland only one shop has them listed, not in stock, and the prices are skyhigh (240€+)

The Netherlands at http://www.alternate.nl:

Sapphire HD6870 € 199,90
Sapphire HD6850 € 149,90

 

[Bouncing Zabaglione Bros.]

I am not complaining really. The cards are solid and fairly priced. Could be a tad better though. Still, a substantial percentage of the money needed for 58XX performance levels, has been shaved off. So I can confidently say "Thank you AMD".

At the end of the day, the 68XX name is still beyond me though.

Here's hoping for some fair performance and prices for the Cayman cards.

Yeah, they've just pushed what was the 5850/5870 class performance down to the £150-£180 price level, but kept the same naming.

This means the Cayman cards should come in above them at higher price and performance (but hopefully still costing less that current 5850/5870), with the 6950/6970 name. These should be the things that end up being really exciting in terms of new single-card performance levels at the top end.

 

[Kaotik]

The Netherlands at http://www.alternate.nl:

Sapphire HD6870 € 199,90
Sapphire HD6850 € 149,90

Yeah, Alternate is the cheapest I've found too, though I went through .de, but I still haven't received activation e-mail for my account (been like 30-60 mins now), so I have no clue (yet) how's their shipping.
Mindfactory.de has the same price for 6870 at least, but their shipping to Finland costs 30€ (apparently regardless how small or big the order is, excluding full computercases)

 

[SimBy]

So Cayman is not just Barts scaled up. It's again slightly tweaked arch.

EDIT: or less slightly

 

[TKK]

EDIT: nevermind, wanted to mention how bandwidth-limited the 6870 appears to be, but you guys were faster...

 

[fellix]

Anyone noted this (from Rage3D's review):

Additionally, Bart's drops support for the narrow and wide tent anti-aliasing modes. These modes were generally unused, as they had a tendency to blur the image and remove details instead of sharpening and cleaning. I'm not sure anyone will notice they're not in the drop-down any more. 'Box' pattern multi-sampling is now known as 'Standard', with Edge-Detect remaining. Using more than one AMD Radeon graphics card allows higher levels of Anti-Aliasing to be selected: 16xMSAA and 24xEDAA.

 

[flopper]

 

[R300King!]

One major issue with the sample I received was that it came with 1120 shaders enabled. The HD 6850 specification requires 960 shaders. Unfortunately neither HIS, nor AMD could provide help with getting the board fixed so I had to reconfigure the shaders on my own. If you see other HD 6850 reviews on the web that show surprisingly high performance, please ask the reviewer to check fillrate or using GPU-Z. I know of one other site who received a Sapphire HD 6850 with 1120 shaders enabled, so the problem might be more widespread.

http://www.techpowerup.com/reviews/HIS/Radeon_HD_6850/

Oh yeah, sounds like a horrible problem..getting more shaders than you paid for..everyone, beware!!!!