AMD: R9xx Speculation

[keritto]

Not without cutting bus width.

How hard is to read the whole section? (and avoid to prejudicially misquote)

 

[Shtal]

Since ATI dies after HD5870, AMD makes new rulers with CaymanXT - making GPU up to 300W DTP.

 

[GZ007]

Posted yet? saw this posted somewhere else.

Chinese site with some new pictures and some benches:
http://tech.sina.com.cn/h/2010-10-16/06171529120.shtml

The 6870 has 65% more shader troughput than 6850 and ends 15% faster in vantage
Overclocking the 6850 to 900 MHz could be more than enough to reach that (and for a lot cheaper).

 

[SimBy]

if it is correct how did they get a 30% perf increase per mm sq?

If those Vantage and 3D mark 06 numbers are legit it's actually more like 40%. Remains to be seen how it does in games.

 

[Jawed]

This looks way more realistic with Barts having 1280SPs rather than measly 960. (but they fu RV840/Juniper specs 144x4 weird)

That site shows no evidence of having official information of any kind, merely some pictures of cards. I think the specifications are pure speculation.

Additionally the Futuremark numbers appear to be a scam. Compare the numbers to those linked way back here:

http://forum.beyond3d.com/showpost.php?p=1480635&postcount=2847

3 of the 4 scores are identical, with only HD6870 Vantage P being different: old = 16270 and new = 17083.

 

[AnarchX]

At Chiphell was posted an interesting 3DMark Vantage screen which is supposed to be from a HD 6850. The interesting stuff are the feature test.


Feature Test 1 (texture) & Feature Test 6 (perlin noise ALUs) ~ HD 5750 level
Feature Test 2 (Pixel ROPs) & Feature Test 5 (GPU particles, with high vertex shader load) ~ HD 5850 level

It seems to turn out that was only a 800SPs @ 650MHz evaluation or developer sample.
According to Fuad and the pictures of XFX HD 6800 boxes, final HD 6850 clocks 775MHz and has 960SPs.
This allows a boost of ~34% over this sample card.

 

[jimbo75]

How can there be a total lack of solid information < 1 week before release?

Have they put a 6-series block in the old drivers and are going to release the working ones with a steam update?

 

[DarthShader]

The 6870 has 65% more shader troughput than 6850 and ends 15% faster in vantage
Overclocking the 6850 to 900 MHz could be more than enough to reach that (and for a lot cheaper).

65%? I calculate only 42%, did I miss something?

Anyhow, this would show vantage isn't compute bound. 42% in shader output and only 5% in memory gives 15% faster vantage. I wonder why didn't they specify the mem clock at 1200mhz, like cypress. Maybe because it would end up faster?

 

[AnarchX]

If HD 6870 and HD 6850 have both 960SPs and only the clocks differs, there could be differences on other parts.

ROPs: HD 6870 32, HD 6850 24 (needs changes in architecture)

Tessellation: HD 6870 4 triangles every 5 clocks, HD 6850 3 triangles every 5 clocks http://forum.beyond3d.com/showpost.php?p=1482643&postcount=3210

 

[Jawed]

According to Fuad and the pictures of XFX HD 6800 boxes, final HD 6850 clocks 775MHz and has 960SPs.

Link? I don't know of any pictures that have legible specifications.

 

[AnarchX]

Link? I don't know of any pictures that have legible specifications.

http://vga.zol.com.cn/199/1999722.html

But only clock rates are "confirmed" by these: 775M and 900M.

 

[Mize]

I'm sorry, but I find myself totally uninterested in Barts (although I realize it's the money maker). Aren't there ANY good leaks on Cayman??

 

[Kaotik]

Posted yet? saw this posted somewhere else.

Chinese site with some new pictures and some benches:
http://tech.sina.com.cn/h/2010-10-16/06171529120.shtml

Assuming the numbers are right, Barts is...
...40% improvement on perf/mm^2 compared to Cypress...
...12% improvement on perf/mm^2 compared to Juniper...
...15.6% improvement on perf/watt compared to Cypress...
...8% improvement on perf/watt compared to Juniper...
...in 3DMark Vantage

(Edit: assuming 155W TDP for Barts, which may be over or under that in reality)

 

[Mianca]

If those Vantage and 3D mark 06 numbers are legit it's actually more like 40%. Remains to be seen how it does in games.

Based on the Vantage scores, Barts XT delivers ~32,3% better performance than [strike]Cayman[/strike] Cypress XT per mm² and clock.

According to Fuad and the pictures of XFX HD 6800 boxes, final HD 6850 clocks 775MHz and has 960SPs.

If Barts Pro has indeed 960SPs, wouldn't that render The_Mask's idea of 1120SPs for Barts XT rather probable after all? (though I still don't like his Cayman/Antilles specs).

 

[ferro]

http://vga.zol.com.cn/199/1999722.html

But only clock rates are "confirmed" by these: 775M and 900M.

It's odd that their table on page 1 says 725MHz, whereas the sticker on their box shot indicates 775MHz. I wonder if the rest of the table on page 1 is correct...

 

[Kaotik]

Based on the Vantage scores, Barts XT delivers ~32,3% better performance than Cayman XT per mm² and clock.

Check your math, per mm^2 the vantage score is just slightly over 40% better than Cypress XT's. (I'm assuming your Cayman XT there is a typo meaning Cypress since there's no Cayman XT to be seen)

 

[Jawed]

http://vga.zol.com.cn/199/1999722.html

But only clock rates are "confirmed" by these: 775M and 900M.

Thanks.

It's odd that their table on page 1 says 725MHz, whereas the sticker on their box shot indicates 775MHz. I wonder if the rest of the table on page 1 is correct...

Yet another nail in that coffin.

 

[Mindfury]

http://www.chiphell.com/thread-129902-1-1.html

BFBC2 1680X1050 4Xaa 4AF:
6870 Avg 71.8fps Min 47fps
460 Avg 61fps Min 43fps

 

[trinibwoy]

Are those numbers from the same part of the game? If you look at reviews BFBC2 numbers are all over the place depending on what the reviewer tested.

 

[Harison]

http://www.chiphell.com/thread-129902-1-1.html

BFBC2 1680X1050 4Xaa 4AF:
6870 Avg 71.8fps Min 47fps
460 Avg 61fps Min 43fps

Since those numbers are barely better than 5770, I very much doubt them. Of course there could many reasons for such score, like diff. scene, drivers, etc., but still IMO Barts will perform better from the get go.