AMD: R9xx Speculation

[Alexko]

Err, I don't see any performance indications for a 384 SP GF104 in that chart, perhaps you can point it out for me…

You're the one who said +20%, which would put it right under the HD 6870 on this chart, a far cry from the HD 6950.

Note that 384 = 336 + 14%, so we'd need at least a 5% clock increase on top of it to reach that +20% target.

Realistically, I suppose we can expect the GTX 560 to be clocked at 750/1500MHz or so (+11%) but that may well take its power draw somewhere between that of the HD 6950 and 6970, with performance somewhere between that of the 6870 and 6950, likely closer to the former.

 

[NathansFortune]

We've seen Nvidia increase clocks on GF110 quite drastically whilst also enabling all of the SMs, so there isn't anything suggesting they couldn't do the same for GF104.

A GF114 with 384 SPs, more/faster GDDR5 with a higher core clock would get a bigger increase than just the 15% increase in shader power.

Think about it, GF104 was a similar sized chip to Cypress, but with parts disabled and a lower clock giving lower performance. By enabling all of the parts and increasing the core clock they would get performance at or above Cypress levels, which would be very competitive with 6870 and probably 6950 depending on how much of Cayman is disabled for it.

Anyway, this is all getting very OT. Bloody hell the release of Cayman can't come fast enough now!

 

[LordEC911]

We've seen Nvidia increase clocks on GF110 quite drastically whilst also enabling all of the SMs, so there isn't anything suggesting they couldn't do the same for GF104.

I don't think Nvidia will be able to increase the cooling on GF114 as much as they did on GF110, GF104 didn't have the same problems with leakage.

Nvidia have managed to improve performance of Fermi without increasing die size or power by around 10-20%

GF110 is 560mm2...

 

[CarstenS]

GF110 is 560mm2...

Really? Nvidia's giving a different number, actually. And contrary to GF100, I haven't seen any heatspreader being pried off.

 

[Kaotik]

Really? Nvidia's giving a different number, actually. And contrary to GF100, I haven't seen any heatspreader being pried off.

Madshrimps say 550mm^2 and "about 4% bigger than GF100"
http://www.madshrimps.be/news/item/76334

 

[CRoland]

Really? Nvidia's giving a different number, actually. And contrary to GF100, I haven't seen any heatspreader being pried off.

What number did they give? I remember reviews only saying it's around the same size as GF100.

 

[CRoland]

Madshrimps say 550mm^2 and "about 4% bigger than GF100"
http://www.madshrimps.be/news/item/76334

That's unattributed but seems to be from Charlie's article.

 

[zshadow]

Looks like retail cards are floating about now Wonder what the switch is for.

http://www.pic-upload.de/view-808390...vorne.jpg.html

More pics

http://www.hardwareluxx.de/community...0-post170.html

GPU-Z screen (looks like incorrect info, though):

http://img196.imageshack.us/img196/3925/gpuzko.jpg

 

[CarstenS]

That's unattributed but seems to be from Charlie's article.

It is, when you click on go to full article below. Word for word the same.
So, I wonder, what did keep Charlie from taking a picture while making his caliper measurements. It really isn't that hard.

 

[Kaotik]

None of your links work there zshadow

 

[CarstenS]

Try this:
http://www.hardwareluxx.de/community/15863940-post170.html

 

[Mize]

there's a bench in the same thread...

 

[Jawed]

By enabling all of the parts and increasing the core clock they would get performance at or above Cypress levels, which would be very competitive with 6870 and probably 6950 depending on how much of Cayman is disabled for it.

When speeding up GF104 NVidia first has to reach HD6870 performance ... which is a 255mm² die

 

[NathansFortune]

there's a bench in the same thread...

4700 in 3DMark 11.

Anyone have the GTX580 or 5870 3DMark 11 scores to hand. Someone in that thread saying 6970 is only 300 points better than 5870...

 

[Jawed]

That's unattributed but seems to be from Charlie's article.

And any such measurement is subject to 0.5mm of linear error...

I think NVidia said GF110 is slightly smaller than GF100 and I don't know of any reason to believe otherwise. I don't remember NVidia ever saying how big GF100 is, but I think the reviews of GF110 had a number supplied by NVidia.

 

[CarstenS]

4700 in 3DMark 11.

Anyone have the GTX580 or 5870 3DMark 11 scores to hand. Someone in that thread saying 6970 is only 300 points better than 5870...

Pick your poison, but remember: Drivers and CPU are important too!
http://www.pcgameshardware.de/aid,8...-Presets-im-DirectX-11-Test/Grafikkarte/Test/

 

[ECH]

Hold up, is the 6970 only able to do 4770 in 3d11? That can't be right, that's no better then an OC 5870.

 

[NathansFortune]

Well the medium benchmark seems to scale quite linearly with game performance and 4700 for 6970 would put it above 5870, but it was probably with older drivers. I don't think we can draw any conclusions until ATi release their new drivers for the VLIW4 architecture. There may be a significant performance boost to be had.

Edit: A little bit of quick maths puts 1600ALUs at 400SPs in Nvidia terms which actually scales almost exactly to the 3D mark score given in the forum, 4750 for my calculation vs 4770 realworld. I hope there are some decent driver based optimisations to be had...

 

[Jawed]

I think Cayman works on Cat 10.11. Cat 10.11 is out isn't it? Admittedly I'm expecting 10.12 to arrive on 15 December.

So then it's a question of whether AMD knobbled prior drivers and gave 10.12 to reviewers.

Gipsel was reporting VLIW-4 compilations a while back. Not sure what driver that was.

 

[CarstenS]

Not even Barts is officially supported by 10.11.

But there's something else leaked: A Driver from Dec 7: 8.79.6 RC2:
http://forums.guru3d.com/showthread.php?t=334207