NVIDIA Kepler speculation thread

From what I've heard the launch is definitely sometime this month..but i havent heard about availability yet.

I hope GTX 680 will finally result in some competition in the market though!

 

No OBR said the launch is not in one week. Nowhere did he say "NOT soon".

 

Hm, I'll be launching tomorrow then.

 

Ok, point taken. Could be tomorrow, could be in two months.

I am wondering though, if the reviewers have the cards already, why do they need to fly over to Santa Clara?

 

OBR is a reviewer, so he obviously has connections.

For example his 7970 review - http://pctuning.tyden.cz/hardware/g...drtic-vykonu-v-testu-asus-radeon-hd-7970-3-gb

 

another die size comparison

Spoiler

http://www.chiphell.com/forum.php?mod=redirect&goto=findpost&ptid=374646&pid=11296632

 

[strike]So the actual die-size could be just a bit over 300mm².[/strike]
edit: They already compared with Tahitis actual die-size.

 

Final spec leaked, so the GK104 will be called 680 GTX after all. Over 2Tflops from the look of it?
http://wccftech.com/nvidia-kepler-gk104-gtx-680-final-specifications-leaked/

 

1536 CCs and a hot clock? I wouldn't doubt either spec by itself, but both?

6.008 GHz memory should give bandwidth essentially equal to that of the GTX 580.

 

Nvidia using faster memory than AMD? Boy, they must have improved their memory controller quite a bit.

1536 CC and hot clock sounds bogus, almost every rumor and tidbit out there says differently.

 

Could be BS to detect leaks. We're talking about 4.3TF/s from a ~320mm^2 die. It would explain Charlie's earlier exuberance but it's way too good to be true.

If they're talking about it in a week it's fair to assume there are final boards and drivers floating about somewhere though. Can't be too long again before final specs are made public.

 

Hmm.. 1536 SP and hot clock. Something does not add up: according to the die size, it would have around 4 billions transistors (compared to GF100's 3b). Now, this is a 33% transistor count increase. Of course, front end could be simplified (but this is unlikely, as there should be more ALUs to be fed), texture units could be coupled to to bigger SP clusters and memory controller is narrower (256 vs 384 bits) but, 300% increase in shader count plus hot clock? I don't know how simplified thay could be made, but this seems quite too optimistic to be true. Now, a 768 SP part with hot clock, I could believe and the transistor count could be even match if we consider they are made à la Fermi.

 

http://www.fudzilla.com/home/item/26179-kepler-256-bit-faster-in-dx11-games-than-7970

Now it's only faster in DX11. Next thing you know it'll only be faster on weekends.

 

Then I wonder if they are some kind of VLIW2 as hinted before, in this case maybe the count could also be right... But weren't they foreseen for the Maxwell architecture?

 

And first time I've heard about gk108 in a long time, suddenly got renamed back from gk107 or do both really exist?
Also the blurb about focus for gk104 being "big OEM, especially notebook manufacturers" doesn't make much sense. Big notebook manufacturers don't have a need for many 100W+ mobile gpus. Looks like there is some confusion with another chip...

 

Yeah GK107 seems to be the big notebook push. GT640m and GT650m are popping up in Samsung and Asus laptops already. Assuming of course that they are actually Kepler based...

 

I loled

 

LOL. I read "should be faster in DX11" as "isn't"

 

GF114 680/670M
GK107 650/640M
GF118/119 630M etc.. later when GK106 is ready nV may launch it as 685M and 675M