NVIDIA Kepler speculation thread

In about 4 hours or so there is a local invite only event where the Green Team CEO will be giving a Q+A session as well as a demo of upcoming Tegra, CUDA and Tesla performance. The preview email gave mention to a smaller process being talked of as well.
Some answers for people.
CUDA demo on unspecified live 28nm card vs GTX 580 was roughly 28% faster. No clue again what card. Also showed 580 vs Tesla cards.
When asked on the new cards launch date I was told "your next 6 paychecks so save up now."
http://www.overclock.net/t/1208362/q-and-a-session-with-nvidia-ceo/0_20#post_16343030
 
Hmm 28% faster sounds really bad. Granted since we know nothing about what chip it was, if it was running (near) final clocks, and not even actually what the cuda code was doing, it doesn't really tell anything, but you'd think nvidia would show something which would show more of an improvement...
 
If it's their midrange chip 28% sounds pretty good to me.

But that's quite unlikely, IMO

edit:
To clarify - this is assuming CUDA performance translates more or less directly to gaming performance
 
Last edited by a moderator:
I guess they aren't planning on getting very competetive with pricing if I need to save my next 6 paychecks? :D
 
From LenzFire, I'm not sure how reliable these alleged specs are (I have some doubt).

Quick summary:
  • Performance: GTX 680 (GK110) @ HD 7970 + ~45%, GTX 660 (GK104) @ ~GTX 580, GTX 650 (GK106) @ ~GTX 560
  • 2:1 hot clock present
 
From LenzFire, I'm not sure how reliable these alleged specs are (I have some doubt).

Quick summary:
  • Performance: GTX 680 (GK110) @ HD 7970 + ~45%, GTX 660 (GK104) @ ~GTX 580, GTX 650 (GK106) @ ~GTX 560
  • 2:1 hot clock present

GK110 performance is clearly way too high if GK104 is @ ~GTX580 level
 
I call it bs, it looks like a mixture of chiphell, obr and a bit charlie and they fillled empty parts like GTX650Ti 224 bit :D how on earth GK104 has more trannies than GF110? it is 256 bit, 32ROPs ok there is extra PCI3 but it's still seems too much..
 
GK110 performance is clearly way too high if GK104 is @ ~GTX580 level
I get a delta of 90% between GK104/GK110. For GF104/GF100 it was roughly 50% and for GF114/GF110 it was even less at 30%.

I dont know why the GK104 is roughly 580 level when it's projected clocks by Lenzfire for core and shader are both exactly 16.5% higher. And they still have hot clocks, not accurate. What Atlantis said above.
 
I call it bs, it looks like a mixture of chiphell, obr and a bit charlie and they fillled empty parts like GTX650Ti 224 bit :D how on earth GK104 has more trannies than GF110? it is 256 bit, 32ROPs ok there is extra PCI3 but it's still seems too much..

It's clear that its bullshit, however it does in fact state 3.4b for GK104 and for GK110 6.4b transistors.
 
Since when is 25% = 50%?

my bad was probably looking athe the 660. I wonder, would you disable 25% of a 550mm2 chip to hit performance 10% above your own 290mm2 chip? That part makes no sense for a number of reasons. (or the surrounding ones don't)
 
Looks stupid. Last time the GF104 was 384 shaders vs 512, it would make more sense for Nvidia to follow the same path and release GK110 with 1024 shaders and GK104 with 768 with GK106 coming in at 512 shaders. Going straight from 1024 to 512 in a single jump would leave Nvidia nursing losses on the 660Ti as they have outlined it. I don't see it happening.
 
Back
Top