For once I agree with gstanford.

tviceman's GK104 die size comparison thread on the "other" forum.
http://forums.anandtech.com/showthread.php?t=2230628

GaiaHunter still can't quite believe it, obviously.

Well, I can think of an easy way to improve transistor density and help mitigate a 256 bus size at the same time -- Cache and lots of it. Cache can be packed more densely onto a chip than just about any other useful logic structure. The more cache you have the easier it is to keep your registers well fed and avoid the need to constantly go to main memory.

Look here at the B3D Kepler thread.
We have a mobile Kepler GT640M scoring P6412, while GT540M scores P3500. Both use DDR3 memory.

Anything? No. Some things? Yes.

And here's that same GT640M in a Gigabyte utrabook apparently at CeBIT, the first ultra book with a dedicated gpu.



http://vr-zone.com/articles/gigabyte-announces-its-first-ultrabook-at-cebit-we-go-hands-on/15125.html

http://wccftech.com/nvidia-kepler-gk104 ... ns-leaked/

Pretty specific numbers, I'd say! Looks like we're getting there!

Why would you have a core clock of 705 mhz and a Processor (CUDA core presumably) clock if there are 1536 shaders? With that amount of shaders I'd presume the "nvidia dropped hot clocks" rumor to be true. Perhaps "core clock" is actually "memory clock"?

Either that or there really are 1536 hot clocked shaders in which case Kepler will be a compute monster.

I also have my doubts.. 1536 cores @1.4Ghz? 1536 cores with hotclock in such a small chip? Or maybe NVIDIA invented something like Hyperthreading, with 768 cores behaving almost like 1536, and thus the system actually reports the latter? Just speculating here, because I could see 768 core with hotclock on such a small chip, but having a hard time coping with 1536 o.O

EDIT - Another interesting thing is comparing the performance delta between the HD7870 and HD7950 or even 7970. There is seems to be a huge efficiency gap between Pitcairn and Thaiti. Thaiti is quite a lot bigger, but doesnt really show that in performance. This might be a problem not only for AMD, in terms of their own sales canibalization, but also for NVIDIA. That is, GK104 needs not only to beat Thaiti, but beat it handily, else it could be in a less than confortable position. If this is the reality, I bet NVIDIA is having headaches deciding GK104's MSRP. All in all, Pitcairn might have been AMD's secret weapon to keep prices high. Like Appopin says they might really have a strategy now, which allows them to drive the market, leaving NVIDIA in a merely reacting position. Once again Appopins "predator" expression makes a lot of sense.

VR-Zone is chiming in now, with the same info.
http://vr-zone.com/articles/nvidia-gefo ... 15137.html

It all seems a bit mind boggling to me, 1536 shaders hot clocked. I predicted 768 hot clocked as an evolution of GF114.

We shall see I guess, if true AMD is in for a world of hurt.

Wild speculation here, but I recall reading that nvidia was working on a method of doing texturing using just the shader cores snf apparently it works quite well, but, I don't think it's meant to hit just yet, possibly not even in time for Maxwell, but, if you removed the texturing units from Fermi you would gain a lot of die space for shader cores.

Hmm.. You might be on to something.. I remember NVIDIA sort of denying going for an Unified Shader approach, and then the G80 hit.
And, given that 32nm was cancelled, there wasnt a "real" Fermi refresh, just the bug fixed GF110, so it might make sense for them to present quite a lot new things on Kepler.