NVIDIA Kepler

[Elvissuperstar007]

http://nvworld.ru/news/2012/02/07/#kepler-specs-leaked

[gabrielefx]

if
3.4B transistors : 1 GTX580 = 6.4B transistors : 1 GTX580*1.88

in few words:
if my hw config needs 8 hours to create a noiseless 3500x2500 pixels render with 4 GTX580 tomorrow in 4.24 hours I will complete my printable renders.

[t_3]

if
3.4B transistors : 1 GTX580 = 6.4B transistors : 1 GTX580*1.88

in few words:
if my hw config needs 8 hours to create a noiseless 3500x2500 pixels render with 4 GTX580 tomorrow in 4.24 hours I will complete my printable renders.

with only 2gb i will complete just nothing

[pixelrush]

I am hoping the GTX660ti will be a shorter card that will fit in my case.
Pity its 1.5 gb and not 2 though but it might be OK for my purposes.
3x faster renders than what I have now would be nice

[gabrielefx]

if
3.4B transistors : 1 GTX580 = 6.4B transistors : 1 GTX580*1.88

in few words:
if my hw config needs 8 hours to create a noiseless 3500x2500 pixels render with 4 GTX580 tomorrow in 4.24 hours I will complete my printable renders.

with only 2gb i will complete just nothing

oh yeah....waiting for the 4GB version
Teslas will double the memory too

[mbetke]

I like my dual 580 for the stuff I do at the moment and on mid-term time line.

So it really depends on how much the amount of RAM will improve in the next generation. And I dont want to get a tesla or another several thousand euro card.

[mainframefx]

Framebuffer does not equals memory size.
I expect atleast 3 or 4GB for the 680 otherwise a memory bandwidth of 512 Mbit would be a total waste of money for Nvidia. 256bit can serve up to ca. 3GB VRAM.
Maybe we can even expect 5GB or 6GB cards at some point.

[gabrielefx]

Framebuffer does not equals memory size.
I expect atleast 3 or 4GB for the 680 otherwise a memory bandwidth of 512 Mbit would be a total waste of money for Nvidia. 256bit can serve up to ca. 3GB VRAM.
Maybe we can even expect 5GB or 6GB cards at some point.

I don't think that Nvidia will release GTXs with 6GB because is the same amount of the Teslas vram. None will buy the Quadro 6000 or the M295.

[mainframefx]

How much VRAM is built into to those cards is depending on the manufactures Asus, MSI, EVGA etc.
There will be a standard design from Nvidia and a few month later we will see custom layouts with diffrent coolers and diffrent VRAM size.
6GB would be pretty huge of course but not impossible. Nvidia has nothing to say here about the size of the VRAM, they don't "bind" the manufacture to their design.
Tesla cards main feature isn't the amount of VRAM their main argument is the double precision computing without performance loss which is a huge bonus for scientific computing. For such purpose you also need a lot of VRAM, thats why they have so much of it.
But thats not very interesting for "standard" GPGPU.
So I don't see the problem there. The only reaons why it could may not happen is the cost factor.

[Jaberwocky]

Actually the main factor for the card manufacturer is the PC games market.Powering a single monitor for this purpose, even with hi res graphics in the game only requires at max around 1gb of graphics memory.Maybe 2 or 3Gb max if it's designed to power multi monitors.Hence why you'll never see a Geforce card with much more than this.I don't think the card manf. consider the GPGPU market when designing Geforce cards.Games are where the money is for them.