In my opinion if anyone is running either SLI'd GTX470s, Crossfired HD6870s or anything higher then there's no reason what so ever to upgrade right now. I'm running Battlefield 3 in either 1280x720p frame sequential 3D (2x 1280x720p) or 1920x1080p with everything on ultra with 60fps Vsynced. Same goes for a single HD7970 or GTX680.
Strictly talking about performance I can agree with that, but I upgraded from SLI 470s. I've not only gained a little more performance, but I have cut heat, noise and power consumption by a huge margin. Plus, after selling the 470s the upgrade only cost $200.Originally Posted by Guilty Bystander
For people running 3x 1920x1080/1200, or 3x 2560x1440/1600 is where it's needed. Anything less than 2GB of VRAM in no way is enough to run either of those, and even 2x 2GB 6870's won't be anywhere near fast enough (2x 6970's aren't even enough).
If you're just running a single 2560 monitor or under, than any more than 2x GTX 580 or 6970's is probably overkill.
I'll still waiting on my 2nd GTX 680 and I've been pretty surprised how well a single card handles BF3 at Ultra on 3 1920x1080 monitors. A second *should* make it flawless.
It's most likely just rumor, though who knows, it could be right too.
However, one has to take into account that unlike previous generation, difference between GK104 and GK110 is far bigger than GF104/114 and GF100/110, as GK104 is stripped from most features helping the GPGPU/HPC speed, while GK110 has to excel on that area more than gaming speed.
What makes things even harder, is the fact that with Kepler (at least for GK104 anyway) nVidia moved from using FP32 and FP32/64 capable CUDA cores to separate FP32-only and FP64-only CUDA cores, and if the same is true for GK110, it'll have hard time to squeeze everything required in.
Following is tired speculation, take it as such, and it assumes that GK110 would share many aspects of GK104 which may or may not be true.
If we assume ~500mm^2 die size and that 512bit membus is true, memory controllers alone will eat ~120mm^2 of it. 44mm^2 is taken by PCIE 3.0 bus, so that leaves us with 336mm^2 to play around with.
If we use the suggested specs, it would have 6 "GPC"s (1 "GPC" = 2x SMX + 1x Rasterizer), meaning 222mm^2 for those using GK104 size - however since GK110 has to be fast in GPGPU/HPC, it has to have more FP64 CUDA cores too, making every SMX that much fatter (sadly it's not known how much space 8 FP64 CUDA cores take in GK104, so it's impossible even to estimate how much fatter they need to be)
For the highend GPGPU/HPC part, one assumes 2:1 FP32/64 ratio as given, as 1:1 would just be too heavy and waste of resources in most cases, that would mean using 96 FP64 CUDA cores opposed to 8 in GK104 SMX.
Completely out of my head, I'd guess the GK110 SMX (identical except for 96 instead of 8 FP64 cores) would be anything from 8 to 12mm^2 bigger than GK104 SMX. That would put us at 53-61mm^2 per "GPC", meaning 6 would suddenly take 318-366mm^2, which means the chip would already be running out of space with ~500mm^2 die size, and that's without ROPs or caches even counted in yet, and other possible changes required for better GPGPU/HPC speed.
(source for GK104 unit sizes http://www.ilsistemista.net/index.ph...s.html?start=1 as well as some random quote from some nV rep I think it was, that 1 GK104 SMX is same sized as 1 GF110 SM)
--
So no, I don't think 2304 FP32 CUDA cores is happening, or GK110 will have to use quite a bit different architecture to that of GK104
Hey CB welcome back. [wave]
I found this image. If true GK110 have 71% larger die.
Last edited by dragomirc; 04-12-2012 at 11:50 PM.
ooh nice chip!
I wonder if it would make sense for me to add another GTX580 when prices drop down instead of getting a GTX680..
That's GF110, not GK ;)Hey CB welcome back. [wave]
I found this image. If true GK110 have 71% larger die.
Nvidia have apparently stopped production of the 580, so they may not be on the shelves much longer.
Hopefully when the 580 is gone, their might be 680s in stock. lol
I'm sure things will improve, obviously 28nm demand is high at the moment too. Even at the silly prices retailers here charge for the 680 the cards are flying off the shelves.
Posting Permissions
Reply With Quote