Kepler lineup

[Leonardo]

To all of us saving money for the new generation of CUDA cards, this might be an interesting read.

http://www.fudzilla.com/home/item/25844 ... p-revealed

The first thing to come to mind (assuming these leaks/ rumours are correct) is that they will have a good increase in the number of cores, nearly doubles the current lineup. For instance, the so-called GTX670, with nearly 900 stream processors, for 500 USD looks nice.

The second thing is a bit of a disappointment regarding the amount of vRAM. Crucial for us, it does not increase a lot, at least in these reference cards.

[t_3]

To all of us saving money for the new generation of CUDA cards, this might be an interesting read.

http://www.fudzilla.com/home/item/25844 ... p-revealed

The first thing to come to mind (assuming these leaks/ rumours are correct) is that they will have a good increase in the number of cores, nearly doubles the current lineup. For instance, the so-called GTX670, with nearly 900 stream processors, for 500 USD looks nice.

The second thing is a bit of a disappointment regarding the amount of vRAM. Crucial for us, it does not increase a lot, at least in these reference cards.

just some important info is missing there: nvidia was to cut down shader clocks by half. so double the cores doesn't mean double the speed compared to fermi. as the kepler design is substantially different from the fermi design, we imo need to wait for first real world examples, what that means for octane...

[Leonardo]

I was looking for more detailed sources with core and shader clocks but they all seem to be fake, or just pure speculation. In any event, you're right in that we have to wait until we see benchmarks with Octane to assess these cards' real value.
However, I don't think these cards are going to have a lot more vRAM, unfortunately.

[face]

2GB on some cards give us later the chance of some 4GB cards.
The same as now the 1.5/3GB cards...

face

[gabrielefx]

waiting for the gtx680 with 4 gb and 1024 cuda cores...probably in Q4

http://lenzfire.com/2012/02/entire-nvid ... ate-43823/

[PeterCGS]

Arent the new cards able to share ordinary RAM? I was under the impression that was one major feature of the new chipset.

[glimpse]

as always =) rumors, speculations, 'leaks' and all other stuff that is neither acurate, nor really trust worthy..

economics works in the way that none of the marketer would like to canibalise their own profits - older products, that stand well enough or other in this case profesional quadro/tesla line.

personaly i don't believe we are going to see gains more than 25-33% of curent line. If the efficiency is going to be touched - this will be nice, but..

Useful cards for us wouldn't seem to reach the market until 4th quarter - reference design only, so when all the tweeked models are going to appear? 1st or 2nd of '13.

i'm looking forward too, maybe for sr-x motherboard to fill up with good cost effetive units fow few years of work, but don't believe that this is going to happen this year.

glimps

[Timmaigh]

So its getting close, gtx680 has 1536 CUDA cores, no hotclocks, 700 -1006 MHz clockspeed and 300 mm2 (compared to 550 mm2 of gtx580). Apparently its performance in games is about the same, perhaps slightly above Radeon 7970. The CUDA performance however, is unknown yet, as the hardware enthusiasts are still cycled in past traditionally test this kind of stuff exclusively in games.

I have heard interesting opinion though, apparently it is basically gtx580 made right. I mean if you could possibly run the gtx580 on 1000 MHz without need for nuclear plant to power it, it would perform about the same as gtx680. BTW there is still distant possibility, there might be other, bigger than 500 mm2 chip in the works as true successor to GF110 chips. Soon we shall see.

[Jaberwocky]

I'd heard that the 1500 Cuda cores have been re-arranged.Theoretically they should provide 3 x the performance in Cuda of the 580 (512 cores).However judging from the historical issues encountered with the Warp Scheduler bottleneck problem encountered on the 460, means that the way they're laid out and utilised on the 680 makes anything from a +10% to a 300% speed up is possible in Octane. Ultimately until someone buys one after Wednesdays launch (assuming they are not all snapped up at launch) and tests it out in octane, we are not really going to know.

[mib2berlin]

Hi, as far I understand this article, the 1536 CUDA cores are split in three parts.

http://www.fudzilla.com/home/item/26393 ... -explained

I think we get 768 CUDA cores for octane, this is 30 % more render power with 50 watt less power consumption.

Cheers, mib.