Kepler lineup

[Timmaigh]

German site tested and it looks really bad.
http://www.computerbase.de/artikel/graf ... ucomputing

mh, but still no cuda test. this is either directx dc or opencl... so we still have hope

EDIT: and yes, by the way refractive/otoy; if you have a minute or two compile an octane cuda benchmark app, rendering a classy scene and make it known to the world - or let one of the current plugin devs do this work - sales will say big thanks for this extra amount of work, because it's the easiest & cheapest possible promotion you can ever have...

don't let luxmark fill this gap alone

great idea, we absolutely need this. I am sick of pointless reviews, which tell me exactly nothing about the octane performance.

[t_3]

great idea, we absolutely need this. I am sick of pointless reviews, which tell me exactly nothing about the octane performance.

it also would tell the world about octane. luxrender already has made its way into lots of reviews, and of course lots of people know slg only because it is mentioned in reviews. and what octane would have in advantage to the few cuda-only benchmarks out there: it will render a mouthwatering image while benching. c4d and cinebench is a perfect example for that.

... and we will see a lot nvidia-benchmarks this year

[gabrielefx]

I think that the current Cuda release in not compatible with Kepler chips.

For Nvidia all Cuda based rendering softwares are a good business. I will never buy 4 GTX680 to play videogames. For sure I will buy 16 new Keplers.

[bb3d]

Anandtech reviewed the new Kepler card too and there you get an explanation, why Kepler sucks in compute benchmarks:
http://www.anandtech.com/show/5699/nvid ... 0-review/1
On the first page of the review there is a table including the "FP64" performance, which means the double precision floating point performance which is also very important for Octane. While the GTX580 has 1/8 FP32 and the GTX560ti has 1/12 FP32, the GTX680 has only 1/24 FP32. So, although Kepler has 3x more shader processors compared to Fermi, the FP64 performance of those shader cores is 3x slower than Fermi !
Later in the article it reads:
"NVIDIA has made it clear that they are focusing first and foremost on gaming performance with GTX 680, and in the process are deemphasizing compute performance."
It may be disappointing, but it seems, that more compute power is reserved for the high end variant of Kepler, which has to be announced yet.

[t_3]

On the first page of the review there is a table including the "FP64" performance, which means the double precision floating point performance which is also very important for Octane.

from a recent post of karba it looks like octane only uses single precision (in fact he wrote "half precision")...
http://www.refractivesoftware.com/forum ... ion#p85094
but afaik slg is as well using sp math, so still no guess possible

[bb3d]

@T_3:
Thanx for the info, I guessed it would be DP.

[pixelrush]

I think the significance is that Nvidia have moved to the same type of specialisation as with gaming and pro cg.
Geforce for gaming, Quadro for CAD etc. with the Geforce deliberately crippled.
So Geforce for single precision compute and Quadro/Tesla for double precision with Geforce limited by the architecture.
The good news is that Octane is single precision....
I hope Kepler is OK with sp You would like it to manage 25-50% more than Fermi.
It has to be more because otherwise Physx would suffer as well.
Ideally it scales with the number of cuda cores, perhaps 3x cores at half effective utilisation = 1.5 x ??
The Luxrender benchmark is especially poor. Crap drivers or something? Lux issues?
Very disappointing if that's an indicator of how Cuda performance is as well.
Don't sell your 580 just yet....

Edit : some sp vs dp testing here
http://www.brightsideofnews.com/news/20 ... x?pageid=4
Not so bad does look like 1.5x
Nvidia definitely has poor dp performance though due to architecture choices.

[elwisoroarke]

still opencl; the picture might change completely for sp cuda apps like octane. and if not, gamers, cheaply selling their 5xx cards will feed the supply chain for a long while

My master plan in finally in motion...soon all their CUDA cores are belong to us <Dr. Yes laughs maniacally as he pets his Persian kitty> : END SCENE

[tehfailsafe]

Edit : some sp vs dp testing here
http://www.brightsideofnews.com/news/20 ... x?pageid=4
Not so bad does look like 1.5x
Nvidia definitely has poor dp performance though due to architecture choices.

Hm.
If I'm reading this right it's great news!

The last test shows the 680 vs 590 on single point calculations in CUDA, and the 680 beats the 590! That's awesome, the 590 SHOULD be able to operate as well as 2 580s if they weren't downclocked lower than 580s (im guessing power issues? never figured that out), but from other things I've read the 680 is a great overclocking candidate.

So if that's correct, I'm predicting the 680 should put out 7ish Ms/s on the benchmark at stock speeds, close to double the 580 at 4.14 Ms/s and still room for overclocking. Pretty cool if that's accurate and octane performs like Sandra.

[bb3d]

Yeah, I noticed this too and it would proof what Nvidia says about the SP performance of Kepler - it's doubled, from 1,5 TFlops (Fermi) to now 3 TFlops (Kepler).
Therefore, if Nvidia did everything right on the driver side, one GTX680 should replace one GTX590 at a fracture of the energy consumption.
BTW, the big Kepler chip (GK110) is expected to have 4,5 TFlops. This would mean a single chip replacing 3x GTX580