590 gtx + two 580 gtx - scalability/performance test

[acc24ex]

scalability testing

590 gtx @ 1186 mhz
580 gtx @ 1507 mhz
both with 512 cores as you all know it

a test at my interior scene @ 1280x720

1x 590 gtx = 2.38 Ms/sec
2x 580 gtx = 2.9 Ms/sec

590 gtx + 2x580 gtx = 4.6 Ms/sec

obviously all GPUs are synchronized to the slower 590 gtx cores, but works pretty much fine on Octane 1.0 beta 2.46b, so far no issues

didn' see anyone test this setup so far (where's that benchmark scene?)

[mib2berlin]

Nice hardware , it is in the demo suite:

http://www.refractivesoftware.com/downl ... oSuite.zip

Cheers, mib

[GeoPappas]

590 gtx @ 1186 mhz
580 gtx @ 1507 mhz
both with 512 cores as you all know it

GTX 590 should have 2 x 512 cores (or 1024 cores). It's basically 2 GTX 580s smashed together.

So the 590 should run at approximately 2x the 580.

Which video driver are you using? Maybe the older drivers don't recognize all of the cores.

[watanabe]

590 gtx @ 1186 mhz
580 gtx @ 1507 mhz
both with 512 cores as you all know it

GTX 590 should have 2 x 512 cores (or 1024 cores). It's basically 2 GTX 580s smashed together.

So the 590 should run at approximately 2x the 580.

Which video driver are you using? Maybe the older drivers don't recognize all of the cores.

the cores run slower which explains the difference, sli gtx 580s beat out 590s in games as well.

[acc24ex]

I meant 512 cores per GPU as 590 has two, and runs at slower frequency, it was to be expected.. now, how do we overclock

[Lyubomir]

acc24ex, could you test the same 3 configurations and with one 580 only, but with the official benchmark scene(s) and show Ms/sec?

[Timmaigh]

I have gainward 590 as well... clocked to the 669/1338/805

at Trench benchmark it gives at 1920*1200 5,82-5,84 Megasamples

i tried it at 700/1400/805 as well, but only for a few minutes, as i was afraid that the card might explode following reading horrible news about dead cards across the net
anyway at 700 the speed bumped from 5,82 to 6,03 so that is really nice

my GTX470 at work scores with CUDA 3.2 version 2,7 Megasamples at default 607Mhz clock (for comparison)

could you guys please benchmark those gtx580s on the Trench scene as well? Those numbers in the initial post are not very telling, as every scene is different.


Last thing, browsing Nvidia forums i found this topic about GTX590:

http://forums.nvidia.com/index.php?showtopic=197878

basically it says, that the card is far less power hungry on some CUDA scientific app than games/3dmark/heaven benchmark...it might be explained by the fact/assumption only? that some parts of the chips like things called rasterizers or polymorph engines are unused during GPGPU comuting. I wonder if this goes for Octane as well? Not a chance that any of you owns the power consumption measuring device?

[acc24ex]

ok, but first, testing the overclocking utility that came with the software , seems to affect all the GPU clocks not just the 590 one

the benchmark scene with pathtracing on *default came with directlighting !?

2x580 + 590gtx = 9.16 Ms/Sec, 15,27 FPS - all clocks are now running at 1.517 mhz (like they should), temp at 85 C for 590 , and 78C for 580

directlighting = 22.5 Ms/sec, 37 FPS

with pathtracing on
590 gtx = 5.83 Ms/sec
single 590gtx core = 3.18 Ms/sec
580 gtx = 3.13 Ms/sec



using the latest drivers
where is that latest benchmarking thread, can't seem to find it
now for a restart and at default settings..

[acc24ex]

and defaults

pathtracing ON:
590 = 4.81 Ms/Sec
2x 580 = 5.74 Ms/Sec
single 580 = 3.18 Ms/Sec
single 590 core = 2.6 Ms/Sec

590 + 2x580 = 7,84 Ms/Sec w pathtrace, 21.2 directlighting

I would guess there is about a drop at about 17 percent from using a 4x580gtx setup (the overclock at 1.5 was about right), now, with a cooling setup I could probably run that at overclock full time, it's too easy to overclock it.. but runs hottter offcourse.. I think I'm quite happy with the setup, the interior scenes still take like an hour, instead of four..
I seem to find the render time increases a bit when you do stuff like, make objects completely or partially transparent, seems like when you make it completely transparent the speed goes back up a bit.. or make a emitter transparent..

[Timmaigh]

You clocked GTX590 to 777MHz?
Did you bump voltage to get there, or was it on stock volts? How long did you test it on these settings? Was it stable?

According to many sources and reviews its almost impossible for the card to be stable past 720MHz at stock voltage/ respectively the aggresive OC protection will start up and downclock the card due the excessive power draw... so you need to overvolt it, but thats a risky business apparently, as the card does have a weak power cascade and can burn/die easily...