Performance comparsion of various GTX cards (including oc)
Posted: Sat Jul 23, 2011 12:16 am
Inspired by this thread (http://www.refractivesoftware.com/forum ... =23&t=7369) and helped by the fact that I have a couple of different Nvidia cards to play with, I did an Octane centered series of performance tests. I used the benchmark scene as posted here (http://www.refractivesoftware.com/forum ... f=9&t=7144), with every card running on the same headless Win7x64 system (with pretty every unneeded process including Aero turned off). I tested with Octane 2.48b and Nvidia 275.33 drivers (CUDA 4.0); the test system had a Core i7 2600K CPU running at 4.4GHz, and 16GB RAM at 1600MHz. I also did some tests at very much lower CPU clock speeds (down to 1600MHz) and RAM speeds (down to 800MHz), even turned off all but one CPU core, but did'nt get any notable differences in terms of Octanes' rendering performance. So it is pretty save to assume that the PC performance has next to no impact on Octane rendering speeds. I also included LuxMark results (available here: http://www.luxrender.net/wiki/LuxMark) to find out, if the performance behaves differently there (and of course ... it did!). One interesting thing was also, that LuxMark seemed to be way more fussy about oc'ed RAM; some clock rates which did'nt hurt Octane let LuxMark crash immediately...
And here are the results of the jury:
PDF: http://t3net.at/misc/octane_bench1.pdf
Remarks: Oc'ing was done with MSI Afterburner (v2.1) only at stock voltage. I did'nt try to extend oc rates to the absolute maximum (therefore all those pretty even numbers), as a few mhz up or down are barely visible at the Ms/sec count; also overclocking of my GTX 550 Ti wasn't possible, as it already sports very high stock clocks. I did'nt have a GTX 580 card for my tests, but as my GTX 590 is exactly twice a GTX 580, i used it to simulate a single GTX 580 (just not for the oc tests, because of the limited max. GPU clock of the GTX 590).
When i posted some thoughts about different cards before doing this test series, I thought RAM clock is pretty much unimportant. Now, i was wrong! I also did some tests to find out, how different clock rates affect the performance, and apparently the performance at a whole is affected by GPU and RAM clock rates differently, BUT – depending on the GPU type – I found that a RAM clock boost may be as useful as oc'ing the GPU core only; sometimes even more important!
Some findings & thoughts:
PS: All that findings and funny numbers may change dramatically while using another than the benchmark scene – I just don't know. If someone lends me a very different scene (preferrably "closed indoor" with more than just 1 light source, which fits into 1GB) I'll try and run the tests against it...
And here are the results of the jury:
PDF: http://t3net.at/misc/octane_bench1.pdf
Remarks: Oc'ing was done with MSI Afterburner (v2.1) only at stock voltage. I did'nt try to extend oc rates to the absolute maximum (therefore all those pretty even numbers), as a few mhz up or down are barely visible at the Ms/sec count; also overclocking of my GTX 550 Ti wasn't possible, as it already sports very high stock clocks. I did'nt have a GTX 580 card for my tests, but as my GTX 590 is exactly twice a GTX 580, i used it to simulate a single GTX 580 (just not for the oc tests, because of the limited max. GPU clock of the GTX 590).
When i posted some thoughts about different cards before doing this test series, I thought RAM clock is pretty much unimportant. Now, i was wrong! I also did some tests to find out, how different clock rates affect the performance, and apparently the performance at a whole is affected by GPU and RAM clock rates differently, BUT – depending on the GPU type – I found that a RAM clock boost may be as useful as oc'ing the GPU core only; sometimes even more important!
Some findings & thoughts:
- At same clock speeds, a GTX 470 is as fast as a GTX 570 – in Octane, but not LuxMark (at least judging benchmark scenes!) – even if it is missing 32 shaders; don't ask me why, I tested it 3 times to be sure...
- All in all the GTX 4xx cards seem to deliver equal or slightly better results than the GTX 5xx cards. Must be some Warp Scheduler effect ;)
- If the RAM clock stays at stock rate, a 25% higher GPU clock gains about 0-20% more Ms/sec, depending on the GPU type.
- With my GTX 570 I had nearly no improvement when I pushed the core clock from 700MHz (3,06Ms/sec) to 800MHz (3,09Ms/sec) ...
but when i boosted the default RAM clock from 1900MHz to 2200MHz, the GTX 570 achieved notably better performance (around +15%) at both 700 and 800Mhz GPU clock (sadly this was way above stable operation); same is true for the GTX 590, where the stock RAM clock is at a very low 1700MHz. - Cards like 460/560/560 Ti achieve a higher performance gain from core frequency overclocking at default RAM clock rates, most probably because the default RAM clock is already relatively high (seems, there is some bandwith headroom for those chips).
- A GTX 560 Ti gains less than 5% additional Ms/sec performance from oc'ing the default 2000MHz RAM clock by 10% to 2200MHz.
- So, the best advice to improve performance with overclocking is: Set every clock rate as high as possible while maintaining stable operation :))
- Based on this results a GTX 470 is with no doubt the card to get, if 1.25 Gigabytes of RAM are enough.
- If you can live with 1.5GB, a GTX 480 (which i sadly couldn't test) should outperform any GTX 560 Ti or GTX 570 and maybe even a GTX 580 for only a little extra money (compared to the GTX 560 Ti), or at an equal price point (compared to the GTX 570).
- If 2GB RAM is needed and the price is the next limiting factor, a GTX 560 Ti is still a solid solution (I personally wouldn't go that way though).
- Last but not least a GTX 570 with 2.5GB is imho preferable over a 560 Ti 2gb because the higher price tag pays back in notable more Ms/sec and also a little extra space for textures ;)
PS: All that findings and funny numbers may change dramatically while using another than the benchmark scene – I just don't know. If someone lends me a very different scene (preferrably "closed indoor" with more than just 1 light source, which fits into 1GB) I'll try and run the tests against it...
which i tried btw (on default volts) and everything looked solid, no crash or explosion 
i am pretty sure it would be unstable in games or 3D mark though, but luckily i do not care about those. Personally, i absolutely love the card, it makes Octane actually usable compared to previous gtx460. I would buy one more, if i had bigger case/better mobo layout.