test performance gain 4 cards vs 1

For school we have to test the performance gain of 4 cards vs 1 card. we have 4 9600GTs. Is there a program where u can easy test the difference, like a render program that uses all the cards.
Or does someone have a cuda test that can use multiple cards?


Others may add programs from their field :)

I’m testing the elcomsoft program now. But how can i see if all the cards ar being used?
They want a proof on school that multiple cards work faster than 1.
Now i’m testing the same pdf, the system with the 4 cards has a avarage speed of 95 752 but it has a quadcore.
My laptop has 16 000 but it only has a dualcore. So maybe it’s just the difference in better cpu that scores a higher performance?
How can i get proof that it may be the cards that accelerate the proces?

Well GPU acceleration for PDF is not released yet, so I’m not sure what are you testing :)

With LHC you can specify which GPUs to use with command line: passing -i0 will use only first device, -i1 will use only the second one; passing -i0123 will use all 4 devices. Download lhc and take a look at test.bat inside archive.

Thx. Ah GPU acceleration for pdf doesn’t exist yet. Only for md5 i guess? Could this program be ported to linux (don’t know how this works)

You definitely should read information on website for list of supported formats. LHC is MD5-only but EDPR supports much wider list of formats.

Probably yes but I don’t have time for this right now.

Is that test batch actually designed to find the code? :P
I does work with the 4 cards and you can see the performance gain. So that good :)
But the teacher also want’s proof that on a stripped down linux, the same code can work faster then on a windows machine.
So i have to search on that now.

Yes. It doesn’t print found passwords to stdout, check test.md5.out file.

I cannot provide you Linux version right now. And I’m not sure if you will see any significant performance difference between stripped-down Linux and Windows.

Then you can tell your professor I said he’s wrong. My software, HOOMD, has identical (within the noise) performance on linux and windows when running on the same hardware machine. See the posted benchmarks here: http://www.ameslab.gov/hoomd/benchmarks.html.

I’m not trying to defend windows at all, just trying to point out a fact. Perhaps some CPU programs might end up faster on linux vs windows (this is debatable and not a topic for this forum) due to compiler differences and the amount of stuff running in the background. But with CUDA, we have the same compiler producing the same code and running on the same hardware. Except for the slightest performance differences in the driver (typically a tad extra launch overhead and a slightly lower host<->device bandwidth on windows) there is no reason for the code to run differently on the different OSes.

But when the cpu is at 100 procent, maybe then it will make some difference?
If you use a pdf file with the elcomsoft program, it wil not use the gpu at all? only the cpu?

It depends.
If you use PDF created with Acrobat 9 security (256bit AES) AND you’re searching for user password AND you’re using latest APDFPR 5.0 AND you have CUDA-compatible card than it will use gpu.

Thx. Where does the M stands for in the speed?

usually Mega or 10e6

To be pedantic about scientific notation: that should be 1e6 or 10^6 (or 10**6, depending on your favorite language) :)

Yes, you are right, I don’t think it is pedantic, my reply was just plain wrong :)

Hi people!

I have an idea about testing performance.

The first I remember from a CUDA doc, you should disable SLI if you want to run many cards together with CUDA (don’t know why :D)

I’ve found a GPU NT-LM bruteforcer (trial version) but it shows PPS and that’s enough for performance test :)


try this ;) (edit test.bat and replace with correct file paths, or else system will hang up :P )

my GTX260 gets 650PPS at 576Mhz, for reference ;)