Benchmarking GPU vs CPU which app to bench and compare?

Hello!

I have been searching through the internet, but can’t find anything useful. I am looking for some application, that can be used to compare the performance between CPU (emu mode) and GPU.

http://www.nvidia.in/page/vmd_on_tesla.html

This looks promising, but how was it done?

It could be some visual / math / hacking (MD5) tool, but not video encoding and Adobe CS4 stuff.

Regards,

Gasper

http://impact.crhc.illinois.edu/parboil.php
https://www.cs.virginia.edu/~skadron/wiki/r…x.php/Main_Page
http://ft.ornl.gov/pubs-archive/shoc.pdf

Thank you for a quick reply!

Will check and report.

G

Any special procedure for these Linux based sources? Because I am doing it on VS 2008.

Thanks!

I have never tried running them on anything other than linux… You could contact the authors of each project directly to see if they can help you with a windows build or you could try modifying the sources themselves (many of them are just a few .cu files that should be more or less platform independent ). I am not currently aware of any releases of CUDA benchmarks for windows.

I am a little curious as to why you want to compare the performance of a CUDA application running on a real GPU vs. running in emulation mode and what value this comparison would have.

Yeah, that comparison is totally meaningless. Emu mode is slow compared to any decent CPU-side app.

OK, thanks for all replies.

First of all, I am just a user with no or minimum C/CUDA knowledge. Let’s go down to my level:)

Secondly, I was told that emu mode is performing well compared to CPU preferred app. I guess it’s not like this.

I am opened to suggestions, all I need to do is compare two, maybe three application that do the same thing, but once on CPU (Q, I7…) and once on a multiGPU (295GTX).

Dare one ask why?

Faculty research.

I am downloading Ubuntu 9.1, is it ok for CUDA?

Never worked with linux before, so maybe I will need assistance;)

Not supported and requires hacking to work. 9.04 is going to be much simpler.

Here or here are the places to ask.

Thanks, I can see that 9.04 is supported on nV page. Will download and try. Now I need a spare HDD:)

if all you really are looking for are some applications that you can run in CPU and GPU mode, HOOMD-Blue ( http://codeblue.umich.edu/hoomd-blue/ ) should be on your list.

hoomd lj_liquid_bmark.hoomd --mode=gpu

(in the benchmarks dir) will run the benchmark on the GPU, while

hoomd lj_liquid_bmark.hoomd --mode=cpu

will run the same job script on the CPU in a relatively well-optimized path.

It runs on windows, linux and mac os x.

Excellent! Just what I was looking for. Thank you.

I will play tomorrow with Parboil on linux. If there is anything like MisterAnderson42 mentioned, it would be great.

If your goal is to demonstrate GPU speedup, then this recent thread has a lot of good advice. On the CPU side, in particular, you really need to build a multi-threaded version with some form of SSE (ideally SSE4) for the comparison to be valid.

Yes, this is my goal. But I am limited to “out of the box” solutions, as I have plenty of other stuff to write down, so I don’t have much time to check the code, try to compile etc. as this is not my primary occupation.

HOOMD is very good and even easier to use. Anything like this would be really cool. Like VMD, but I don’t know how to use it properly t benchmark (like in nV page).

p.s.: can’t get in touch with Parboil developers

I’ve built it a few times and used it in a couple papers. Do you have any specific questions about it?

I never played with linux before.

I installed ubuntu 9.04 yesterday, then playing around nv drivers, then I installed CUDA toolkit, and then system path got somehow corrupted. No bin commands, nothing. Will try tomorrow again.

Could you tell me in quick steps how to run and benchmark with this Parboil?

Thank you!

Hello again,

I am getting this error, when trying to run the benchmark.

[codebox]gas@cuda-test:~/parboil$ ./parboil run mri-fhd cuda large

Parboil parallel benchmark suite, version 0.1

build/cuda/mri-fhd: error while loading shared libraries: libcudart.so.3: cannot open shared object file: No such file or directory

Run failed![/codebox]

Any hint?