Volunteers needed

[axn]

I have a prototype CUDA program (sadly, not mersenne TF) that I need to benchmark on some serious hardware -- preferably kepler, but any mid-to-high end GPU will do. I need some volunteers who can run the program and report back the performance numbers.

I can provide either a Windows build, or, if you can compile CUDA under windows, I can give you the source code. Those who can spare a few minutes of compute time can either PM me or post here.

Thanks in advance.

[Dubslow]

Will Linux do? I have a GTX 460.

[axn]

Originally Posted by Dubslow

Will Linux do? I have a GTX 460.

I currently only have the Windows build. I am planning to port the changes to Linux over the weekend - at that time, I'll definitely require Linux. But not yet.

[flashjh]

I have GTX 580s and HD 4870 and HD 5870s. If you can use either of these, let me know where to get the file. I can use a build for Win64 except for the 4870 which would require Win32.

[axn]

Originally Posted by flashjh

I have GTX 580s and HD 4870 and HD 5870s. If you can use either of these, let me know where to get the file. I can use a build for Win64 except for the 4870 which would require Win32.

It is CUDA, so only the GTX 580 will work. I have Win32 build (which should work under Win 64 also) for CUDA 4.2. What is the CUDA version that your Nvidia driver supports?

EDIT:- If you can compile CUDA (i.e. have CUDA toolkit and VC++ express installed), I can give you the source.

[flashjh]

Originally Posted by axn

It is CUDA, so only the GTX 580 will work. I have Win32 build (which should work under Win 64 also) for CUDA 4.2. What is the CUDA version that your Nvidia driver supports?

EDIT:- If you can compile CUDA (i.e. have CUDA toolkit and VC++ express installed), I can give you the source.

I can compile CUDALucas, so I should be ok to compile your program. I have CUDA 4.2 drivers installed right now.

[axn]

Originally Posted by flashjh

I can compile CUDALucas, so I should be ok to compile your program. I have CUDA 4.2 drivers installed right now.

Thanks. I've attached the source files and compilation / run scripts for two versions of the code. You might have to change the path to point to you VS installation.

After compiling, use the test script to run the programs. Ctrl-C the program after it runs for about a minute or two, but don't abort the batch. Also note down the second number in the progress line -- that is the performance number I'm interested in (it will fluctuate -- just note down the approximate range of values). After the test is run, open the f<>.txt file and check if both version produced the same output. The "63" should be faster than "64", but should otherwise produce identical test.

In case there is any issue in compiling it, I can send you the pre-built binaries.

[axn]

Posting Win32 binaries and the test scripts

[axn]

Updated binaries with bigger work chunks. For faster GPUs, smaller chunks are giving too much variable performance numbers.

[flashjh]

Originally Posted by axn

Updated binaries with bigger work chunks. For faster GPUs, smaller chunks are giving too much variable performance numbers.

On GTX 580 with AMD 1055T I get:

63: ~5950
64: ~4112 or 4376 (it spent equal time on both)

I used your 'bigger chunks' binary for the test

[axn]

Originally Posted by flashjh

On GTX 580 with AMD 1055T I get:

63: ~5950
64: ~4112 or 4376 (it spent equal time on both)

I used your 'bigger chunks' binary for the test

Thanks. I got set of numbers for GTX 480 also (about 10% slower). Do I hear anyone volunteering a GTX 680? Anyone? Anyone? Bueller?