Min nVIDIA driver for E@h CUDA app?
15 Feb 2011 2:05:07 UTC
Topic 195652
(moderation:
Hello community!
It's a pity, I needed to post here in the Getting Started subforum, I couldn't post in the CUDA thread in the News subforum (0 RAC). :-(
Hmm.. maybe a forum mod could move this message to there:
News : New Einstein@Home Radio Pulsar Search and NVIDIA GPU Code
I was happy as I hear that E@h have finally a CUDA app.
I attached my PCs and was sad that BOINC got:
NVIDIA display driver 26000 or later needed
..from the server.
My machines need to stay with 190.38, because best performance @ S@h/stock cuda23 app.
Why need the E@h CUDA app at least 260.00 driver?
cuda23 started with 190.x
At S@h my machines are slower with the drivers after 190.38 .
If I look here:
http://einstein.phys.uwm.edu/apps.php
Binary Radio Pulsar Search
1.05 (BRP3cuda32)
Arecibo Binary Pulsar Search (STSP)
3.11 (ABP2cuda23)
E@h have two CUDA apps.
Maybe the E@h scheduler could be adjusted?
Which nVIDIA driver support cuda32? 260.x?
So maybe like this:
190.38 to < 260.x send out the cuda23 app.
260.x+ send out cuda23 and cuda32 app.
If not possible..
Maybe it's possible to make an app_info.xml file with the cuda23 app? So I could crunch E@h on my 5 GTX260.
Thanks!
Min nVIDIA driver for E@h CUDA app?
)
Ohh..
Maybe the cuda23 app is this the old which outsourced a small calculation to the GPU?
Only ~ 10 % GPU core usage, or something?
And the new cuda32 app use the whole GPU chip?
RE: Ohh.. Maybe the cuda23
)
Did you know that if you post within 1 hour you can click to edit your old post instead of creating a new one? Yours were only 8 minutes apart.
Hi! The old Arecibo CUDA
)
Hi!
The old Arecibo CUDA 2.3 app is no longer used, the only active CUDA app at E@H is BRP3
The reasons for the minium requirement of CUDA 260,* driver are manifold. The code uses some crucial features that were not yet available for CUDA 2.3. Some other CUDA drivers between 2.3 and 3.2 had bugs on one or another platform that would prevent their use with BRP. I know that the devs had a really hard time finding a way so BRP CUDA would work on all platforms and would cross compile on Linux for the Windows Build (some scientific libraries that are used by BRP are not supported to be compiled other than with gcc I think).
Anyway, the CUDA library for FFT that comes with later versions of CUDA seems to be significantly faster that the 2.3 version, so that alone calls for newer drivers.
Sorry.
But I wonder if there might be optimized SETI CUDA apps which are
a) faster than Seti CUDA 2.3 stock app
b) run with later drivers.
Then you could upgrade your driver, crunch at E@H and even get better performance at SETI.
E.g. check out this site: http://lunatics.kwsn.net/
Happy crunching
HB
If they're still looking for
)
If they're still looking for a version of gcc that will run under Windows, they could check if the Cygwin version of gcc is adequate:
http://cygwin.com/
Or possibly the Cygwin gcc4, if they want a later version.
The current solution is to
)
The current solution is to use mingw to cross-compiler under Linux (so the build tool chain is under Linux, code produced runs under Windows). Never touch a running build system :-)
CU
HBE