For those who might be considering a Tesla K20 for crunching, without modifications to the code the card is too slow to justify.

I have one for development reasons and am trying to improve the performance over at www.primegrid.com. There are screenshots of CUDA-Z below.
For double precision work I still recommend a GTX580.
For single precision a card from the 6xx series may be best.

For those who might be considering a Tesla K20 for crunching, without modifications to the code the card is too slow to justify.

Which Cuda app have you been trying?, with your computers hidden no-one can tell.

Claggy

Which Cuda app have you been trying?, with your computers hidden no-one can tell.

Thank you for pointing that out, I had not noticed that. Over at primegrid they are visible with the k20. I am working on optimizing GeneFerCUDA. For Seti I have used the astropulse v6 6.04 cuda_opencl_100 application, which takes about 10 minutes per percent, thus 1000 minutes per workunit, 16 hours 40 minutes.

Well, seems like Seti does not like Kepler GPUs...

Message 1300486

Waited 45 minutes to have this thing download before it crashed.
I am aware of the envrionment setting now.

Well, seems like Seti does not like Kepler GPUs...

Message 1300486

Waited 45 minutes to have this thing download before it crashed.
I am aware of the envrionment setting now.

Your Stock OpenCL AP Wu's were going maximum time exceeded, the parameters supplied are very low so a wide range of GPUs should be able to run them,
so could try putting the following parameters: -unroll 10 -ffa_block 6144 -ffa_block_fetch 1536 into the ap_cmdline_6.04_windows_intelx86__opencl_nvidia.txt file in your project directory and see that improves matters,
and for best performance the app needs a core free too.

Claggy

Thank you for trying to help me on that. My comment regarding Kepler was because of the Fermi work units. K20 is a Kepler GPU with which the Seti Fermi program apparently has an issue.

See thread:
http://setiathome.berkeley.edu/forum_thread.php?id=69857

There is mention of the optimized lunatics versions, but they only list AMD CL on the website. Should I consider these to be applicable for Nvidia as well. I'm confused here.

Hydropower.

Thank you for trying to help me on that. My comment regarding Kepler was because of the Fermi work units. K20 is a Kepler GPU with which the Seti Fermi program apparently has an issue.

See thread:
http://setiathome.berkeley.edu/forum_thread.php?id=69857

There is mention of the optimized lunatics versions, but they only list AMD CL on the website. Should I consider these to be applicable for Nvidia as well. I'm confused here.

Hydropower.

Because of licensing issues with the CPU apps the Lunatics Installer isn't available for download at the moment,
The rest of the apps in the download section are (mostly) rather old, newer OpenCL apps are in the Windows GPU section of the forum, but they are only a few revisions newer than Stock (at the moment),

The Optimised Cuda apps are available elsewhere, at: http://jgopt.org/index.html

For your Kepler GPU you'd want to run the Cuda 5 variant.

Claggy

Thank you for that pointer !

63 seconds per workunit. I'll have to tweak the settings a bit.

Thank you for that pointer !

63 seconds per workunit. I'll have to tweak the settings a bit.

They are High Angle Range Wu's, my factory overclocked GTX460 does those in around 120 seconds, and about 8 minutes for a Normal AR task (of an AR of around 0.44)

Claggy