Hey all, I will be sending under this WU name some workunits to test a new method that we are developing. If everything goes well it will allow us to simulate more "intelligently" allowing us to see even slower biological events.
Otherwise the system itself is of little interest as we have studied it before. These WU's are for testing the method as we know what to expect from this system. Therefore, the simulations will be very few and rare for the moment.

It's a heads-up in case you are one of the lucky few to get one of these WU's, hehe.

Edit: I forgot to mention, they should be relatively short for the long queue (estimated 3 hours on a 680)

So we're beta testing in the long queue now
I've just downloaded one

Hahah. Nah it's not much of a test on your side. Since it's a well tested system it is only a test on our side. 15 WU's of one step shouldn't harm too much I think ;) The method works locally, we just need to work out any kinks on our side related to using it on GPUGRID.

lol i was one of the lucky guys ^^

http://www.gpugrid.net/workunit.php?wuid=4674960

---

DSKAG Austria Research Team: http://www.research.dskag.at

Looks like I got one back on the 15th.

http://www.gpugrid.net/workunit.php?wuid=4672222

Seems to have gone through with no issues. Have these all been completed now?

"Large" (150 WU's) batch incoming! Getting to some serious stuff now :) Should have same failure rate (close to 0)

I got one these units, and finished it successfully.


http://www.gpugrid.net/workunit.php?wuid=4686336


Wouldn't mind getting another one.

Received one more today.

http://www.gpugrid.net/workunit.php?wuid=4703804

Again, went through smoothly.

Had no failures with these work units on both card types. Links are below. I had not been able to find a reference in the forums, but was curious why the faster 680 card takes more CPU time to finish a work unit as compared to the slower 460 card. This occurs on two different machines. Would think that it would take the same number of CPU cycles to complete a unit. This is similar on all the work units...SANTI, NATHAN, etc. Generally inverse relationship. Takes 3x the GPU time, but 1/3 the CPU time.



GTX460 Run time CPU time
http://www.gpugrid.net/workunit.php?wuid=4703892 27,488.55 3,661.81
http://www.gpugrid.net/workunit.php?wuid=4692713 27,451.38 3,039.59
http://www.gpugrid.net/workunit.php?wuid=4686243 27,534.66 2,908.34


GTX680
http://www.gpugrid.net/workunit.php?wuid=4708332 10,864.27 10,826.50
http://www.gpugrid.net/workunit.php?wuid=4708253 11,257.73 11,142.95
http://www.gpugrid.net/workunit.php?wuid=4708236 11,562.67 11,496.92
http://www.gpugrid.net/workunit.php?wuid=4708233 11,149.81 11,112.88
http://www.gpugrid.net/workunit.php?wuid=4703779 11,015.36 10,970.59
http://www.gpugrid.net/workunit.php?wuid=4679679 10,948.89 10,925.45

Completed another BenAdapt today on my 680, a bit slower than the examples Jeremy gave, but I am running other projects on my CPU as well.

http://www.gpugrid.net/workunit.php?wuid=4708280 12,725s

Are you seeing anything new from these, Stefan, or will it be a while yet before the results are in?

@Matt: Yes, they are actually working brilliantly well :) Next week I will be preparing a presentation for a conference. But as always, it will take long to make the results public through a publication.
@Jeremy: Post it in the GPU subforum. Maybe it's a configuration thing and someone can help you out.

Jeremy Zimmerman wrote:

I had not been able to find a reference in the forums, but was curious why the faster 680 card takes more CPU time to finish a work unit as compared to the slower 460 card.

This has been discussed when the Kepler based cards came out. Before the CUDA4.2 client came out, there was a parameter called SWAN_SYNC, which controlled whether the client should take a whole CPU thread to feed the GPU (speeding up computation a little bit), or not. Now this parameter is automated: it's always effective on Kepler based cards (because they need it to work as fast as they can), and ineffective on others. So this is not a bug, this is a feature. However there were some workunits in the past which didn't activate this feature on Keplers by accident, therefore those workunits were significantly slower than those with the activated feature.