Folding@home: Recent instability in GPU2 WU's

We've released a series of new WU's (announced in this forum thread http://foldingforum.org/viewtopic.php?f=52&t=5452&start=0 ). From our statistics, there are fewer EUE's with these WU's than previous ones, but it definitely looks like people who weren't seeing EUE's before are seeing them with this one. I should stress that we release WU's from beta to advanced to all of FAH based on EUE statistics and these were not just "ok", but better than previous ones. I think what was different here is the fact that the EUE profiles were different, i.e. people who didn't get EUE"s before were getting them.

These WU's are important as they represent the major next step from the first generation GPU2 WU's. However, as we've moved to more interesting and complex (and important) protein systems, we're seeing new behavior in the GPU2 core code. We've been working on core modifications to help address this and they are in QA right now.

It will take a few days for the new core to get out of QA, but we think this new core should help, although it's not clear it it will solve all the issues people are seeing. Thus, we are continuing to work to see what's up. However, as I mentioned above, these WU's appear to be EUE-ing less than previous WU's (based on the fraction of returned WU's that EUE'd), which means we are getting a lot of useful WU's back as well. Hopefully with these working WU's as well as the knowledge of which WU's EUE'd, we can work to improve the core further.

It's important for us to stress that the GPU2 client is a "high performance client", which means that it's pretty bleeding edge technology. Indeed, just running any sort of production calculation on a GPU is pretty new ground. For this reason, we've tried to make it clear that this software is experimental and still very much under development (see the discussion on our high performance client download page http://folding.stanford.edu/English/DownloadWinOther ), but we wanted to remind people of this designation and its implications. In time, we should be able to further improve the software, but these are still very early days for running complex calculations like FAH on GPUs.

Finally, we have some some exciting news on the GPU2 front science-wise. With the results we're getting back so far, we've been able to submit our first scientific results from the GPU2 client for peer review. We're very excited about the GPU2 client in general and look forward to working with donors to improve the software further.