Is there any chance that those GPU based calculations will be running on WCG - maybe as 'Nutritious Rice for the World Phase 2'?
Thanks for your answer in advance!

We are focussing on examining the results that we have obtained first to see what we've done right, what we've done wrong and how this can be improved. It is premature to talk about a phase 2 until we have looked more closely at the results from phase 1.

From a technical viewpoint, the gpu version of the folding process runs quite stably on ATI cards under Linux. We will be porting it to open-CL for compatibility and testing on NVIDIA cards and to allow the same process to run on unaccelerated machines. It uses more accurate energy functions but is essentially the same process otherwise (except about 100x faster). I'd like to modify the algorithm a bit more to better utilize the gpu but that'll take some time and some funding...

Hong

Hello lhhung -- Nutritious Rice for the World Scientist

Thank you for the update. I've noticed a lot of feedback coming from scientists across all WCG projects and I very much welcome that development. Your update is not an exception. A close scientist-to-cruncher connection/communication bodes well for all of us here at WCG.

As for the use of the GPU to do calculations, I too welcome any push to make that possibility a reality. My ATI Radeon is itching to do some WCG crunching exercises!

[WCG techs: Is there already GPU alpha code? Any chance for a GPU to be on beta test (as alpha code is developed to beta status) as the algorithm is being fashioned towards finalization?]

The promise of a GPU-accelerated crunching (100x faster, you say?) is one chip you could use on the bargaining table to get that needed funding.

Good day
;

We should be submitting a paper in the next month or so regarding the application of GPUs to the analysis of the NRFTW data. One of the major methods involved in evaluating the quality of the structures is to compare them with other structures that have been generated. Accurate structures will tend to share similar features, whereas outliers will be tend to be divergent. This requires the comparison of each structure with every other structure in the ensemble which is a bottleneck when you are talking about 100-200 thousand structures per sequence and 60K+ sequences. The paper describes the use of GPUs for these structural comparisons have resulted in a 500-fold increase in speed over previous implementations.

Terrabytes of compressed data were obtained from the project. As we process this data, we will be making available on the website the most accurate structures from each set of 100-200,000 structures that were generated per sequence. We hope to have some of them up by the end of the summer.

Once we have winnowed through the data and identified the best structures, we will compare them to the structure of known families of proteins and try to identify the function of these unknown genes. The sequences that we chose to model in the NRFTW project have little sequence similarity to known genes and hopefully a structure comparison will give us some insight into their function. As there is so much data to process, we are consulting with rice biologists as to which sequences might be of the most interest and prioritize the analyses to those genes.

Hong

Can you tell us more about which type of ATI hardware are you thinking or you are using?
Are you going to go for accuracy (SP or DP), only OpenCL (4xxx and above, or directly with 5xxx and above), or maybe even Brook+ (for 2xxx and 3xxx).
Any news is good news :)