Projects 6802-6804 Experiences.

[GreyWhiskers]

There's a thread on the beta forum discussing this topic. But, since these projects have been thrown to the general populace, and since they are all I've been getting for the last several days, and since I've been able to analyze a good number of these WUs through an Excel pivot table derived from the HFM.net work unit history database, I thought I'd share the numbers and some observations.

a. re: format of pivot table below. Columns are ppd; rows are primary number of points (e.g., 1298), with the projects (e.g., 6802) grouped under the points. The items in the cells are counts of completed WUs in that project getting that ppd efficiency. The data set is 840 successfully completed work units since April up through this morning, July 3 2011.

b. My hardware is a Sandy Bridge i7 2600k on an ASUS p8p67 Deluxe MOBO, with an MSI 560GTX-Ti Twin Frozr GPU@950MHz with 384 CUDA Cores.

c. The benchmarking is done on a GTX 460 768MB GDDR5 with 336 CUDA cores.

d. I think there are enough samples to draw conclusions. On my 560 card, I get quite comparable TPF for all of the 6800, 6801, 6802, 6803, 6804 - 1:21 or 1:22, even though there are fewer points for 6800, 6802, 6803, 6804 projects. That means that I get the exact same number of WUs per day for the whole family, but a differing number of points for each.

e. P6805 is in a class of its own - getting 1:07 tpf, and consequently larger ppds.

f. I'm not sure what to conclude for how projects 6802-6804 run on the reference GTX 460 with fewer cores than on my 560 with more cores - or if the differences are due to the scientific value of the output.

g. I've only included the lower rows of the pivot table for the 68xx projects. On my hardware, the 1xxxx WUs consistently get the right two columns of the pivot table - over 19k ppd.

Interesting exercise. I think these GPU WUs give a lot of stats to analyze. I don't think the differences in points are earthshattering, since the project leads do benchmark them all against the reference GPU, and strive to balance points vs. science produced.

One final note. There is no way I could do analysis like this under the FAH Client V7 yet - by and large, all the stats info are buried in the archived log files. I'm hoping that once v7 comes out of beta and the API/data formats stabilize that the third party developers will produce the appropriate data mining tools.

[GreyWhiskers]

I updated the data set to include some 36 additional WUs in the last few days, included rows in the pivot table which summarize the 912/925 point projects (there are dozens of each I've processed one or two WUs from - so I didn't list each project individually, although they are in the data set, and in the Excel Pivot Table).

Bruce has explained that the project leads benchmark each project on designated hardware, in the case for these Fermi core 15 projects on a GTX460 768 Mbyte. Since my GTX 560 Ti GPU has a different number of shaders than the benchmark, and more memory, it's relative behavior on the different WUs will change. Due to the number of atoms that map into the shaders in the GPU itself, or memory, or clock rates, my mileage on my hardware isn't the same as that on the benchmark machines.

I don't think there is any reason to re-benchmark these projects - it's a miracle that the code works so well on so many disparate sets of hardware.