Points calculations

[matt__]

I'm not necessarily new to Folding@Home, but I'm still new to points calculations - I don't really understand them at all.

From the faq I do understand that the formula is "Points = PPD * (DaysPerWU)". Wouldn't "DaysPerWu" award more points to a system that took longer to complete a work unit?

PPD for the SMP client is 1760 and the GPU3 is 9787. So would this (typically) mean that the GPU client will be awarded more points per work unit? The reason I ask is that one of my team members running the GPU client is really crushing work units relative to the rest of the team (who are SMP only), but his points are significantly lower.

To put it in perspective, as of now he's got 158938 points for 218 WUs. I have 613875 points for 177 WUs.

I'm clearly misunderstanding the points formula - could someone point out what? I realize there are other bonuses for various resources contributed but I think I'm misunderstanding some very basic aspects of the calculation.

[7im]

No, total points assigned to a CPU client work unit is 110 * (DaysPerWu), when they are benchmarking that project #.

For example, they fold a brand new type of work unit on the benchmark computer. That computer takes 4 days to complete the work unit. So then we have 110 * 4 DaysPerWU = 440 total points to complete that work unit.

If you fold that work unit in 4 days, you get 110 PPD, for a total of 440 points. In 2 days, you get 220 PPD, but still a total of 440 points. In 8 days, you get 55 PPD, but still get a total of 440 points.


Where as the GPU2 client is benched at 1500 * DaysPerWU.

Get it?

[matt__]

Ah ha. So if it took the GPU2 reference machine 2.4 hours to complete a given work unit, said work unit would be assigned 150 points. If my system were significantly faster than the ref. machine it would score 150 points at a higher rate (ie: in less than 2.4 hours.)

So are the GPU2 work units easier to compute, relatively speaking? Or is it just really easy for a GPU to compute the work unit? The "110x" and "1500x" weight the scores - I guess the use of a GPU is considered a smaller contribution (despite the greater weight) than CPUs in terms of resources donated to the project? (Accounting for the discrepancy I referred to with the difference in points/WU between me and my teammate?)

[7im]

The benchmark points base (110 or 1500) is a reflection of how much science each client is capable of contributing in a fixed amount of time.

Size of work unit has nothing to do with the relative amount of contribution. It's simply a logical place to slice the data in to smaller pieces so it can be sent out to all of us.

And thus points per work unit also has no comparative value.

PPD is about the only value that is comparable across different types of hardware and different types of clients.

[P5-133XL]

Also, the SMP WU's points are not fixed like the uniprocessor or the GPU WU's. The SMP WU's have a base value and then also get a bonus for completing the WU early. For many, the bonus can be 6x-10X, or even more for very fast machines, the value of the base. So that 110 * 4 = 440 is really not a valid estimate for the vast majority of SMP folders.

My Q6600 (using only 3 cores) takes about 13 hours to complete a normal SMP WU valued at 481 points. Using the base value alone I should be getting about 900 PPD but in actuality I'm getting 481 base points + 1,700 bonus points for a total of aprox. 2,200 points every 13 hours which calculates out to PPD of 3,600

[Baowoulf]

Thanks for the info P5-133XL. I also have a Q6600 and was thinking about trying to get it to do SMP and GPU(Nividia) but hadn't gotten around to see if it was good enough to do it.

[P5-133XL]

If you do, make sure you get a Passkey. Also, if running a SMP/Uni client in addition to the GPU client you should set the SMP/UNI priority to the lowest possible while the GPU client core priority should be slightly higher to get optimum PPD.

Using 4 cores of your q6600 you should do considerably better (You should get a much higher bonus) than me using only 3 of the 4.

[matt__]

Thanks for the clarification!