Primegrid > GIMPS?

[MooMoo2]

I've just found this on the GIMPS front page:

Teams 355
Users 46858
CPUs 305642
TFLOP/s 37.669
GHz-Days 18834.648

The stats for Primegrid over here: http://boincstats.com/stats/project_graph.php?pr=pg
show that Primegrid averages 67.361 TeraFLOPS.

Does Primegrid really have more computing power than GIMPS, or does BOINC overestimate the flops for their projects?

[mdettweiler]

Originally Posted by MooMoo2

I've just found this on the GIMPS front page:

Teams 355
Users 46858
CPUs 305642
TFLOP/s 37.669
GHz-Days 18834.648

The stats for Primegrid over here: http://boincstats.com/stats/project_graph.php?pr=pg
show that Primegrid averages 67.361 TeraFLOPS.

Does Primegrid really have more computing power than GIMPS, or does BOINC overestimate the flops for their projects?

I would not at all be surprised that PrimeGrid has more overall computing power. While GIMPS is older and better established, PrimeGrid uses BOINC, which is very popular with a lot of novice crunchers due to its universality across multiple projects. Since BOINC scores can be compared across projects, it lends itself more readily to contributors primarily/solely interested in credit competition (a large contingent of the DC community), whereas projects like GIMPS with standalone applications attract more because of interest in the underlying math.

[cheesehead]

Originally Posted by MooMoo2

The stats for Primegrid over here: http://boincstats.com/stats/project_graph.php?pr=pg
show that Primegrid averages 67.361 TeraFLOPS.

Does Primegrid really have more computing power than GIMPS, or does BOINC overestimate the flops for their projects?

There seems to be some confusion about credits over at BOINC. :-)

A recently-edited sentence on their "Computation credit" page ( http://boinc.berkeley.edu/wiki/Computation_credit ) seems to have at least one obvious error. A look at the revision history suggests that there may be more to it than that.

The sentence is:

Originally Posted by BOINC

BOINC's unit of credit, the Cobblestone (named after Jeff Cobb of SETI@home), is 1/200 day of CPU time on a reference computer that does 1,000 FLOPS based on the Whetstone benchmark

There are 86,400 seconds in a day, so 1/200 day = 432 seconds. At a rate of "1,000 FLOPS", that would amount to less than half a million floating-point operations! Hmm...

However, later on that same page is:

(Remember that a 1 GigaFLOP machine, running full time, produces 200 units of credit in 1 day).

So it seems that the first-paragraph sentence should have "... that does 1,000 MegaFLOPS ..." instead of "... that does 1,000 FLOPS ...".

But, not so fast ...

When one consults the Wikipedia article linked from that first quoted sentence, one finds:

Originally Posted by http://en.wikipedia.org/wiki/Whetstone_(benchmark)

The Whetstone benchmark originally measured computing power in units of kilo-Whetstone Instructions Per Second (kWIPS). This was later changed to Millions of Whetstone Instructions Per Second (MWIPS).

So it seems that the Whetstone unit is (ignoring various complications) about 1 MegaFLOPS, and 1,000 MWIPS would be about 1 GigaFLOPS.

Perhaps someone mistook "1,000 MWIPS" for "1,000 FLOPS".

... and then there's some business about incorporating memory use and storage space into BOINC credits ...

[ixfd64]

Another thing is that BOINC now has GPU support, which GIMPS doesn't yet have.

[imwithid]

Originally Posted by ixfd64

Another thing is that BOINC now has GPU support, which GIMPS doesn't yet have.

Which is unfortunate. I just got two new video cards last week (one of my systems' onboard GPU has been acting up with video distortions and a rebate offer led me to get two, one for a slower computer to free up resources). Since I don't play games, it's a shame that these powerful cards cannot be put to use towards GIMPS.

[Uncwilly]

Originally Posted by imwithid

Since I don't play games, it's a shame that these powerful cards cannot be put to use towards GIMPS.

They can. There is factorizing potential in those cards.

See this thread. http://www.mersenneforum.org/showthread.php?t=12827

[imwithid]

See this thread. http://www.mersenneforum.org/showthread.php?t=12827

Thanks for the link. Unfortunately for ATI users, this is work being done on CUDA (Nvidia based cards). It seems that it will be a long time before any progress is done on CAL.

[ixfd64]

There's actually an effort to write factoring code for OpenCL: http://mersenneforum.org/showthread.php?t=13621

You might want to ask the OP about it.

[MooMoo2]

Originally Posted by MooMoo2

I've just found this on the GIMPS front page:

Teams 355
Users 46858
CPUs 305642
TFLOP/s 37.669
GHz-Days 18834.648

The stats for Primegrid over here: http://boincstats.com/stats/project_graph.php?pr=pg
show that Primegrid averages 67.361 TeraFLOPS.

Does Primegrid really have more computing power than GIMPS, or does BOINC overestimate the flops for their projects?

I thought that this would be interesting to revisit exactly 10 years later. As of today -

Primegrid:
https://www.boincstats.com/stats/11/project/detail/
1,900.227 TeraFLOPS
Teams - 3,199 (416 active)
Users - 350,860 (3,224 active)

GIMPS:
https://www.mersenne.org/
1,306.009 TFLOP/s
Teams - 1,438
Users - 227,053

(Other GIMPS Stats: 2,155,555 CPUs today, 653,004 GHz days today, 630 first prime tests yesterday, 192 verified prime tests yesterday, and 3,163 newly factored exponents yesterday)

So it does look like Primegrid is still in the lead, though the percentage of their lead has decreased.

[retina]

All those stats are pointless and useless.

Look for the results. That is all that matters.

[MooMoo2]

Originally Posted by retina

All those stats are pointless and useless.

Look for the results. That is all that matters.

It's useful for comparing like vs. like. As a rough estimate, one month's worth of work in 2010 is approximately the same as one day's worth of work in 2020, which I find to be quite interesting (and expected by the "doubling every ~2 years" statement in Moore's law/observation)