GHz-days

[houding]

Quick question - nothing to serious.

I finished a LL earlier the week (M65806621), and received 159.6268 credits.

On mersenne.ca it shows "takes 166.345 GHz-days to do one L-L test".

Why is there a difference?

[Mark Rose]

Each exponent will take a different amount of time depending on how big it is. Larger exponents take more processing.

[ET_]

Originally Posted by Mark Rose

Each exponent will take a different amount of time depending on how big it is. Larger exponents take more processing.

But the question is: why 2 distinct sites give different values for GHz/day on the same exponent...

[Mark Rose]

Originally Posted by ET_

But the question is: why 2 distinct sites give different values for GHz/day on the same exponent...

Well that I can't answer :)

[petrw1]

Originally Posted by ET_

But the question is: why 2 distinct sites give different values for GHz/day on the same exponent...

I am free to stand corrected but I believe it can differ for some of the following:
- The FFT size used is different than expected (a round off test done before the test starts might pick a different one)
- Whether you used AVX or not (32 vs 64 bit, etc)
- What version of Prime95 is used
- Maybe if you use another software product

[VictordeHolland]

M65806621 is right on the edge of the SSE2 FFT size used for credit calculations.

Max expo => FFT
63970000 =>  3440640  (3360K)
65790000 =>  3538944  (3456K)
68130000 =>  3670016  (3584K)

Mersenne.ca assumes a FFT size of 3584K, since 65,806,621 > 65,790,000 . http://www.mersenne.ca/credit.php?showsource=1

With the timings from the same source code you can calculate credits yourself:

Timings
3440640 => 2.0296E-1,
3538944 => 2.0958E-1,
3670016 => 2.1840E-1,

M65,806,621 FFT 3456K:
return ($timing * $exponent / 86400.0);
2.0958E-1 * 65806621 / 86400 = 159.6267549.....

M65,806,621 FFT 3584K:
2.1840E-1 * 65806621 / 86400 = 166.34451.....

[Gordon]

Originally Posted by houding

Quick question - nothing to serious.

I finished a LL earlier the week (M65806621), and received 159.6268 credits.

On mersenne.ca it shows "takes 166.345 GHz-days to do one L-L test".

Why is there a difference?

Of course, back in the good old days the credit used to be given in units that actually meant something you could relate to - P90 CPU years, not this nebulous GHz day thing. Means nothing really, I've checked in 111,615 of these mysterious Ghz day thingies, try putting that in perspective, you know is that a lot or not?

Back when I was ranked 64 in the world (August 2001) I had done 123.72 YEARS worth of work (checking in about 18 months of cpu time per week), that I can relate to and get a feel for.

[Mark Rose]

I think the only fair calculation is flops.

[petrw1]

Originally Posted by Mark Rose

I think the only fair calculation is flops.

No "flops" would bring back bad memories of many of my programming attempts in University.

[science_man_88]

Originally Posted by Gordon

Means nothing really, I've checked in 111,615 of these mysterious Ghz day thingies, try putting that in perspective, you know is that a lot or not?.

Originally Posted by http://www.mersennewiki.org/index.php/FLOPS

The work accomplished by one core of a hypothetical 1GHz Core 2 Duo CPU in one day. One P90 year equals 5.075 GHz-days. 1 TFLOPS equals 500 GHz-days, thus 1 P90 year is about 10 GFLOPS

so lets see (111,615)/5.075 = 21993.10 ....... P90 years.