New work types for GPU72?

[nucleon]

Originally Posted by firejuggler

M3321935017 takes 91 630.041 Ghz/day to LL once

M60082823 takes 133.476 Ghz/day to LL once ( exemple of reasonnable work -one month?)

ratio is 686.5
Supposing that Moore law keep working for as long as needed, ( double computer output each 2 year) that would put the reasonnable range in 18 to 20 years.
Also, one vote for the P-1 on DC

I get about 10 years until computers get the same rate of what we have now vs those 100M digit exponents.

-- Craig

[diamonddave]

The work saved for DC P-1 is a bit high. I don't think a 45M factor saves 144 GHz/Days

[chalsall]

Originally Posted by diamonddave

The work saved for DC P-1 is a bit high. I don't think a 45M factor saves 144 GHz/Days

Thanks. Missed that in my Observation spider -- it was still assuming that all P-1s were LLs, and thus gave credit for both a LL and a DC saved. Fixed.

[petrw1]

Originally Posted by nucleon

I get about 10 years until computers get the same rate of what we have now vs those 100M digit exponents.

-- Craig

Well if history repeats...
About 10 years ago I got my first LL assignment 33M (10M digits) and it took 1.5 years on a fairly modern PC. Today's 100M digit assignments are again taking about 1.5 years.

[Dubslow]

Originally Posted by petrw1

Well if history repeats...
About 10 years ago I got my first LL assignment 33M (10M digits) and it took 1.5 years on a fairly modern PC. Today's 100M digit assignments are again taking about 1.5 years.

With AVX or CUDALucas it's actually just under a year... bcp19?

[nucleon]

Originally Posted by Dubslow

With AVX or CUDALucas it's actually just under a year... bcp19?

I wouldn't trust a video card without ECC to have a reliable calculation over a year.

I'm getting about say (1 in 7) to (1 in 5) LL DC mismatch. (that's my wild guess estimate). That's per 18-20hours for each DC LL. I'm pretty keen to keep my non TF cards to DC LLs.

BTW If there is an easy way analyze stats, to get exact mis match figure. I'm all ears. Considering all DCs I've done in the last 6 months were GPU completed.

-- Craig

[bcp19]

Originally Posted by Dubslow

With AVX or CUDALucas it's actually just under a year... bcp19?

Mine is slated to finish in Jan 13, which will have been a little under 1.2 years. (Started with 26.6) .075 sec/ iter atm (288 days for a full run atm)

[NormanRKN]

hi !

i´ve a few generell (newbie) questions ;)

1.) if i found a faktor with p.e. mfaktX what means that and what is the next step ?
2.) is the way of factoring better then a LL test or a requirement ?
3.) what do the other programs better or lesser good ?
4.) is p-1 factoring not effective enough then mfaktX or is that "two pair of socks ;) )
5.) why p.e. gpu72 can save LL test with this trial factoring methods.

Norman

[Dubslow]

Originally Posted by NormanRKN

hi !

i´ve a few generell (newbie) questions ;)

1.) if i found a faktor with p.e. mfaktX what means that and what is the next step ?
2.) is the way of factoring better then a LL test or a requirement ?
3.) what do the other programs better or lesser good ?
4.) is p-1 factoring not effective enough then mfaktX or is that "two pair of socks ;) )
5.) why p.e. gpu72 can save LL test with this trial factoring methods.

Norman

We are trying to determine of Mersenne numbers, 2^p-1 for prime p, are themselves prime.

If a number has a factor, it is obviously not prime.

~60% of Mersenne numbers have a small factor. That is what mfaktc tries to do.

P-1 is an alternate method of finding smallish factors that works separately from mfatkc, which is purely Trial Factoring (TF).

If neither TF nor P-1 finds a factor, then the Mersenne number is still probably composite -- a LL test will say for sure. On the other hand, if a factor is found, then the Mersenne number is definitely composite, and there is no need to run an LL test, hence "LL tests saved".

Report all mfaktc results (factor or no factor) to PrimeNet (be sure to log in first for credit).

By the way, the authors of both mfaktc and mfakto (for AMD/OpenCL) are German, and there are a few other German speakers around if English is tough. (Hell, English is tough even for native speakers.)

[chalsall]

Originally Posted by NormanRKN

1.) if i found a faktor with p.e. mfaktX what means that and what is the next step ?
2.) is the way of factoring better then a LL test or a requirement ?
3.) what do the other programs better or lesser good ?
4.) is p-1 factoring not effective enough then mfaktX or is that "two pair of socks ;) )
5.) why p.e. gpu72 can save LL test with this trial factoring methods.

In order:

1. It means that the candidate does not have to be LL tested (twice) in order to prove that it's composite (i.e. It's not prime). The "next step" is to simply report it to PrimeNet, and then that candidate will be considered "cleared".

2. The "bit level" we're trial factoring to has been carefully chosen such that it's more "economical" to run a Trial Factor (TF) test than two LL tests. This is based on the amount of CPU/GPU resources required to TF and the statistical likelihood of finding a factor, vs. simply running the LL tests.

3. Not really sure what you mean by this, but to try to answer... mfaktc is for nVidia cards. mfakto is for OpenCL cards (mainly AMD/ATI). Prime95/mprime can do CPU based trial factoring, but since GPUs came onto the scene it's no longer "cost" effective to do so.

4. Different methodologies. Generally P-1 is done after appropriate TFing, but before LLing.

5. GPU72 is simply a coordination tool for those with GPUs to do the newly appropriate TFing without "stepping on each other's toes" (read: two or more people doing the same work on the same candidate). Later on the other worktypes (P-1, LL, DC) were made available from the system..

5.1. GPU72 saves LL tests by eliminating candidates before they are assigned for LL testing.

Hope that helps clarify things for you.

[NormanRKN]

it´s much clearer now.
thank you for your(s) explanations Dubslow/Chalsall !
forward to the digits
it is very interesting !

Norman