GPU to 72 status... - Page 159

Prime95 · 2013-01-24, 23:08

Quote:

Originally Posted by chalsall

But, at the same time, it doesn't make sense right now to take the lowest available 31M candidates to 70 if that means that higher candidates will be assigned for DCing which are only at 69.

I agree. Your statement implies that GPU72 is not ahead of the DC wavefront and is in "catch-up" mode. Your TF to 2^70 working downward until you meet the wavefront is consistent with my suggested rules.

The alternative strategy (and maybe better strategy?) would be for GPU72 to grab all the 31M exponents so that the server hands out 32M exponents that are already TFed to 2^70. This assumes there are enough 32M exponents TF'ed to 2^70 for Primenet to hand out while GPU72 catches up in the 31M area.

chalsall · 2013-01-24, 23:33

Quote:

Originally Posted by Prime95

Your statement implies that GPU72 is not ahead of the DC wavefront and is in "catch-up" mode.

We are no longer ahead of the wave. Not as of the release of mfaktc version 0.20; thanks to Oliver, you and rcv. New tech changes the game.

And please don't forget that we haven't yet heard from those running mfakto, those who are running CC1.x, and those who we hope might be interested in doing this work.

Pete et al, what say you? Should we take all of 31M to 70 first? We can release it immediately upon completion. It would make sense.

petrw1 · 2013-01-25, 03:39

Quote:

Originally Posted by chalsall

We are no longer ahead of the wave. .

This:
http://www.mersenne.info/exponent_st..._7/2/30000000/
tells me that in the last week DC was working only in the 30M range; while TF was in the 33 and 34M ranges. Looks ahead to me???

Quote:

Originally Posted by chalsall

Pete et al, what say you? Should we take all of 31M to 70 first? We can release it immediately upon completion. It would make sense.

If you are only talkng about DC then I (who has an opinion but no real power without a GPU) says TF DC to 70 only where/if you are ahead.

As far as GPU72 project factoring as a whole goes; my "opinion" is more LL-TF.

chalsall · 2013-01-25, 04:00

Quote:

Originally Posted by petrw1

This: http://www.mersenne.info/exponent_st..._7/2/30000000/ tells me that in the last week DC was working only in the 30M range; while TF was in the 33 and 34M ranges. Looks ahead to me???

You also have to consider the PrimeNet Activity Summary report. Please note the 636 current DC assignments in the 31M range.

Quote:

Originally Posted by petrw1

As far as GPU72 project factoring as a whole goes; my "opinion" is more LL-TF.

Indeed. That is optimal.

flashjh · 2013-01-25, 04:21

I'm going to stick to WMS LL-TF 7x-73 unless you need a lot of DC done quickly.

LaurV · 2013-01-25, 05:42

Talking from my angle of view only, and related to "high end" nvidia cards (this excludes mfakto, cuda sm older then 1.3, etc, who can't do GPU-LL):

I take the opportunity to reaffirm again what I am talking here since years: that DC-TF you are arguing here, like 30M to 70 or 33M to 71, for expos with P-1 done? C'mon! That makes no sense! Neither of it.

One LL 30M takes under 20 hours on a gtx580 and a bit longer on 570. To find a factor for this range/bitlevels you need 25-30 hours in average.

Which one is better?

And this, of course, if you don't hit a dry path (like I just did, over 40 hours without any factor! on the LMH range/bitlevel where I am supposed to find a factor every 2 hours or so). Of course, one can try forcing his/her luck.

But high-bit GPU TF for the DC range is not worth. Even if you are "ahead".

For LL-front range the story is different, as TF is still finding factors (eliminating exponents) much faster then TWO LL tests will do. If you find one factor per week, you are still faster than TWO-LL's can eliminate exponents.

But for DC, you compete against ONE LL only, so half of the time, which is also much shorter, as the exponents are lower. You need to find factors about 3-4 times faster to "justify" the TF. Which you never will, especially for exponents which survived P-1, their chances to have factors in your TF range which were missed by P-1 is micro-thin...

LaurV · 2013-01-25, 06:46

Backing it up with numbers: (from the GPU72's "factoring cost" report)

Code:

          70bit          71 bit
expo    trials factors trials factors
-------------------------------------
30M     10,117   115      27    0       
31M         83     0      53    0       
32M     20,716   241     264    5       
33M      9,949   108     422    8       

total   40,865   464     766   13

On a summary calculus, that come to about 420 GHzDays per factor, which is consistent with the "factoring cost" table if you take in calculus the "infinities" that appear in the table (here as "0" factors found). This would be a FULL DAY (and a bit more) work of a gtx580. Therefore, you can clear one exponent per day doing TF here, if you are lucky, or more, or less if you jinx it.

Finding factors might be fun, but doing DC-LL is the "safe" process (not affected by probability/luck/jinx)*: you clear one expo every ~20 hours.
And you are few hours faster than doing TF.

My advice would be that GPU72 keeps a bunch of few 30-33M TF to 70-71 bits for the people who might like to get such assignments (mfakto users, whatever), eventually rotate them regularly as the DC front is progressing, but do not make such a big deal of it. There is no gain TF-ing here for the "heavy" GTX users, they better do CuLu-DC, be faster, and still have a slim chance to find a missed prime, which would INDEED be a wonderful hit!

*beside of the situation when your computer takes fire :D

ckdo · 2013-01-25, 09:30

You know, you can basically read anything into numbers. Relevant example:

"GPU72 has thus far completed 4,607 DCs. On the other hand we have only found 2,748 factors by means of DCTF. Evidently we are doing way too much DCTF already."

On the other hand, I myself have saved 26,157 GHzd (or around 60 GHzd/d) worth of DC tests using a single mid-range GPU. That's around 25% of GPU72's total DC(!) throughput, and I'm not going to get anywhere near that throughput by actually doing those DCs on all the hardware I have available (18 cores and the GPU).

But this is getting off topic. The question at hand was whether we should take 30-32M to 70 or skip that and take everything to 71 starting at 34M. My vote is on the first option.

LaurV · 2013-01-25, 11:43

Of course you can read anything in the numbers, if you trample the logic. You did 68 and 69, that is where you saved your 20k GHzD. It has nothing to do with our discussion. Look to flashjh, for example, who did a DOUBLE amount of 70 bit compared with you, and only saved 2000 GHzD.

Of course, everybody is free to do whatever work he likes. Be my guest to do as many DCTF as you want.... If you look in that table, you will see that I even took 11 expos to 72 myself. But that was "ages ago" when I joined the project and I did not realize how I am wasting my resources. Axn and few others convinced me (the posts are on the forum).

If your card can do DCLL, then 69 is MAX you may want to TF, for this range. Over 69, you clear them faster doing DCLL.

ckdo · 2013-01-25, 14:27

Quote:

Originally Posted by LaurV

If your card can do DCLL, then 69 is MAX you may want to TF, for this range. Over 69, you clear them faster doing DCLL.

Since my CPU does in fact do DCs faster than my GPU (yes, I did benchmark that), I consider myself part of the "my card can not do DCLL" team. How far should I be factoring?

LaurV · 2013-01-25, 14:40

Quote:

Originally Posted by ckdo

Since my CPU does in fact do DCs faster than my GPU

Whoops!

Then you can do TF to how high you like :D as long as you have no alternative...

You may consider doing LLTF however... which is worth doing more then DCTF to 71, (my opinion, I can't argue here, and other people can contradict me; my argument was GPU-DC-TF against GPU-DC-LL only, but as long as you can't do the last...).

But of course, it is entirely up to your preference.

2013-01-25, 05:42	#1744
LaurV Romulan Interpreter Jun 2011 Thailand 2⁶·151 Posts	Talking from my angle of view only, and related to "high end" nvidia cards (this excludes mfakto, cuda sm older then 1.3, etc, who can't do GPU-LL): I take the opportunity to reaffirm again what I am talking here since years: that DC-TF you are arguing here, like 30M to 70 or 33M to 71, for expos with P-1 done? C'mon! That makes no sense! Neither of it. One LL 30M takes under 20 hours on a gtx580 and a bit longer on 570. To find a factor for this range/bitlevels you need 25-30 hours in average. Which one is better? And this, of course, if you don't hit a dry path (like I just did, over 40 hours without any factor! on the LMH range/bitlevel where I am supposed to find a factor every 2 hours or so). Of course, one can try forcing his/her luck. But high-bit GPU TF for the DC range is not worth. Even if you are "ahead". For LL-front range the story is different, as TF is still finding factors (eliminating exponents) much faster then TWO LL tests will do. If you find one factor per week, you are still faster than TWO-LL's can eliminate exponents. But for DC, you compete against ONE LL only, so half of the time, which is also much shorter, as the exponents are lower. You need to find factors about 3-4 times faster to "justify" the TF. Which you never will, especially for exponents which survived P-1, their chances to have factors in your TF range which were missed by P-1 is micro-thin... Last fiddled with by LaurV on 2013-01-25 at 05:45

2013-01-25, 06:46	#1745
LaurV Romulan Interpreter Jun 2011 Thailand 2⁶×151 Posts	Backing it up with numbers: (from the GPU72's "factoring cost" report) Code: 70bit 71 bit expo trials factors trials factors ------------------------------------- 30M 10,117 115 27 0 31M 83 0 53 0 32M 20,716 241 264 5 33M 9,949 108 422 8 total 40,865 464 766 13 On a summary calculus, that come to about 420 GHzDays per factor, which is consistent with the "factoring cost" table if you take in calculus the "infinities" that appear in the table (here as "0" factors found). This would be a FULL DAY (and a bit more) work of a gtx580. Therefore, you can clear one exponent per day doing TF here, if you are lucky, or more, or less if you jinx it. Finding factors might be fun, but doing DC-LL is the "safe" process (not affected by probability/luck/jinx): you clear one expo every ~20 hours. And you are few hours faster than doing TF. My advice would be that GPU72 keeps a bunch of few 30-33M TF to 70-71 bits for the people who might like to get such assignments (mfakto users, whatever), eventually rotate them regularly as the DC front is progressing, but do not make such a big deal of it. There is no gain TF-ing here for the "heavy" GTX users, they better do CuLu-DC, be faster, and still have a slim chance to find a missed prime, which would INDEED be a wonderful hit! beside of the situation when your computer takes fire :D Last fiddled with by LaurV on 2013-01-25 at 06:55

2013-01-25, 11:43	#1747
LaurV Romulan Interpreter Jun 2011 Thailand 2⁶·151 Posts	Of course you can read anything in the numbers, if you trample the logic. You did 68 and 69, that is where you saved your 20k GHzD. It has nothing to do with our discussion. Look to flashjh, for example, who did a DOUBLE amount of 70 bit compared with you, and only saved 2000 GHzD. Of course, everybody is free to do whatever work he likes. Be my guest to do as many DCTF as you want.... If you look in that table, you will see that I even took 11 expos to 72 myself. But that was "ages ago" when I joined the project and I did not realize how I am wasting my resources. Axn and few others convinced me (the posts are on the forum). If your card can do DCLL, then 69 is MAX you may want to TF, for this range. Over 69, you clear them faster doing DCLL. Last fiddled with by LaurV on 2013-01-25 at 11:46

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Status	Primeinator	Operation Billion Digits	5	2011-12-06 02:35
62 bit status	1997rj7	Lone Mersenne Hunters	27	2008-09-29 13:52
OBD Status	Uncwilly	Operation Billion Digits	22	2005-10-25 14:05
1-2M LLR status	paulunderwood	3*2^n-1 Search	2	2005-03-13 17:03
Status of 26.0M - 26.5M	1997rj7	Lone Mersenne Hunters	25	2004-06-18 16:46

2013-01-25, 04:21	#1743
flashjh "Jerry" Nov 2011 Vancouver, WA 1,123 Posts	I'm going to stick to WMS LL-TF 7x-73 unless you need a lot of DC done quickly.

2013-01-25, 09:30	#1746
ckdo Dec 2007 Cleves, Germany 2·5·53 Posts	You know, you can basically read anything into numbers. Relevant example: "GPU72 has thus far completed 4,607 DCs. On the other hand we have only found 2,748 factors by means of DCTF. Evidently we are doing way too much DCTF already." On the other hand, I myself have saved 26,157 GHzd (or around 60 GHzd/d) worth of DC tests using a single mid-range GPU. That's around 25% of GPU72's total DC(!) throughput, and I'm not going to get anywhere near that throughput by actually doing those DCs on all the hardware I have available (18 cores and the GPU). But this is getting off topic. The question at hand was whether we should take 30-32M to 70 or skip that and take everything to 71 starting at 34M. My vote is on the first option.