Getting others to do the work on exponents I like (was: Trial Factoring Progress) - Page 2

tuckerkao · 2021-03-06, 12:40

Quote:

Originally Posted by axn

On what hardware?

I only have 1 computer at home, Intel Core i7 4930K CPU @3.40 GHz and Nvidia GeForce GTX 780 which were all purchased in year 2013.

kriesel · 2021-03-06, 14:50

TF depth decisions in GIMPS are properly made based on relative performance of TF, P-1, and PRP on the device doing the TF, which should whenever practical be a GPU, with high TF performance relative to the other computation types.

100M exponent TF 77-78 306.08 GHD https://www.mersenne.ca/credit.php?exponent=100000000&frombits=77&tobits=78&worktype=TF
100M exponent PRP or LL 381.39 GHD https://www.mersenne.ca/credit.php?e...78&worktype=LL

Radeon VII
TF 1113 Ghd/day, 3.712 GHD/d/W, 300 W TDP https://www.mersenne.ca/mfaktc.php
PRP 296 GHD/day, 1.053 GHD/d/W, 300 W TDP https://www.mersenne.ca/cudalucas.php
ratio TF/LL per day = 1113/296 = 3.76; log2(3.76) ~+2 bits more TF depth than for cpu factoring
306.08/1113 = 0.275 days = 6.6 hours TF 77-78 estimate on 100M exponent
381.39/296 = 1.29 days LL estimate on gpu
This gpu should be used on PRP/GEC/proof or P-1 with a recent version of gpuowl, not on TF

GTX780
TF 344 Ghd/day, 1.377 GHD/d/W, 250 W TDP https://www.mersenne.ca/mfaktc.php
LL 38.7 GHD/day, 0.166 GHD/d/W, 250 W TDP https://www.mersenne.ca/cudalucas.php
ratio TF/LL per day = 344/38.7 = 8.89; log2(8.89) ~+3 bits more TF depth than for cpu factoring
306.08/344 = 0.89 days = 21.35 hours TF 77-78 estimate
381.39/38.7 = 9.85 days estimate to LL on gpu
PRP/GEC/proof with recent gpuowl version may be a better use of this gpu than TF.
The low TF and LL GHD/d/W figures make this ~8 year old gpu a candidate for replacement or non-use on GIMPS.
https://www.techpowerup.com/gpu-spec...-gtx-780.c1701
Perhaps replace with a GTX1650, 75W TDP and save a lot on power costs; similar TF/LL ratio to the RTX2080. At $0.12/kwhr, 24/7 running, each watt of load reduction returns $1/year, paying for the upgrade in a bit over a year, a very good return on investment.

GTX1070
TF 747.6, 4.88, 150
LL 47.8, 0.34, 150
ratio TF/LL per day = 747.6/47.8 = 15.64; log2(15.64) ~+4 bits more TF depth than for cpu factoring
this gpu probably should be used on TF not P-1 or primality testing

RTX2080
TF 2623.5, 12.202, 215
LL 60.5, 0.301, 215
ratio TF/LL per day = 2623.5 / 60.5 = 43.36; log2(43.36) ~5.4 bits more TF depth than for cpu factoring
306.08/2623.5 = 0.117 days to TF 77-78 estimate for 100M exponent (2.8 hours)
381.39/60.5= 6.3 days estimate to LL
This gpu should be used on TF, not P-1 or primality testing

LL performance above and at mersenne.ca is reflective of CUDALucas, cllucas and very early gpuowl.
Faster gpuowl PRP performance with recent versions than the above shifts the tradeoff point toward less TF depth than the above, by almost 1 bit level. (In the case of Radeon VII, log2(525/296) ~0.83 bits lower.)

axn · 2021-03-06, 15:20

Quote:

Originally Posted by kriesel

The low TF and LL GHD/d/W figures make this ~8 year old gpu a candidate for replacement or non-use on GIMPS.
https://www.techpowerup.com/gpu-spec...-gtx-780.c1701
Perhaps replace with a GTX1650, 75W TDP and save a lot on power costs; similar TF/LL ratio to the RTX2080. At $0.12/kwhr, 24/7 running, each watt of load reduction returns $1/year, paying for the upgrade in a bit over a year, a very good return on investment.

I second this recommendation. Upgrading to a 1650 (or 1650 super, if possible) is a quick and painless way to increase you TF thruput 3x and drastically reduce your power consumption at the same time. You might even be able to sell off the 780 and recoup a little bit of the cost.

tuckerkao · 2021-03-08, 11:04

Quote:

Originally Posted by axn

quick and painless way to increase you TF thruput 3x and drastically reduce your power consumption at the same time.

My current computer can perform P-1 factoring without too much problems. Most exponents still only require up to 2^76 of the trial factoring, so my GPU can handle those. I felt comforted because Ben Delo is finally willing to run PRP tests for lots of my exponent guesses now -

M103,353,143
https://www.mersenne.org/report_expo...exp_hi=&full=1

M103,378,153
https://www.mersenne.org/report_expo...exp_hi=&full=1

M103,427,143
https://www.mersenne.org/report_expo...exp_hi=&full=1

M103,674,113
https://www.mersenne.org/report_expo...exp_hi=&full=1

M103,737,103 and M103,737,143 and M103,737,173
https://www.mersenne.org/report_expo...3737173&full=1

Viliam Furik · 2021-03-08, 11:13

Quote:

Originally Posted by tuckerkao

I felt comforted because Ben Delo is finally willing to run PRP tests for lots of my exponent guesses now

How many times do people have to explain it to you? Ben Delo is not "willing to run PRP tests for your guesses". He probably doesn't even know you exist, because he is not an active forum user. He just has a lot of AWS machines that get assigned many exponents, including your guesses.

He is willing to run PRP tests for GIMPS. Not for you.

tuckerkao · 2021-03-08, 11:15

Ben Delo skipped M103,737,157 and M103,737,167, this cannot be the coincidence. As far as I understand, he goes by the numerical orders, not jumping around during the regular situation especially when these.2 exponents are not assigned to other people.

Okay, I finished up the P-1 factoring for M103,358,341 around 6 minutes ago and Ben Delo just picked it up several seconds before I edit this comment, he didn't take any other exponents between M103,358,000 and M103,359.000 -
https://www.mersenne.org/report_expo...exp_hi=&full=1

If these were the random server assignments, he would be on numerous exponents across the several thousand blocks, not only on my guesses. Unless you are informing me that I have the invisible fingers which can magically type those exponents into Ben Delo's computers through the Space Impaler thousands of miles away.

kriesel · 2021-03-08, 14:05

https://en.wikipedia.org/wiki/Grandiose_delusions
https://rationalwiki.org/wiki/Don%27t_feed_the_Troll

Ben Delo and many others run mprime or prime95 via PrimeNet connection. It's likely Ben does not see much less select his many many assignments. Billionaires have far more enjoyable ways to spend their time. For throughput of order Ben Delo or curtisc level, lots of scripting runs the show.

Viliam Furik · 2021-03-08, 15:08

Quote:

Originally Posted by tuckerkao

Ben Delo skipped M103,737,157 and M103,737,167, this cannot be the coincidence. As far as I understand, he goes by the numerical orders, not jumping around during the regular situation especially when these.2 exponents are not assigned to other people.

That's most probably because these exponents don't have P-1 done, yet.

He gets what PrimeNet gives him, please, correct your knowledge.

Happy5214 · 2021-03-08, 15:14

Unless any GIMPS users happen to know you personally, the only contributors to the project who know anything about you are the unfortunate forumites who have to sit here and watch you bloviate about your supposed hold on our top crunchers, who needless to say don't fall into either category. Nobody in GIMPS is your personal PRP tester. We contribute because we want to, and we help others (when we do) because we want to. Ben Delo just does a ton of PRP tests, which happens to include your exponents. He has no idea he's doing your exponents. He doesn't know who you are. He's just doing a huge service to GIMPS and the mathematical community. To PrimeNet, you are just another user (don't get me wrong, every useful GHz-day helps), and the fact that you did part of the TF has no bearing on the future assignments of those exponents.

tuckerkao · 2021-03-08, 21:53

Quote:

Originally Posted by Viliam Furik

That's most probably because these exponents don't have P-1 done, yet.

He gets what PrimeNet gives him, please, correct your knowledge.

No doubt why curtisc is still running the PRP test for this already known composite number -
https://www.mersenne.org/report_expo...exp_hi=&full=1

I'll try to finish up the P-1 factoring for more of my exponent guesses, so the server will assign them for more billionaire PRP tests, cannot miss the opportunities.

Viliam Furik · 2021-03-08, 22:08

Quote:

Originally Posted by tuckerkao

No doubt why curtisc is still running the PRP test for this already known composite number -
https://www.mersenne.org/report_expo...exp_hi=&full=1

One of the many computers that are being used at the school where he works simply didn't get the news of it being factored... That happens.

2021-03-06, 14:50	#13
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 2⁴×461 Posts	TF depth decisions in GIMPS are properly made based on relative performance of TF, P-1, and PRP on the device doing the TF, which should whenever practical be a GPU, with high TF performance relative to the other computation types. 100M exponent TF 77-78 306.08 GHD https://www.mersenne.ca/credit.php?exponent=100000000&frombits=77&tobits=78&worktype=TF 100M exponent PRP or LL 381.39 GHD https://www.mersenne.ca/credit.php?e...78&worktype=LL Radeon VII TF 1113 Ghd/day, 3.712 GHD/d/W, 300 W TDP https://www.mersenne.ca/mfaktc.php PRP 296 GHD/day, 1.053 GHD/d/W, 300 W TDP https://www.mersenne.ca/cudalucas.php ratio TF/LL per day = 1113/296 = 3.76; log2(3.76) ~+2 bits more TF depth than for cpu factoring 306.08/1113 = 0.275 days = 6.6 hours TF 77-78 estimate on 100M exponent 381.39/296 = 1.29 days LL estimate on gpu This gpu should be used on PRP/GEC/proof or P-1 with a recent version of gpuowl, not on TF GTX780 TF 344 Ghd/day, 1.377 GHD/d/W, 250 W TDP https://www.mersenne.ca/mfaktc.php LL 38.7 GHD/day, 0.166 GHD/d/W, 250 W TDP https://www.mersenne.ca/cudalucas.php ratio TF/LL per day = 344/38.7 = 8.89; log2(8.89) ~+3 bits more TF depth than for cpu factoring 306.08/344 = 0.89 days = 21.35 hours TF 77-78 estimate 381.39/38.7 = 9.85 days estimate to LL on gpu PRP/GEC/proof with recent gpuowl version may be a better use of this gpu than TF. The low TF and LL GHD/d/W figures make this ~8 year old gpu a candidate for replacement or non-use on GIMPS. https://www.techpowerup.com/gpu-spec...-gtx-780.c1701 Perhaps replace with a GTX1650, 75W TDP and save a lot on power costs; similar TF/LL ratio to the RTX2080. At $0.12/kwhr, 24/7 running, each watt of load reduction returns $1/year, paying for the upgrade in a bit over a year, a very good return on investment. GTX1070 TF 747.6, 4.88, 150 LL 47.8, 0.34, 150 ratio TF/LL per day = 747.6/47.8 = 15.64; log2(15.64) ~+4 bits more TF depth than for cpu factoring this gpu probably should be used on TF not P-1 or primality testing RTX2080 TF 2623.5, 12.202, 215 LL 60.5, 0.301, 215 ratio TF/LL per day = 2623.5 / 60.5 = 43.36; log2(43.36) ~5.4 bits more TF depth than for cpu factoring 306.08/2623.5 = 0.117 days to TF 77-78 estimate for 100M exponent (2.8 hours) 381.39/60.5= 6.3 days estimate to LL This gpu should be used on TF, not P-1 or primality testing LL performance above and at mersenne.ca is reflective of CUDALucas, cllucas and very early gpuowl. Faster gpuowl PRP performance with recent versions than the above shifts the tradeoff point toward less TF depth than the above, by almost 1 bit level. (In the case of Radeon VII, log2(525/296) ~0.83 bits lower.) Last fiddled with by kriesel on 2021-03-06 at 14:51

2021-03-08, 11:15	#17
tuckerkao "Tucker Kao" Jan 2020 Head Base M168202123 2⁴·5·11 Posts	Ben Delo skipped M103,737,157 and M103,737,167, this cannot be the coincidence. As far as I understand, he goes by the numerical orders, not jumping around during the regular situation especially when these.2 exponents are not assigned to other people. Okay, I finished up the P-1 factoring for M103,358,341 around 6 minutes ago and Ben Delo just picked it up several seconds before I edit this comment, he didn't take any other exponents between M103,358,000 and M103,359.000 - https://www.mersenne.org/report_expo...exp_hi=&full=1 If these were the random server assignments, he would be on numerous exponents across the several thousand blocks, not only on my guesses. Unless you are informing me that I have the invisible fingers which can magically type those exponents into Ben Delo's computers through the Space Impaler thousands of miles away. Last fiddled with by tuckerkao on 2021-03-08 at 11:57

2021-03-08, 15:14	#20
Happy5214 "Alexander" Nov 2008 The Alamo City 399₁₆ Posts	Unless any GIMPS users happen to know you personally, the only contributors to the project who know anything about you are the unfortunate forumites who have to sit here and watch you bloviate about your supposed hold on our top crunchers, who needless to say don't fall into either category. Nobody in GIMPS is your personal PRP tester. We contribute because we want to, and we help others (when we do) because we want to. Ben Delo just does a ton of PRP tests, which happens to include your exponents. He has no idea he's doing your exponents. He doesn't know who you are. He's just doing a huge service to GIMPS and the mathematical community. To PrimeNet, you are just another user (don't get me wrong, every useful GHz-day helps), and the fact that you did part of the TF has no bearing on the future assignments of those exponents. Last fiddled with by Happy5214 on 2021-03-08 at 15:24 Reason: Expand

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Automatic fetch of Trial Factoring work for GPU mfakt*	LaurV	GPU to 72	81	2020-12-02 05:17
Simple Script to get Trial Factoring Work	jfamestad	PrimeNet	3	2016-11-06 20:32
Why trial factoring work chopped into chunks?	lidocorc	PrimeNet	4	2008-11-06 18:48
How does the trial factoring work with 15K*2^n-1	jocelynl	15k Search	0	2003-07-11 14:23
How does trial-factoring work?	ThomRuley	Software	5	2003-05-30 20:34

2021-03-08, 14:05	#18
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 16320₈ Posts	https://en.wikipedia.org/wiki/Grandiose_delusions https://rationalwiki.org/wiki/Don%27t_feed_the_Troll Ben Delo and many others run mprime or prime95 via PrimeNet connection. It's likely Ben does not see much less select his many many assignments. Billionaires have far more enjoyable ways to spend their time. For throughput of order Ben Delo or curtisc level, lots of scripting runs the show.