mersenneforum.org  

Go Back   mersenneforum.org > New To GIMPS? Start Here! > Information & Answers

Reply
 
Thread Tools
Old 2020-03-23, 17:43   #12
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2·3·11·67 Posts
Default

Quote:
Originally Posted by Petnek View Post
Thanks for all your answers.

So far I didn't find answer for this one... For which tests is wise to use GPU? Are GPUs faster on something else than TFing?

In my case I have GTX960, Ryzen 5 2600, i5-2410M and partially i7-6820HQ
Ratios of benchmarks, TF/other, for cpus are typically 0.7 to 1.4 in my experience, while for gpus across many models are generally 11. to 42. This is why TF to 4 or 5 bits deeper on gpus than on what we used to do on cpus a decade ago before mfaktc etc, makes sense. You can see the benchmarks for your GTX 960 at

https://www.mersenne.ca/mfaktc.php
https://www.mersenne.ca/cudalucas.php
256.4/15.5 = ~16.5

You could get over 3 times the throughput in TF for less electricity cost with a GTX1650.
kriesel is offline   Reply With Quote
Old 2020-03-23, 18:24   #13
Petnek
 
Petnek's Avatar
 
Mar 2020
Czechia, Tábor

24 Posts
Default

These sites I'm already looking at. Time when I'll change my GPU is coming really fast.

Is here some way how to compare GPU and CPU performance? GPUs has GHz/day and CPUs ms/iter. Is here some way how to unify these units?

One question regarding to worktodo.txt.
When I have PC connected to PrimeNet and manually added some test. PrimeNet will send another work to that PC when it reaches setted cache? Or will it take into account the time needed for a manually added test and will not request additional test?

@Uncwilly: Yes, I saw some double/triple test on found primes using GPUs.
Petnek is offline   Reply With Quote
Old 2020-03-23, 19:52   #14
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2×3×11×67 Posts
Default

Quote:
Originally Posted by Petnek View Post
These sites I'm already looking at. Time when I'll change my GPU is coming really fast.

Is here some way how to compare GPU and CPU performance? GPUs has GHz/day and CPUs ms/iter. Is here some way how to unify these units?

One question regarding to worktodo.txt.
When I have PC connected to PrimeNet and manually added some test. PrimeNet will send another work to that PC when it reaches setted cache? Or will it take into account the time needed for a manually added test and will not request additional test?

@Uncwilly: Yes, I saw some double/triple test on found primes using GPUs.
GhzD/day throughput changes mildly as exponent and TF bit level change, while the effort of the various TF assignments change as 1/p and 2bitlevel. GhzD/day doesn't change much in other computation types either.

But ms/iter in P-1 or primality testing changes faster than exponent does, due to the required fft length increasing. Computational effort per iteration in a primality test is ~p log p log log p; about p1.1.
Comparing GhzD/day, ms/iter, and hours run time, is sort of like comparing meters to hectares/minute and cubic kilometers. The relationships are task-dependent. The units are not dimensionally comparable.
You might find the various posts linked in https://www.mersenneforum.org/showthread.php?t=24607 concerning "run time" helpful.

Prime95 and mprime when PrimeNet connected will communicate expected completion dates for all items in the worktdo file to the server. More work will not be requested until what's queued for some worker falls below the threshold, which you can adjust.

Last fiddled with by kriesel on 2020-03-23 at 19:53
kriesel is offline   Reply With Quote
Old 2020-03-23, 21:15   #15
Petnek
 
Petnek's Avatar
 
Mar 2020
Czechia, Tábor

1610 Posts
Default

From what I understand, CUDALucas can run LL test on GPU. I can run this also on CPU. How I'll figure out on which it will be faster or more efficient. Will it be better to do this test for example on GTX2080Ti or i9-9900. Or GTX260 () vs. Threadripper. This is what I'm trying to understand.

Thanks for that link. I'm going trough all of that, but its a looooot of informations...
Petnek is offline   Reply With Quote
Old 2020-03-23, 22:28   #16
Petnek
 
Petnek's Avatar
 
Mar 2020
Czechia, Tábor

24 Posts
Default

One more thing. Is here some way how to calculate how long it will take to my computer to do a test? In every calculator I found is again GHz-days, but CPU performance is in ms/iter. Should I use "P4 effective equivalent" written on my CPU page and when using 3 cores for one worker multiply that 3 times?

Ryzen 5 2600 has 3.823 GHz (3.392 GHz P4 effective equivalent) but old i5-2410M has 2.579 GHz (7.378 GHz P4 effective equivalent). New Ryzen has worst effectivity than P4 and almost 10 years old i5 from laptop is 2 times more effective than Ryzen

So in case of Ryzen, 3,392 * 3 = 10,176GHz-days when running 24/7? If yes, what about CPU I'm using just for manual tests? In benchmark page is again ms/iter but that is not in GHz-day unit..
Petnek is offline   Reply With Quote
Old 2020-03-23, 23:17   #17
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

23·311 Posts
Default

Quote:
Originally Posted by Petnek View Post
P4 effective equivalent
This value is highly suspect -- practically worthless
Prime95 is offline   Reply With Quote
Old 2020-07-19, 05:17   #18
nemonusquam
 
Jul 2020

3 Posts
Default

Hello.

In the same line of the original question, I was looking in the Work Distribution Map and there are 244 exponents available to double-check in the 52M - 53M range but as I try to get manual assignments at that range (Double-check LL/PRP tests) I get the error:

Error code: 40
Error text: No assignment available meeting CPU, program code and work preference requirements

Does anyone have a solution?

Thanks.

nemonusquam
nemonusquam is offline   Reply With Quote
Old 2020-07-19, 08:10   #19
Happy5214
 
Happy5214's Avatar
 
"Alexander"
Nov 2008
The Alamo City

3·53 Posts
Default

According to https://www.mersenne.org/thresholds/, exponents in that range (Cats 0-1) are not available for manual testing.
Happy5214 is offline   Reply With Quote
Old 2020-07-19, 14:32   #20
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2×3×11×67 Posts
Default

Quote:
Originally Posted by nemonusquam View Post
there are 244 exponents available to double-check in the 52M - 53M range but as I try to get manual assignments at that range (Double-check LL/PRP tests) I get the error:

Error code: 40
Error text: No assignment available meeting CPU, program code and work preference requirements
Set your prime95/mprime to request LL DC. Depending on its history, it will perhaps get cat 0 or cat 1 DC assignments.
Set your manual assignment requests to get lowest exponents' DC below 57885161. That will contribute to softening up the 55M, 56M, or 57M range in advance.
If enough of us hit LL DC hard we may DC everything below M57885161 within a year.
https://www.mersenne.org/report_milestones/ says 61,525 DC to go. That's an average of less than one per month per active GIMPS participant.

https://www.mersenne.org/thresholds/ shows that above ~57.3M is cat 4 territory currently, so everyone can help.

Last fiddled with by kriesel on 2020-07-19 at 14:35
kriesel is offline   Reply With Quote
Old 2020-07-19, 16:23   #21
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

860610 Posts
Default

Quote:
Originally Posted by Happy5214 View Post
According to https://www.mersenne.org/thresholds/, exponents in that range (Cats 0-1) are not available for manual testing.
Quote:
Originally Posted by kriesel View Post
Set your prime95/mprime to request LL DC. Depending on its history, it will perhaps get cat 0 or cat 1 DC assignments.
To get the Cat 0 (and Cat 1, IIRC), you also need to go to https://www.mersenne.org/thresholds/ and agree to take the lowest exponents. It is an opt-in. Each machine has to prove it is reliable and that it is producing fast enough to meet the standards. I have one that I converted to DC recently that likely won't produce fast enough, but it can still work in Cat 2.
Uncwilly is online now   Reply With Quote
Old 2020-07-21, 16:18   #22
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2·3·11·67 Posts
Default

Quote:
Originally Posted by Petnek View Post
One more thing. Is there some way how to calculate how long it will take to my computer to do a test? In every calculator I found is again GHz-days, but CPU performance is in ms/iter.
For an exponent p using an fft length L benchmarked as x ms/iter,

x ms /iter x p iterations / 1000 / 3600 / 24 = n days per exponent.
At https://www.mersenne.ca/credit.php for the same p, select LL test, get GhzD/exponent.
Divide GhzD/exponent / days/exponent to get GhzD/day for your cpu worker.
Then you can use that figure to roughly estimate the time to complete any p, with a scaling according to about p2.1.
For example, if your cpu takes 2. weeks to do a 111M primality test, expect about 2 x (332/111)2.1 = 20. weeks for a 332M exponent (100Mdigit)

Run time scaling and much more is covered at https://mersenneforum.org/showthread.php?t=24607

Last fiddled with by kriesel on 2020-07-21 at 16:19
kriesel is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Best type of work for lowest exponent work? cappy95833 Software 5 2019-11-21 04:24
how to get cudaLucas work for large exponent? fairsky GPU Computing 5 2013-11-22 10:33
Ill-chosen password restrictions James Heinrich Puzzles 42 2011-11-24 16:03
Different B1 & B2 values chosen on P-1 resume azhad Software 0 2006-12-30 06:36
I've chosen primality test task, but info shows P-1 factoring! Why ? Unregistered Software 3 2004-02-23 16:45

All times are UTC. The time now is 23:30.

Wed Sep 23 23:30:13 UTC 2020 up 13 days, 20:41, 0 users, load averages: 1.91, 1.77, 1.67

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.