mersenneforum.org  

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

Reply
 
Thread Tools
Old 2012-01-26, 21:11   #12
NBtarheel_33
 
NBtarheel_33's Avatar
 
"Nathan"
Jul 2008
Maryland, USA

111510 Posts
Default

Is that 50-60 years on a single core, or on multiple cores? If the former, how long do you get if you run on multiple cores?
NBtarheel_33 is offline   Reply With Quote
Old 2012-01-27, 02:34   #13
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

859610 Posts
Default

Quote:
Originally Posted by Dubslow View Post
Using James' estimate of 95K GHzDays,
http://home.earthlink.net/~elevensmooth/Billion.html

Uncwilly is online now   Reply With Quote
Old 2012-01-27, 03:06   #14
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

3×29×83 Posts
Default

Classic.

When was that estimate made?

Last fiddled with by Dubslow on 2012-01-27 at 03:10 Reason: sed -e s/.\ /.\\n\\n/ -i
Dubslow is offline   Reply With Quote
Old 2012-01-27, 09:50   #15
ET_
Banned
 
ET_'s Avatar
 
"Luigi"
Aug 2002
Team Italia

2·2,383 Posts
Default

Quote:
Originally Posted by Dubslow View Post
Classic.

When was that estimate made?
AFAIK, the estimate (852 years) was done many years ago, using just one Pentium IV.

And for the records, there is another program actually faster than Prime95 on LL tests: it is CUDALucas, working in parallel on GPUs, with a throughput about double than a 8 cores I5. The limitation is that it still doesn't handle very big exponents (but neither Prime95 does).

Luigi
ET_ is offline   Reply With Quote
Old 2012-01-27, 10:27   #16
axn
 
axn's Avatar
 
Jun 2003

469310 Posts
Default

Quote:
Originally Posted by Dubslow View Post
Using James' estimate of 95K GHzDays
That estimate is way off. At minimum, we should have 100x the GHz-days of a 100M digit exponent (which comes to around 5000 GHz-days). So 500,000 GHz-days will be in the right ballpark.
axn is offline   Reply With Quote
Old 2012-01-27, 12:49   #17
science_man_88
 
science_man_88's Avatar
 
"Forget I exist"
Jul 2009
Dumbassville

26·131 Posts
Default

Quote:
Originally Posted by axn View Post
That estimate is way off. At minimum, we should have 100x the GHz-days of a 100M digit exponent (which comes to around 5000 GHz-days). So 500,000 GHz-days will be in the right ballpark.
at 5 GHz-day/day per core 8 cores would be sufficient to do it in 12500
days by my count about 34 years.
science_man_88 is offline   Reply With Quote
Old 2012-01-27, 14:41   #18
bcp19
 
bcp19's Avatar
 
Oct 2011

2A716 Posts
Default

Quote:
Originally Posted by science_man_88 View Post
at 5 GHz-day/day per core 8 cores would be sufficient to do it in 12500
days by my count about 34 years.
Except that adding additional cores doesn't add lilke that... 4 cores that can do 5 GHzD/Day tied together may only get 12-15 GHzD/Day.
bcp19 is offline   Reply With Quote
Old 2012-01-27, 16:17   #19
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

11100001101012 Posts
Default

Quote:
Originally Posted by axn View Post
That estimate is way off. At minimum, we should have 100x the GHz-days of a 100M digit exponent (which comes to around 5000 GHz-days). So 500,000 GHz-days will be in the right ballpark.
Yes, that does make sense, so that's around 250 years on one of my cores. (500K/100K*50yrs)

@_ET, a GTX 460 gets around 2-2.5x the throughput of one of my cores, but a 580 would obviously get quite a bit more.


If I used all four CPU cores, I think I could get 15 GD/d (being generous here) so that's 250/3=~ 80 years. Either way, OP still shouldn't even think about it.
Dubslow is offline   Reply With Quote
Old 2017-07-26, 19:12   #20
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

26×3×23 Posts
Default Extrapolation from GTX480 run times gives ~14 years for current fastest GPUs w/ software changes.

From running a selection of exponents p under 90M, I obtained a best regression fit to a power law p^2.03 and a (widely extrapolated) forecast of 1305 days ~ 3.57 years for p~10^9 on a GTX480 running CUDALucas 2.05.1. More modern GPUs are up to three times as fast at a given fft length per the mersenne.ca benchmarks, so a billion-bit candidate could be run in about 1.2 years now.

CUDALucas currently supports up to about 1.14 billion as an exponent (1143276383) with its maximum implemented fft length of 65536K. (A GTX480 or other card of 1.5GB or 2 GB VRAM or less won't run that max length, but other existing cards with 3GB or more will run threads benchmarks to the maximum implemented length.) The NVIDIA cufft supports up to twice that, so conceivably CUDALucas could be expanded by some determined and talented programmer to about p=2.25 billion for cards with adequate memory (6-8GB?).

Beyond that it would require either a future increase in what NVIDIA chooses to support in their library, using a custom created fft, or alternate approaches (perhaps a dash of Karatsuba, which would slow things). If NVIDIA were to double the supported cufft length again, and it fit on existing fastest cards, and CUDALucas or equivalents were modified to support larger exponents, run time on one fast GPU would be somewhere around an estimated 13.7 years for p~3.322 billion, an exponent corresponding to a billion decimal digit candidate.

That's one exponent, with tiny chance of being prime. Trial factoring software mfaktc is ready up to 4,294,967,291, but P-1 factoring software CUDAPm1 has the same fft length limits as CUDALucas and so would also require extension, and run times for them longer than current typical LL test times would be justified to qualify a candidate for such a lengthy run.

I think "prime95" might recommend saving money on electricity until faster more electrically efficient gpus come out in the next few years, then paying for the new faster card with the saved cost of electricity. Running a 180W GPU for 14 years is currently in my neighborhood at about a dollar per watt year, ~$2500.
kriesel is online now   Reply With Quote
Old 2017-07-26, 21:46   #21
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2·13·443 Posts
Default

@KenK: Your ~15 years is about the same time estimate I get for my code running on a cutting-edge manycore Xeon server or Knights Landing workstation. (And I actually timed out some 10Kiter partial runs on the smallest several 1-Gdigit-M-number exponents at the smallest FFT length needed for such, 192M.) More cost-efficient would be an 8-core AMD Ryzen, on which such a test would need ~60 years. Not sure how the lifetime cost of ownership math there stacks up vs your GPU.
ewmayer is offline   Reply With Quote
Old 2017-07-26, 23:52   #22
retina
Undefined
 
retina's Avatar
 
"The unspeakable one"
Jun 2006
My evil lair

3×23×83 Posts
Default

Tests can be migrated. You don't need to run the entire test on the same machine. As the years pass you can upgrade the hardware and copy over the current state and continue from there. So the time estimates would be maximums, and will have lower run times when taking into account the expected future advancement in hardware capabilities.
retina is online now   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
New Mersenne primality test Prime95 Miscellaneous Math 19 2014-08-23 04:18
A (new) old, (faster) slower mersenne-(primality) PRP test boldi Miscellaneous Math 74 2014-04-17 07:16
The fastest primality test for Fermat numbers. Arkadiusz Math 6 2011-04-05 19:39
LLT Cycles for Mersenne primality test: a draft T.Rex Math 1 2010-01-03 11:34
Mersenne Primality Test in Hardware Unregistered Information & Answers 4 2007-07-09 00:32

All times are UTC. The time now is 00:31.

Tue Sep 22 00:31:17 UTC 2020 up 11 days, 21:42, 0 users, load averages: 1.50, 1.63, 1.66

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.