mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   News (https://www.mersenneforum.org/forumdisplay.php?f=151)
-   -   Merry Christmas and a prime! (https://www.mersenneforum.org/showthread.php?t=22837)

ewmayer 2017-12-27 07:02

[QUOTE=Prime95;474943]A day after Christmas a computer has reported a new Mersenne prime. Unlike the numerous CudaLucas false positives we've had over the last few years, this one shows promise.[/QUOTE]

[QUOTE=Madpoo;474957]Otherwise, my verification won't technically count as a double-check since it's another Prime95 run and we'd still want one with another program (mprime, cudalucas, cllucas, etc).[/QUOTE]

When I read George's OP his wording led me to understand that the prime-reporting run was done using CuLu ... but Aaron, you say it was in fact a Prime95 (or mprime) run?

@Serge: I don't recall if Gord Palameta (a.k.a. GP2) did timings for 16-or-more-core C5 instances, and if so, how those compared to the EC2-instance timings you got for your verify run of the last prime 2 years ago. I'll try to dig up some numbers tomorrow.

a1call 2017-12-27 07:27

A positive PRP test result for the candidate you are testing, would be a relativity quick test with a very strong indication that you do not have a false-positive.
Would it not?

ewmayer 2017-12-27 07:48

[QUOTE=a1call;474966]A positive PRP test result for the candidate you are testing, would be a relativity quick test with a very strong indication that you do not have a false-positive.
Would it not?[/QUOTE]

I don't see the advantage of using a PRP test in this context ... no faster than LL, and on the highly-reliable (typically with ECC RAM) systems usually used by George & Aaron for their 'first look' DCs the kinds of system glitches the Gerbicz error check is designed to catch are unlikely to come into play. Also, if the initial run used LL, using LL for the DC allows direct cross-comparison of intermediate Res64s, if the user's logfile is available. (At least I would hope it would work this way.)

Even if the original run didn't output interim residues in the manner I describe, when George/Aaron do their DC they set the flag to do so, which means if their run also says 'prime', then the official independent-software DCs can do such interim-residue cross-matching ... this is especially important for the GPU clients, which are very fast but error-prone, to be able to tell if the GPU DC has gone off the rails, in which case it is restarted from the most-recent interim savefile whose Res64 matches the Prime95 DC one.

axn 2017-12-27 10:00

[QUOTE=Madpoo;474957]What's the fastest GPU out there now for LL testing... Titan of some flavor?[/QUOTE]

Titan V. Probably will do this in under 18 hours. IIRC, TheJudger posted some benchies on it.

Prime95 2017-12-27 10:04

[QUOTE=Uncwilly;474956]Did the e-mail system work? [/quote]

Of course not. Chalk up yet another failure for this particular PHP code. I've asked Aaron and James to investigate.

What's weird is I copied that code for the recent PRP changes. I've successfully received several Mersenne-cofactor-is-PRP emails over the last few months.

[quote] Did the residue masking work? (And is there a clue somewhere?) [/quote]

The quick look around I did on the web site showed nothing of interest. Perhaps someone here can find a web page that doesn't mask the result as well as it should.

[quote] Can we assume that this is a WR or should we assume that this is a Double Check? Also, I would assume, based upon the times your are talking about, this is not in the 100M digit range.[/QUOTE]

You may assume anything you like :smile:

axn 2017-12-27 10:05

[QUOTE=TheJudger;473281]CUDA 9.0, CUDA driver 384.98, CUDALucas 2.05.1 (SVN rev. 99)

Benchmark FFT sizes './CUDALucas -cufftbench 2048 32768 20'
[CODE]Device Tesla V100-PCIE-16GB
Compatibility 7.0
clockRate (MHz) 1380
memClockRate (MHz) 877

fft max exp ms/iter
<snip>
4375 80897867 0.9595
4608 85111207 0.9649
5184 95507747 1.0124
[/CODE]

Oliver[/QUOTE]

Assuming 82M expo, Titan V should complete it in under 24 hrs.
EDIT:- The above benchmark is for a Tesla V100 which ought to be faster than Titan V.

Prime95 2017-12-27 10:08

@Ernst: Yes, the initial run was done with Prime95. Madpoo's unofficial DC is also with prime95.

An mlucas run and a CudaLucas run would be a good idea. Volunteers?

ET_ 2017-12-27 10:24

[QUOTE=Prime95;474947]Prime95 version 28.9 was used. Thus, no Jacobi checking during the run or at the end of the run.[/QUOTE]

Then I'm not the one who found it :sad:

As you know, I am maintaining the DoubleMersenne site. It would be nice to know the exponent in advance (once confirmed) to start the sieving (it should be tested up to at least 600T, or 210 CPU/days) and set up the updated pages.

Prime95 2017-12-27 10:29

[QUOTE=ET_;474975]Then I'm not the one who found it :sad:[/QUOTE]

So sorry, maybe next time.

I reached the potential discoverer by email. Of course, he/she is traveling and cannot send the last intermediate file to me. We have to wait for madpoo.

ATH 2017-12-27 11:11

[QUOTE=Prime95;474974]@Ernst: Yes, the initial run was done with Prime95. Madpoo's unofficial DC is also with prime95.

An mlucas run and a CudaLucas run would be a good idea. Volunteers?[/QUOTE]

I would like to start a Titan Black run and I could start a mlucas run on a C5 instance, I just need to compile mlucas on it.

The fastest GPU right now is the Tesla P100 and TheJudger might have one, as he has posted benchmarks.

TheJudger 2017-12-27 12:11

Fastest GPU today should be Tesla [B]V[/B]100 - TITAN V has "crippled" memory bandwidth and thus should be slower because CUDAlucas is memory bandwidth bound on this type of hardware.

Oliver


All times are UTC. The time now is 16:03.

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