mersenneforum.org  

Go Back   mersenneforum.org > Factoring Projects > FermatSearch

Reply
 
Thread Tools
Old 2019-01-16, 13:45   #34
feromant
 
"Roman"
Dec 2016
Everywhere

2×13 Posts
Default New version of program feromant_CUDA (v2.018) for Linux is released!

With a little delay, but with great joy I present the Linux build of
feromant_cuda (v2.018). The build is compiled under CUDA 10.0 and
supports GPUs generation Kepler (sm_3.0 and sm_3.5), Maxwell (sm_5.0 and
sm_5.2), Pascal (sm_6.1), Volta (sm_7.0) and Turing (sm_7.5). To run the
program, you will need a CUDA driver version 410 or higher.

You will find it in the download page.
feromant is offline   Reply With Quote
Old 2019-04-06, 18:00   #35
Gary
 
Gary's Avatar
 
"Gary"
Aug 2015
Texas

72 Posts
Default

Is there a version of feromant (CPU only) for Linux? Thanks!
Gary is offline   Reply With Quote
Old 2019-04-14, 08:57   #36
feromant
 
"Roman"
Dec 2016
Everywhere

110102 Posts
Default

Quote:
Originally Posted by Gary View Post
Is there a version of feromant (CPU only) for Linux? Thanks!
Gary, I apologize for not responding promptly. No version feromant under Linux yet. If I have free time, I will definitely build under Linux. The difficulty lies in the large amount of asm code that will have to be converted from wasm to gas.
feromant is offline   Reply With Quote
Old 2019-04-19, 20:23   #37
Gary
 
Gary's Avatar
 
"Gary"
Aug 2015
Texas

72 Posts
Default

Quote:
Originally Posted by feromant View Post
Gary, I apologize for not responding promptly. No version feromant under Linux yet. If I have free time, I will definitely build under Linux. The difficulty lies in the large amount of asm code that will have to be converted from wasm to gas.
Thanks Roman. I would like to try out Feromant on one of my servers at work, since (according to Fermatsearch) Feromant is the fastest CPU only program for many ranges of N and is multi-threaded. Unfortunately the system is Linux only and no GPU. If you find time to create a Linux version I would be happy to help with testing and benchmarking. The server is a 64 core Epyc system, which should make things interesting.
Gary is offline   Reply With Quote
Old 2019-04-23, 14:04   #38
mathwiz
 
Mar 2019

7×13 Posts
Default

For a given range of N and k to test, is feromant_cuda faster than mmff? Does feromant_cuda do some pre-sieving that mmff does not? Thanks!
mathwiz is online now   Reply With Quote
Old 2019-04-23, 17:16   #39
feromant
 
"Roman"
Dec 2016
Everywhere

328 Posts
Default

Quote:
Originally Posted by Gary View Post
Feromant is the fastest CPU only program for many ranges of N
Probably Yes.


Quote:
Originally Posted by Gary View Post
Feromant is multi-threaded.
No.
feromant is offline   Reply With Quote
Old 2019-04-23, 17:39   #40
feromant
 
"Roman"
Dec 2016
Everywhere

2×13 Posts
Default

Quote:
Originally Posted by mathwiz View Post
For a given range of N and k to test, is feromant_cuda faster than mmff? Does feromant_cuda do some pre-sieving that mmff does not? Thanks!
No, because mmff uses pre-sieving on the GPU, and feromant_cuda does pre-sieving on the CPU, and the surviving candidates are transferred to the GPU and the divisibility is checked on GPU. The scope of these programmes is important. If factorbit<252 (maximum mmff), better use of the mmff. In the range 252<factorbit<640 (approximately 200<N<600) feromant_cuda is out of competition.
feromant is offline   Reply With Quote
Old 2019-04-24, 00:00   #41
mathwiz
 
Mar 2019

7×13 Posts
Default

Quote:
Originally Posted by feromant View Post
No, because mmff uses pre-sieving on the GPU, and feromant_cuda does pre-sieving on the CPU, and the surviving candidates are transferred to the GPU and the divisibility is checked on GPU. The scope of these programmes is important. If factorbit<252 (maximum mmff), better use of the mmff. In the range 252<factorbit<640 (approximately 200<N<600) feromant_cuda is out of competition.
Thanks! In fact I just discovered the table and info at http://www.fermatsearch.org/productivity.html which confirms what you said.

I'm curious how up-to-date the numbers in that table are? (and of course they depend heavily on the CPU+GPU used). For example, is it still the case that around n=100, mmff and feromant_cuda are about equally fast?
mathwiz is online now   Reply With Quote
Old 2019-04-24, 09:06   #42
ET_
Banned
 
ET_'s Avatar
 
"Luigi"
Aug 2002
Team Italia

2×2,383 Posts
Default

Quote:
Originally Posted by mathwiz View Post
Thanks! In fact I just discovered the table and info at http://www.fermatsearch.org/productivity.html which confirms what you said.

I'm curious how up-to-date the numbers in that table are? (and of course they depend heavily on the CPU+GPU used). For example, is it still the case that around n=100, mmff and feromant_cuda are about equally fast?
Yes. Since mmff was delivered many years ago and never updated since, an Feromant_CUDA was modified to make it working faster and deeper at high ranges, the data should be correct. Maybe mmff could have a small boost from the data shown using Nvidia RTX 20- series graphics cards.
ET_ is offline   Reply With Quote
Old 2019-04-27, 14:40   #43
Gary
 
Gary's Avatar
 
"Gary"
Aug 2015
Texas

3116 Posts
Default

Quote:
Originally Posted by feromant View Post

Originally Posted by Gary
Feromant is multi-threaded.

No.
In the table at http://www.fermatsearch.org/download.php Feromant has a "y" in the M/T column, which I interpreted to mean Feromant is multi-threaded. Thanks for the correction Roman.

Luigi, should the table be changed to Feromant M/T = n? Also, does the M/T column just refer to the CPU portion of the program? I assume that all the GPU based programs use many GPU cores / threads. Maybe add a table footnote defining what M/T means. Thanks!
Gary is offline   Reply With Quote
Old 2019-04-28, 13:36   #44
ET_
Banned
 
ET_'s Avatar
 
"Luigi"
Aug 2002
Team Italia

2×2,383 Posts
Default

Quote:
Originally Posted by Gary View Post
In the table at http://www.fermatsearch.org/download.php Feromant has a "y" in the M/T column, which I interpreted to mean Feromant is multi-threaded. Thanks for the correction Roman.

Luigi, should the table be changed to Feromant M/T = n? Also, does the M/T column just refer to the CPU portion of the program? I assume that all the GPU based programs use many GPU cores / threads. Maybe add a table footnote defining what M/T means. Thanks!
Working on it, thanks. I obviously considered CUDA capable programs as m/t.
The following programs are CPU M/T:

GFNDsieve (siever)
ppsieve (siever)
SrSieve (siever)
LLR
PMFS
Prime95

Should I add it as a footnote?

Last fiddled with by ET_ on 2019-04-28 at 13:45
ET_ is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Fermat code performances ET_ FermatSearch 29 2018-03-15 19:06
Rho code Happy5214 YAFU 3 2015-11-01 21:54
Please help me with my code daxmick Programming 15 2014-02-14 11:57
Code Help Andrew Programming 12 2013-02-16 20:53
New Code JohnFullspeed Programming 20 2011-09-04 04:28

All times are UTC. The time now is 02:09.

Thu Oct 1 02:09:24 UTC 2020 up 20 days, 23:20, 1 user, load averages: 1.63, 1.49, 1.45

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.