mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > GPU Computing

Reply
 
Thread Tools
Old 2013-03-14, 02:57   #221
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

863510 Posts
Default

Quote:
Originally Posted by ewmayer View Post
Based on all the tests done to date, what is the error rate?
Unknown. Some people do "in house" double checks before reporting CL results, and "spidering" won't help. I have about 10% error rate for the tests I do. I never counted exactly, but I still remember Feb-March 2012, when CuLu changed to non-power-of-2 FFT and the intermediary versions used to give wrong residues, after that date I always run two copies of CuLu in two different cards, and maybe with two different FFT sizes, and report the result only if I have a match. This is still faster than using P95 in the loop, and anyhow it will need a third party to do the DC. So, the results reported to the server are not relevant. Also, if someone only do DCLL, they won't report mismatching results, and they will TC first.

For me, running in parallel allows me to catch the error in the earliest phase, and save the time which would be wasted if the test is allowed to finish. When I have two non matching residues, I re-do both from the last saving.

From my experience, with overclocking (water cooled) the errors are around 10%, maybe more. Without overclocking the value is much lower, maybe 2% or so, but I am still not confident enough to giveup "in-house double checking".

The 10% can look a lot, but you have to think about the fact that if you test 10 exponents of 50M, then you do 500M iterations, and if only one iteration in those 500M is wrong, you got one of your 10 expos wrong, so here is your 10%.

There is no mystery into it. For me the real mystery is how P95 (and the CPUs in general) can do 500 MILLIONS iterations, without getting any of them wrong....
LaurV is offline   Reply With Quote
Old 2013-03-14, 08:15   #222
ET_
Banned
 
ET_'s Avatar
 
"Luigi"
Aug 2002
Team Italia

11·433 Posts
Default

I am running DC and LL on my cards. Since I started, I only had 3 non-matching results, two of them were subsequently found correct.

Luigi
ET_ is offline   Reply With Quote
Old 2013-03-14, 10:24   #223
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

B4B16 Posts
Default

Quote:
Originally Posted by nucleon View Post
Another Titan double check match.

Processing result: M( 29325073 )C, 0x23c2d7bc0d08e8d4, n = 1835008, CUDALucas v2.03
LL test successfully completes double-check of M29325073
CPU credit is 29.1622 GHz-days.

To me looks like underclocking the RAM did the trick.

The hard part now is to work out the most efficient use of this card.

-- Craig
What options do you have for RAM clocking? Max is 6.008 Ghz? (According to the review articles):
Quote:
Memory Clock 6.008GHz GDDR5
2500 Mhz now seems stable, have you tested the stability of anything between those?

Last fiddled with by ATH on 2013-03-14 at 10:25
ATH is offline   Reply With Quote
Old 2013-03-14, 11:43   #224
nucleon
 
nucleon's Avatar
 
Mar 2003
Melbourne

10038 Posts
Default

Quote:
Originally Posted by ATH View Post
What options do you have for RAM clocking? Max is 6.008 Ghz? (According to the review articles):


2500 Mhz now seems stable, have you tested the stability of anything between those?
It seems I can adjust in roughly 1MHz increments. The 2500 figure is based on a starting figure of 3000MHz. I guess the app I used measures the clock waveform, and the 6000MHz figure is the sexed-up DDR figure.

The reduction in performance was about 10+% ish. Given it takes a long time to verify stability, as LL double checks seem to be the only reliable way to verify. I'm too lazy to fiddle to optimize ram speed. (It's going to take a while :) )

-- Craig
nucleon is offline   Reply With Quote
Old 2013-03-14, 16:05   #225
Redarm
 
Redarm's Avatar
 
Apr 2012
Berlin Germany

3·17 Posts
Default

2900 mhz produces errors (3.1ms per I M48)
2750mhz seems stable for now (3,35ms per I M48)
Redarm is offline   Reply With Quote
Old 2013-03-14, 16:42   #226
frmky
 
frmky's Avatar
 
Jul 2003
So Cal

23×11×23 Posts
Default

Quote:
Originally Posted by ewmayer View Post
Based on all the tests done to date, what is the error rate?
I'm using older cards that are not over clocked, but in what must now be well over 100 tests, I've had one known bad result. That's under a 1% error rate.
frmky is offline   Reply With Quote
Old 2013-03-14, 16:47   #227
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

72×59 Posts
Default

Quote:
Originally Posted by Redarm View Post
2900 mhz produces errors (3.1ms per I M48)
2750mhz seems stable for now (3,35ms per I M48)
Ok, so 2500mhz is pretty close to the unstable crossover. Wierd that these benchmark programs showed no errors.

Quote:
Originally Posted by nucleon View Post
Doing further testing...

Furmark - no issues
MemtestG80 - no issues
occt - no issues
ATH is offline   Reply With Quote
Old 2013-03-14, 17:52   #228
Redarm
 
Redarm's Avatar
 
Apr 2012
Berlin Germany

3×17 Posts
Default

found a way to find errors produced by vram
instructions coming soon ... (memtestg80 is essential)
Redarm is offline   Reply With Quote
Old 2013-03-14, 19:32   #229
Brain
 
Brain's Avatar
 
Dec 2009
Peine, Germany

33110 Posts
Default 100M

Quote:
Originally Posted by Redarm View Post
2000h
2500MHz VRAM = 2070h:
Code:
E:\Eigene Dateien\Computing\CUDALucas\2.03\D0>CUDALucas-2.03-5.0-sm_35-x64.exe -threads 512 -f 20971520 -t 332192831
Warning: No ini file detected. Using defaults for non-specified options.
Starting M332192831 fft length = 20971520
Iteration 10000 M( 332192831 )C, 0x7e591d0cbd938d73, n = 20971520, CUDALucas v2.03 err = 0.0332 (3:44 real, 22.4112 ms/iter, ETA 2067:55:55)
Iteration 20000 M( 332192831 )C, 0xc2eda19e28f00bf4, n = 20971520, CUDALucas v2.03 err = 0.0339 (3:44 real, 22.4587 ms/iter, ETA 2072:14:52)
Iteration 30000 M( 332192831 )C, 0x4e96b2da36fcce56, n = 20971520, CUDALucas v2.03 err = 0.0352 (3:44 real, 22.4425 ms/iter, ETA 2070:41:49)
No guarantee for the residues... ;-)
Brain is offline   Reply With Quote
Old 2013-03-14, 19:38   #230
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

32×1,019 Posts
Default

Quote:
Originally Posted by Brain View Post
2500MHz VRAM = 2070h
Less than three months...

Nope. Still doesn't Make Sense(tm)....
chalsall is offline   Reply With Quote
Old 2013-03-14, 21:02   #231
kracker
ἀβουλία
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

1000011011112 Posts
Default

Quote:
Originally Posted by chalsall View Post
Less than three months...

Nope. Still doesn't Make Sense(tm)....
In one way it still makes sense to do them on CPU... I mean, 22 ms for a 30M exp.?? On a $1000 card?
kracker is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Titan's Best Choice Brain GPU Computing 30 2019-10-19 19:19
Titan Black ATH Hardware 15 2017-05-27 22:38
Is any GTX 750 the GeForce GTX 750 Ti owner here? pepi37 Hardware 12 2016-07-17 22:35
Nvidia announces Titan X ixfd64 GPU Computing 20 2015-04-28 00:27
2x AMD 7990 or 2x Nvidia Titan ?? Manpowre GPU Computing 27 2013-05-12 10:00

All times are UTC. The time now is 18:57.

Wed Aug 12 18:57:26 UTC 2020 up 26 days, 14:44, 0 users, load averages: 1.76, 2.06, 2.20

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.