mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > PrimeNet

Reply
 
Thread Tools
Old 2014-04-09, 16:52   #1
tapion64
 
Apr 2014

8510 Posts
Default Theoretical Evaluation of a GPU

In terms of trial factoring? In terms of LL testing? I guess what I would like to know is how to estimate a card's GHz-days/day from its specs.

From what I've gathered, CUDALucas relies on double precision calculations, so is it simply the card with the greatest number of cores and highest TFLOPS for double? Would mfaktc also benefit from high double precision, or is single precision better? I guess I'd like to see benchmarks for cards like the Tesla K series vs. consumer cards like GTX. I have a GTX 560 Ti which is more than capable for gaming, I don't really need or want a higher GTX card at the moment. If I'm interested in upping GPU compute abilities, is going for a Tesla card like K20 or K40 worth it in terms of GHz-days/day/dollar compared to say the GTX Titan?
tapion64 is offline   Reply With Quote
Old 2014-04-09, 17:32   #2
kracker
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

1000011110012 Posts
Default

http://www.mersenne.ca/mfaktc.php
kracker is offline   Reply With Quote
Old 2014-04-09, 17:50   #3
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

944710 Posts
Default

yes, and this too: http://www.mersenne.ca/cudalucas.php

and to answer to your question, no, it doesn't worth. Teslas are Titans, same chips, sorted for reliability, consumption, underclocked a bit and having ECC RAM, for exact calculus, more stability, longer life (well, not really, but you got the picture). If you are not in nuclear reactors business, go for Titans. They are few percents faster and 4 times cheaper. You will get DC/LL mismatches sometimes. Live with it :D

[edit: if you want to hit the jackpot at TF, then go for HD cards. They are also good for mining, but they are worse at LL]

Last fiddled with by LaurV on 2014-04-09 at 18:00
LaurV is offline   Reply With Quote
Old 2014-04-09, 18:15   #4
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11×157 Posts
Default

The need for double precision is just in LL. TF only uses single precision so there is no advantage to having DP for your TF tests.

For this reason, a GPU will always be better at TF than LL because double precision takes a massive performance hit. The Titans have 1/8 DP performance compared to the usual 1/24 for the other cards. So really while a Titan is three times better in LL than any other card, right off the bat, it's really three times less bad.

So for LL, you'll very likely be leaning the way of a Titan but even that one will perform better in TF.
TheMawn is offline   Reply With Quote
Old 2014-04-09, 18:32   #5
kracker
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

41718 Posts
Default

Quote:
Originally Posted by TheMawn View Post
For this reason, a GPU will always be better at TF than LL because double precision takes a massive performance hit. The Titans have 1/8 DP performance...
*1/3 DP.

@LaurV: Frankly though, I think titans are not worth it and are a waste of money compared to Haswells. Assuming a 2K FFT...

The GTX Titan costs 1000$, 250W TDP, 1.8 ms/iter.
A i5 4670k costs ~$240, 84W TDP, 11.5 ms/iter x4
kracker is offline   Reply With Quote
Old 2014-04-09, 19:41   #6
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

172710 Posts
Default

Quote:
Originally Posted by kracker View Post
@LaurV: Frankly though, I think titans are not worth it and are a waste of money compared to Haswells. Assuming a 2K FFT...

The GTX Titan costs 1000$, 250W TDP, 1.8 ms/iter.
A i5 4670k costs ~$240, 84W TDP, 11.5 ms/iter x4
Plus motherboard (x4) plus RAM (x4) plus PSU (x4) plus heatsink (x4) plus USB stick (or something to boot, x4)...

Whereas you can supplement your existing Haswell system with multiple Titans. All that takes is a 100% bigger expense for a PSU and maybe a step up in RAM capacity, and a better motherboard depending on what you have.

I did the math at some point, and I think it was worth $100 more for the Motherboard, $100 more for the PSU, and $1000 worth of GTX 670 even.
TheMawn is offline   Reply With Quote
Old 2014-04-09, 19:53   #7
tapion64
 
Apr 2014

5516 Posts
Default

So I guess the major downside to the HD cards is LL is out. Best of both worlds you'd just buy an R9 295 X2 and a Titan Black and call it a day? I am curious to see how much better the Maxwell cards are going to be though, especially with CUDA 6.
tapion64 is offline   Reply With Quote
Old 2014-04-09, 20:43   #8
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

22·1,873 Posts
Default

Quote:
Originally Posted by tapion64 View Post
From what I've gathered, CUDALucas relies on double precision calculations, so is it simply the card with the greatest number of cores and highest TFLOPS for double?
CUDALucas also has memory bandwidth (on the GPU) limitations, mfaktc does not. This means you can't simple count cores and DP throughput to pick the best card.
Prime95 is offline   Reply With Quote
Old 2014-04-09, 20:44   #9
kracker
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

32×241 Posts
Default

Quote:
Originally Posted by TheMawn View Post
Plus motherboard (x4) plus RAM (x4) plus PSU (x4) plus heatsink (x4) plus USB stick (or something to boot, x4)...

Whereas you can supplement your existing Haswell system with multiple Titans. All that takes is a 100% bigger expense for a PSU and maybe a step up in RAM capacity, and a better motherboard depending on what you have.

I did the math at some point, and I think it was worth $100 $80 more for the Motherboard, $100 $50 more for the PSU, and $1000 worth of GTX 670 even. ??
No, it really doesn't have to. I'm just talking about pure LL output. I'll calculate more later.

Also, buying a 6xx series is seriously a bad choice for computing.
kracker is offline   Reply With Quote
Old 2014-04-09, 21:19   #10
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11×157 Posts
Default

My point was that according to the expected throughput, it was more cost effective to buy one system with a motherboard capable of housing three GTX 670's (I did say GTX 670 even because I know it's a bad choice) than to buy multiple systems without GPUs.
TheMawn is offline   Reply With Quote
Old 2014-04-09, 22:21   #11
kracker
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

216910 Posts
Default

Quote:
Originally Posted by TheMawn View Post
My point was that according to the expected throughput, it was more cost effective to buy one system with a motherboard capable of housing three GTX 670's (I did say GTX 670 even because I know it's a bad choice) than to buy multiple systems without GPUs.
Hmm, perhaps. But I'm not going to think about the electric bill
But still, it is possible to build a Haswell system around $550.
kracker is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Theoretical Experiment Design c10ck3r Homework Help 7 2015-02-03 08:54
maximum theoretical speed of LL test w/o bandwidth limitations? ixfd64 Hardware 30 2012-03-05 06:16
Theoretical Model aymanmikhail Math 16 2011-09-06 01:50
Maximum theoretical MPG Mini-Geek Lounge 9 2008-07-14 22:45
Theoretical dream farm Lumly Hardware 25 2003-11-03 01:12

All times are UTC. The time now is 11:02.

Fri May 14 11:02:42 UTC 2021 up 36 days, 5:43, 0 users, load averages: 1.36, 1.47, 1.56

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, 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.