MULTI PULE GPU ON ONE BOARD HOW IS IT WORKING


log in

Advanced search

Message boards : Number crunching : MULTI PULE GPU ON ONE BOARD HOW IS IT WORKING

Author Message
Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1309720 - Posted: 24 Nov 2012, 11:30:11 UTC

I am running sabertooth fx990 rev 1 motherboard with fx8150 Zambezi 8 core cpu.
My current video is 1 hd6870 and 1 hd5770.
I am running catalyst 12.11 and amd streaming sdk.
I have 8 gig of ram and am running 2 1.5 Terra byte wd hard drives.
after much sweat and tears every thing is working good.
what are your experience's with 2 or more cards in 1 system,
and how about the apu's any one have comments on these with 2 or more additional gpu's?
Let's here about massive systems and what work's and what doesn't.


looking forward to hearing about what works DANCER42.

____________

John
Send message
Joined: 21 May 99
Posts: 49
Credit: 5,131,097
RAC: 6,213
United States
Message 1309786 - Posted: 24 Nov 2012, 15:05:11 UTC

Run Nvidia gtx 670 and Gtx 470 310.54 beta driver
Cool MAster 1200watt PS
ASUS P6t V2 motherboard
H100 cpu cooler
I7 920 cpu

Problems with the first driver where 600 series only. After a few driver updates can now run different cards on same machine.

edit XML for 2 wu on 470 and 4 wu on the 670

____________

MikeN
Send message
Joined: 24 Jan 11
Posts: 301
Credit: 32,337,886
RAC: 17,913
United Kingdom
Message 1309877 - Posted: 24 Nov 2012, 18:40:24 UTC - in response to Message 1309720.

I am running sabertooth fx990 rev 1 motherboard with fx8150 Zambezi 8 core cpu.
My current video is 1 hd6870 and 1 hd5770.
I am running catalyst 12.11 and amd streaming sdk.
I have 8 gig of ram and am running 2 1.5 Terra byte wd hard drives.
after much sweat and tears every thing is working good.
what are your experience's with 2 or more cards in 1 system,
and how about the apu's any one have comments on these with 2 or more additional gpu's?
Let's here about massive systems and what work's and what doesn't.


looking forward to hearing about what works DANCER42.

I gave up trying to run 2 different GPUs on the same motherboard. The problem I had was that SETI only calculates an average task duration factor for your whole system (CPU+GPUs) and the GPUs had such a speed difference that SETI kept timing units out as they had taken more than the allowed time over the expected run time. I think it is possible to avoid this problem if you run two instances of BOINC at a time with 1 GPU each, but that was too much like hard work for me.
____________

juan BFBProject donor
Volunteer tester
Avatar
Send message
Joined: 16 Mar 07
Posts: 5394
Credit: 305,667,114
RAC: 320,885
Brazil
Message 1309892 - Posted: 24 Nov 2012, 19:31:39 UTC
Last modified: 24 Nov 2012, 19:32:35 UTC

See this post at Vyper site:

http://vyper.kafit.se/wp/index.php/2011/02/04/running-different-nvidia-architectures-most-optimal-at-setihome/

It teaches whow to do run 2 different NVIDIA on the same hosts. Don´t know if works with ATI
____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1310068 - Posted: 25 Nov 2012, 7:28:54 UTC - in response to Message 1309877.

I am running sabertooth fx990 rev 1 motherboard with fx8150 Zambezi 8 core cpu.
My current video is 1 hd6870 and 1 hd5770.
I am running catalyst 12.11 and amd streaming sdk.
I have 8 gig of ram and am running 2 1.5 Terra byte wd hard drives.
after much sweat and tears every thing is working good.
what are your experience's with 2 or more cards in 1 system,
and how about the apu's any one have comments on these with 2 or more additional gpu's?
Let's here about massive systems and what work's and what doesn't.


looking forward to hearing about what works DANCER42.

I gave up trying to run 2 different GPUs on the same motherboard. The problem I had was that SETI only calculates an average task duration factor for your whole system (CPU+GPUs) and the GPUs had such a speed difference that SETI kept timing units out as they had taken more than the allowed time over the expected run time. I think it is possible to avoid this problem if you run two instances of BOINC at a time with 1 GPU each, but that was too much like hard work for me.




--------------------------------------------------------------------------------
when i first started i ran hd6870,hd5770,hd5570 and was cranking about 110,000 seti work units a day mostly primegririd have had no timeouts while computing.
the hd6870 would crank about 23 jobs a day worth about 3300 work units the hd 5770 about 11 jobs and the hd5570 about 5 jobs, 2 cpu cores were ocupied feeding the gpu's 5 cpu core's did the remainder of the work.
i set gpu and cpu max usage at 85 % so i could do other things still get a little stutter when throttling accrues. have not neded to run more than 1 boinc manager yet give it a try now may work fine.
misplaced hd 5570 while tweaking so am only running 2 now.
was running boinc 7.0.31 runng current client have only been running multiple gpu's for about 1 month but did 3,000,000 work unit's in that time.
come in the water is fine lol.


DANCER42


____________

Profile Ageless
Avatar
Send message
Joined: 9 Jun 99
Posts: 12324
Credit: 2,629,532
RAC: 1,085
Netherlands
Message 1310092 - Posted: 25 Nov 2012, 9:12:11 UTC - in response to Message 1310068.

...only been running multiple gpu's for about 1 month but did 3,000,000 work unit's in that time.

Yeah, why don't I believe that. That means that you run one task for 1.16 seconds each second and I can tell you that your 2 HD68x0s aren't that fast.

Now, before you yell, "no, look, here", that's credits that you got over at Primegrid. Credits is not equal to work units. You get multiple credits for each task. 3,371.00 credits for each Primegrid task. With the credit being 3371, it's only 889 tasks that you ran to get to your 3 million credits. A lot less than your continued claims of fame.

See Computation_Credit for more information.

And oh, you cannot compare Primegrid's method of giving out credit, until they use the same version of CreditNew that Seti is using. As long as they give out fixed credit, it's incomparable.
____________
Jord

Fighting for the correct use of the apostrophe, together with Weird Al Yankovic

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1310112 - Posted: 25 Nov 2012, 10:13:57 UTC - in response to Message 1310092.

...only been running multiple gpu's for about 1 month but did 3,000,000 work unit's in that time.

Yeah, why don't I believe that. That means that you run one task for 1.16 seconds each second and I can tell you that your 2 HD68x0s aren't that fast.

Now, before you yell, "no, look, here", that's credits that you got over at Primegrid. Credits is not equal to work units. You get multiple credits for each task. 3,371.00 credits for each Primegrid task. With the credit being 3371, it's only 889 tasks that you ran to get to your 3 million credits. A lot less than your continued claims of fame.

See Computation_Credit for more information.

And oh, you cannot compare Primegrid's method of giving out credit, until they use the same version of CreditNew that Seti is using. As long as they give out fixed credit, it's incomparable.




--------------------------------------------------------------------------------I concede all that,between tweaking, and the server crashes i have yet to get good numbers about seti, will post when I can.
why I started this thread was that while several people are running rigs that are cranking out huge numbers there hasn't been any one place i could go to find out howto.
So I hoped to share my experiences and attract others as well so it would be easier for the next guy.
Right now after mostly finishing my build on an 8 core monster I am questioning whether I would have been better off building around a 4 core apu to get more bang for the buck.
1 thing I have learned is that even the lowly under rated hd 5570 which can be found for under $25 used can at least come close to equaling my $180 fx8150.
I hope this thread will be about finding answers together so we can all crunch better, faster,and cheaper by learning from each others mistakes rather than duplicating them.

____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2860
Credit: 215,135,125
RAC: 177,129
United States
Message 1310278 - Posted: 25 Nov 2012, 20:15:13 UTC - in response to Message 1310112.



1 thing I have learned is that even the lowly under rated hd 5570 which can be found for under $25 used can at least come close to equaling my $180 fx8150.



I am surprised it is only equaling it.

Several things you should know (I'm trying to be helpful):

Gpus outperform Cpus in the projects that can use the Gpus.

Some video cards, on some projects, require a little more "help" from the CPU. On those projects it is usually best to leave one or more cores idle on the CPU.

I use the "resource monitor" in Windows to help determine when I am over-taxing the system resources and then limit the CPU (and therefore memory) use. What I am most interested-in is feeding the video cards efficiently. I don't really care if it "looks like" I'm wasting 20% of the CPU's ability to crunch, or 30% for that matter. I never, ever, want to see the CPU's usage go to 100%, or see it run out of physical memory. That just means I'm slowing something down and if that "something" is a video card, I'm doing more harm than good. The video cards are at least 10 times better at crunching than the CPUs, often 20 times.

Recently I found that on one project I need to leave at least half of the CPU's resources idle (depending on the CPU) in order not to throttle the video card's crunching.

Intel CPUs are vastly superior to AMD CPUs at crunching the numbers we crunch here; so-much-so that putting an AMD to crunching-work is almost an exercise in futility. Having said that, you'll find I am a fan of AMD CPUs because they are relatively cheap and can feed a GPU just fine.

I have an 8120. I find that the "shared FPU" on the FX CPU makes it a less-good cruncher than a Phenom II and not much better than an Athlon II. The FX processors have superior overclocking potential, but I, personally, don't care for the extra power consumption and heat produced by overclocking the FX processors. I have settled on crunching my FX on three cores (allowing it to go into "boost" on its own) and letting the other five cores feed video cards, although that machine currently only has one card in it.

After experimenting with my FX processor, I went back to buying less expensive Phenom II processors and even Athlon IIs, and using the money saved to get more and better video cards.

I haven't seen any tangible benefit to "overclocking" the RAM on any of my AMDs, with the best overclocking performance coming from *small* increases to the front side bus.

Yes, you can buy three or four inexpensive video cards and equal the number-crunching performance of one moderate video card. In my experience, it isn't worth it. There is a "sweet spot" of price/performance somewhere in the mid-range cards, usually in the x60 numbers with NVIDIA. I have not found that with SETI@Home that the AMD/ATI cards are as productive for the money despite the fact that they do some things better.

You mentioned "huge numbers." There are people here who have nearly $4,000 worth of video cards in their computers. That's how they are doing it. Most of us can't, or won't, spend the money to do that.

When you put four video cards in a computer you are asking for heat problems, so you have to be careful about how your case will "breathe" with that many cards in it and it *will* heat your room, causing a sort of heat feedback loop, so you have to be able to ventilate the room, also.

My limited experience, not to be confused with mega-crunchers, is that without spending gigantic dollars it is possible to cheaply adequately cool three video cards in one case, if you have the right case. If you have the wrong case, it is practically impossible to keep two cool.

Power supplies to run four big cards in a single case are extremely expensive. *I* find that *for me* I'd rather buy less expensive power supplies and use the savings to buy better GPUs.

So, to sum-up: It's all about the GPUs, not the CPUs. It's better to crunch on NVIDIA GPUs and Intel CPUs. It's better to crunch on one GPU than four if the output is ultimately the same. Once you get into the under $100 GPUs, you are usually better-off buying fewer $150-200 GPUs. You are usually better off if you will crunch two work units (or sometimes three) at a time on your GPUs. You are always much, much, much better-off using "optimized applications" from the SETI Lunatic's website when they are available. You need to be sure not to run your system out of resources (RAM or CPU cycles) by trying to do too much. The least expensive way to crunch is to purchase "last year's models" of GPUs rather than buy the newest/hottest/bestest thing available right now. (for instance, a 560Ti can be had for under $200 and will at least equal a 660Ti that costs $300).

Yes, there are "fans" of various brands and cards and cases and power supplies. You can get into an argument about anything if you want-to. I don't think I'm the be-all and end-all of crunching knowledge or anything like that, but I am trying to share with you, all at once, what I've learned from my own experiences (and errors) and those of other people who have posted their experiences here over the years.

SURE the AMD CPUs could be said to be a good deal in bang/buck if you overclock the heck out of them and keep them well fed and cool and optimize every setting under the sun. That doesn't change the fact that the programs we run here run better on an Intel.

YES, the ATI cards are wonderful at double-precision OpenCL applications that call for _____. That doesn't change the fact that the NVIDIA cards running CUDA applications are what you want to use if you are buying the cards to crunch numbers here.

YES, you can buy three $90 video cards and get them all to crunch more than one $270 card --- under specific circumstances, maybe. Or you might have an output of 15,000 "credits" whereas one $200 card is much cooler and much easier to power and would do 20,000 "credits."

Anything is arguable. (see also: lawyers)
____________

juan BFBProject donor
Volunteer tester
Avatar
Send message
Joined: 16 Mar 07
Posts: 5394
Credit: 305,667,114
RAC: 320,885
Brazil
Message 1310291 - Posted: 25 Nov 2012, 21:18:59 UTC - in response to Message 1310278.
Last modified: 25 Nov 2012, 22:17:07 UTC


YES, the ATI cards are wonderful at double-precision OpenCL applications that call for _____. That doesn't change the fact that the NVIDIA cards running CUDA applications are what you want to use if you are buying the cards to crunch numbers here.

An excelent explanation...

You just forget to mention...

- SETI only use single precision to do it´s wonderful work... so who needs double-precision capacity for crunching???

- Even for experienced heavy-crunches (remember what happens with our Burnig Bird friend), manage 3 or 4 cards per MB is something almost catastrophic most of the times, to much heat, to much power hungry, to much what else... a time bomb ready to explode...and it´will explode, sooner or later... you could bet all your chips on that!

Still without power, thanks my netbook have a large battery...
____________

Profile MikeProject donor
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 24498
Credit: 33,829,681
RAC: 24,175
Germany
Message 1310304 - Posted: 25 Nov 2012, 22:05:35 UTC

YES, the ATI cards are wonderful at double-precision OpenCL applications that call for _____. That doesn't change the fact that the NVIDIA cards running CUDA applications are what you want to use if you are buying the cards to crunch numbers here.


Wrong.

With new apps HD 7 cards are equal to nvidia cards.
A HD 7970 can outperform a 680 for example.
Of course it uses more power.

____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2860
Credit: 215,135,125
RAC: 177,129
United States
Message 1310310 - Posted: 25 Nov 2012, 22:19:37 UTC - in response to Message 1310304.
Last modified: 25 Nov 2012, 22:46:50 UTC

YES, the ATI cards are wonderful at double-precision OpenCL applications that call for _____. That doesn't change the fact that the NVIDIA cards running CUDA applications are what you want to use if you are buying the cards to crunch numbers here.


Wrong.

With new apps HD 7 cards are equal to nvidia cards.
A HD 7970 can outperform a 680 for example.
Of course it uses more power.


What new apps? The new standard apps? Running what? Astropulse or MB or vlars or what?

I'm very interested. Please show me something / anything to illustrate the equality or superiority to NVIDIA crunching SETI work.

You know I'm cheap, so if there's value in the new generation of AMD/ATI cards, I think I need to know about it.

EDIT: Here's what I was able to find. I found several so that there wasn't a one-off problem with the comparisons.

Here’s a 7970 vs a 550Ti: http://setiathome.berkeley.eduworkunit.php?wuid=1119805086

…and a 7970 vs a 430: http://setiathome.berkeley.eduworkunit.php?wuid=1119617589

vs a 480: http://setiathome.berkeley.eduworkunit.php?wuid=1119617589

vs a 590: http://setiathome.berkeley.edu/workunit.php?wuid=1119592106

vs a 560: http://setiathome.berkeley.edushow_host_detail.php?hostid=6831358

vs a 560Ti: http://setiathome.berkeley.eduworkunit.php?wuid=1119546643

vs a GTS 450: http://setiathome.berkeley.eduworkunit.php?wuid=1119530039

vs a GTX 460: http://setiathome.berkeley.eduworkunit.php?wuid=1119525627

vs a different 560Ti: http://setiathome.berkeley.eduworkunit.php?wuid=1119119424

vs a 570: http://setiathome.berkeley.edu/workunit.php?wuid=1119020786

vs a different 460: http://setiathome.berkeley.eduworkunit.php?wuid=1118956594

I'm not seeing GTX 680 competition here, but then, I couldn't care less about frame-rates in Batman.

Every time I say this you say "Wrong!" and I never see any indication that I'm wrong at all. I've been wrong before. I know how to be wrong. I know how to appreciate someone's enlightening me and then thank them for curing my ignorance.

I just am not able to find evidence of my ignorance.

Are the ATI cards doing six at once or something? I don't know how to tell that.
____________

Profile MikeProject donor
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 24498
Credit: 33,829,681
RAC: 24,175
Germany
Message 1310311 - Posted: 25 Nov 2012, 22:37:03 UTC

I think you`ll have to wait til it gets released.

____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2860
Credit: 215,135,125
RAC: 177,129
United States
Message 1310318 - Posted: 25 Nov 2012, 22:51:34 UTC - in response to Message 1310311.

I think you`ll have to wait til it gets released.


Okay, so what I was looking at wasn't a 7970.

...it was a 79-something with the same clocks, maybe different SMUs?

Anyway, okay, so an ATI card that hasn't been released yet is the equal of a NVIDIA card that's been out for a year.

I have no problem believing that.

____________

Profile MikeProject donor
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 24498
Credit: 33,829,681
RAC: 24,175
Germany
Message 1310323 - Posted: 25 Nov 2012, 23:00:35 UTC

LOL

The card is available but the apps not yet.

____________

juan BFBProject donor
Volunteer tester
Avatar
Send message
Joined: 16 Mar 07
Posts: 5394
Credit: 305,667,114
RAC: 320,885
Brazil
Message 1310325 - Posted: 25 Nov 2012, 23:12:23 UTC
Last modified: 25 Nov 2012, 23:17:08 UTC

Don´t want to enter in any discussion but, thats makes me remember the 1GB at SETI lab, is avaiable (and even paid every month) but can´t use because don´t have the optical fiber cable needed allready wired to the lab...
____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1310343 - Posted: 26 Nov 2012, 0:19:35 UTC

thank you for the lively discussion!!!
You may well be right that running three or more cards on 1 mb is problematic I did notice some instability while running an hd6870, a hd5770, and a hd5570.
the 6870 and 5770 co-reside well so far.
the 5770 runs twice as fast as the 5570, and the 6870 run's about twice as fast as the 5770.
a point to be made is that every one has there own set of circumstances and how much or little will depend on that.
The best bang is a moving target, and for some one just starting an hd5570 with 400 streaming cores beat's the heck out of no gpu at all. I had 2 in my junk box,and it ran on the same catalyst 12.11 driver the other cards run on just had to install the "AMD streaming sdk" to get opencl back.
As to cooling my case is open with a fan blowing in the cpu has a XIGMATEK Aegir SD128264 Mega Killer Double HDT cooling it and while massive is barely doing the job. Not only the cost of the card but of the power bill for the top card's needs to be taken in to account.


____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2860
Credit: 215,135,125
RAC: 177,129
United States
Message 1310347 - Posted: 26 Nov 2012, 0:26:11 UTC - in response to Message 1310323.

LOL

The card is available but the apps not yet.



I'm sorry, I'm so lost...

So a card that has been released is as good as another card that has been released at running applications that haven't been.

And we know this because someone's beta-testing them somewhere?

____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1310406 - Posted: 26 Nov 2012, 5:57:53 UTC - in response to Message 1310347.

Welcome to the bleeding edge.
____________

Profile dancer42
Volunteer tester
Send message
Joined: 2 Jun 02
Posts: 436
Credit: 1,159,602
RAC: 186
United States
Message 1310868 - Posted: 27 Nov 2012, 23:11:22 UTC

on advice from posters will leave first computer with only 2 gpu's have found second hd6870 so running 2.
am now building athalon 64 to run hd5770 out of junk box.
I have 2 nas servers as well to getup and running, my hope is to get enough network up and running to learn about then and be a test bed for idea's.
so if anyone has interesting things in a junk box I take donations.
And will put to use what i can so if you have wondered how or if post idea's
and i will try to find answers.
____________

Message boards : Number crunching : MULTI PULE GPU ON ONE BOARD HOW IS IT WORKING

Copyright © 2014 University of California