Advanced search

Message boards : Number crunching : Noelia ^^

Author Message
Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38188 - Posted: 30 Sep 2014 | 11:03:30 UTC

Hi all.

Since I started folding for GPUGRID, there has been nothing but trouble with the Noelia WUs.

A quick search in the forum tells me I`m not the only one.

Nobody wants to play with her.

SO....

Can`t you make it possible to choose which WUs to run!?

The first "short" runs from her took around 4 hours.....
Lately it`s 7+ hours.

We are talking (2-3 hours on fastest cards) runs.

What is going on!?

Best regards :)

Betting Slip
Send message
Joined: 5 Jan 09
Posts: 670
Credit: 2,498,095,550
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 38189 - Posted: 30 Sep 2014 | 12:29:19 UTC - in response to Message 38188.
Last modified: 30 Sep 2014 | 12:29:45 UTC

I can't see because you have your computers hidden but do you use graphic cards that only have 1gig of RAM?

Noelia's WU's tend to use more than 1gig of video ram and will therefore run much slower on a card that only has 1gig and may cause problems on computers that are used for other tasks at the same time.

Solution only run GPUGRID when computer is not in use.

eXaPower
Send message
Joined: 25 Sep 13
Posts: 293
Credit: 1,897,601,978
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 38194 - Posted: 30 Sep 2014 | 13:50:03 UTC - in response to Message 38189.
Last modified: 30 Sep 2014 | 13:57:25 UTC

If system is an SLI, running two tasks, one on each card, the memory can only be used for one card.

For a 2048Mb GDDR5 card if total memory go's near 2000Mb, you'll notice for one card MCU will drop to under 15% and one GPU will be pinned at 99% GPU usage due to starved MCU (this why you saw a jump form 3hr to 7hr)

I had this happen a while back when Noelia tasks were about 1000~Mb each, my total time skyrocketed, until I released more memory. If you look for 2Gb cards SLI system in Host section, even for high end cards-- Runtimes are way higher when MCU is starved. (normal MCU GDDR5/128bit usage for my two cards for all different type of Noelia is 21-31%for short or Longs)

I shut off every aero feature to save GPU RAM. If I run two long Noelia's (depending on type I can get away with 2024Mb usage before I starve MCU.)

Trial and error will help out testing, Noelia work units are currently being researched the most with numerous task types.

Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38197 - Posted: 30 Sep 2014 | 15:25:08 UTC

I`m crunching on a gtx 780 TI 3gb.

"Noelia's WU's tend to use more than 1gig of video ram....."

They have NEVER exceeded 400-500mb vram according to GPU-Z.

No Sli setup and 1 WU at a time.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 2
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 38199 - Posted: 30 Sep 2014 | 17:19:25 UTC - in response to Message 38197.
Last modified: 30 Sep 2014 | 17:20:05 UTC

I`m crunching on a gtx 780 TI 3gb.

Is this a Gigabyte Windforce card?
You should make your computers visible to get more detailed help, or a link to the failed workunits would be helpful.

Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38201 - Posted: 30 Sep 2014 | 17:37:49 UTC

It`s a MSI card.

The WU is not failing, it´s aborted.

Don`t want to use 7+ hours on a run that should be completed in 2-3 hours :)

No point in choosing a short run then imo.

I hope it does`nt sound like I`m bitching :)

eXaPower
Send message
Joined: 25 Sep 13
Posts: 293
Credit: 1,897,601,978
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 38203 - Posted: 30 Sep 2014 | 17:47:31 UTC - in response to Message 38201.
Last modified: 30 Sep 2014 | 17:50:52 UTC

Do you have an Intel iGPU running any other project tasks? This will (5-20%)slow down any GPU crunching concurrently on GPUGRID. If you look at performance chart most GTX780ti are between 2-3hr total runtimes for short task--SH2 Noelia.

Matt
Avatar
Send message
Joined: 11 Jan 13
Posts: 216
Credit: 846,538,252
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 38211 - Posted: 1 Oct 2014 | 2:35:45 UTC - in response to Message 38203.

Thomas, I'm running two 780Ti cards and have never had a problem with Noelia WUs. Making your computer visible to us would really help us with a diagnosis for you. Her recent Long Task WUs take between 19k and 20k seconds (5.28 - 5.56 hours) for me running on W7 x64. I don't typically do Short Runs unless the Long Queue is out of work.

I see you're using GPU-Z to monitor your cards. That's a pretty good program but Nvidia Inspector can give you a lot more information on your cards, at least from my experience.

Nvidia Inspector

I'm currently running two NOELIA_5MG tasks (1 per card) in SLI (I do a lot of gaming and don't like to switch SLI on and off). My MCU (Memory Controller Unit) usage is at 29% per card, GPU usage is at 85% per card, and memory usage is around 1730MB per card (remember I'm in SLI, so WU memory is x2).

I very, very rarely have a failed GPUGrid WU and when I do it's usually due to CPU tasks I'm running from other BOINC projects.

Which drivers are you using currently? Really, we just need to see your computer setup and I think we may be able to help you out further.

Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38256 - Posted: 2 Oct 2014 | 13:09:54 UTC

First of all.

Thank you guys for your patience :)

I have made my setup visible, and I hope it will add some clarity to the issue.

Could a part of it be the EVGA Precision?! I run it to control the amount of heat allowed.

At 60° celsius/65% fan speed, GPU clock is 915 Mhz/Voltage 1049mn.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 2
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 38261 - Posted: 2 Oct 2014 | 17:51:43 UTC - in response to Message 38256.

Could a part of it be the EVGA Precision?! I run it to control the amount of heat allowed.

At 60° celsius/65% fan speed, GPU clock is 915 Mhz/Voltage 1049mn.

These settings make your card to not boost, which causes some slowdown, and probably the card throttles itself down a little more when it reaches 60°C. The another reason is that these NOELIA workunits are CPU intensive, so you should set the SWAN_SYNC environment variable (to 0) to make your workunits faster, but it's make no sense when you have such low temperature and fan speed limits.

Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38276 - Posted: 3 Oct 2014 | 10:12:49 UTC

So if I want better performance I should get some waterblocks for the card?

5-10% cpu utilization isn`t really cpu intensive imo.

How do I add the SWAN_SYNC? Folding noob alert :)

(0.985cpus + 1 Nidia gpu) I have chosen 3 cpus(cores)?

Thomas
Send message
Joined: 17 Aug 14
Posts: 16
Credit: 4,464,650
RAC: 0
Level
Ala
Scientific publications
watwatwat
Message 38410 - Posted: 11 Oct 2014 | 11:29:10 UTC

Hmm, I did a Google search, and apparently the boost curve for a GTX 780 ti kicks in at 88% utilization.

I have been able to push my gpu to just 80%, while folding, watching 1080p video and playing FM 2013 at the same time.

SO...either you have to utilize the gpus better, or I have to get some waterblocks and overclock my card, forcing it to boost.

Jacob Klein
Send message
Joined: 11 Oct 08
Posts: 1127
Credit: 1,901,927,545
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 38417 - Posted: 12 Oct 2014 | 5:31:18 UTC
Last modified: 12 Oct 2014 | 5:33:16 UTC

I would read through this thread, where I provide a software-based solution to force max boost at Windows startup, using NVIDIA Inspector.
http://www.gpugrid.net/forum_thread.php?id=3647
... specifically this post includes some detailed instructions:
http://www.gpugrid.net/forum_thread.php?id=3647&nowrap=true#36320

511513y
Send message
Joined: 15 Apr 14
Posts: 5
Credit: 6,433,630
RAC: 0
Level
Ser
Scientific publications
wat
Message 38550 - Posted: 16 Oct 2014 | 18:53:58 UTC
Last modified: 16 Oct 2014 | 18:55:43 UTC

Could part of the problem be the strict adherence to running under 60 C?
To me that seems quite cold.. I run on a laptop so I'm used to seeing GPU temps around 80 C.

Perhaps if you increased the allowed thermal limit a bit, your card would go faster without the need to overclock it or get a new cooling block.
If you've already hit a predetermined thermal limit, there's little that overclocking or forcing boost would do.

If you feel 60 C is the top mark with what you're comfortable with, that's your call. As for SLI, All I know is both GPUs usually keep copies of the same memory, so if each card has 2 GB of memory, you really only have 2 GB available (not 4 GB) Although I will yield to someone who has a system like that to experiment on

Post to thread

Message boards : Number crunching : Noelia ^^

//