Printable View
Yesterday, just for kicks, I changed the <count> setting on my 980X+7970 to <count>.083</count> to try and get 12 GPU WUs going.
So far, 24 hours later, it seems to be doing nicely. It went from 8 in ~3.5 minutes to 12 in ~4.5 minutes.
10/28/2012 0:009:16:46:05 978,709 4,128 (running 12)
10/27/2012 0:007:19:44:55 906,234 3,887 (running 8)
Some of that gain is from PVs catching up, but it seems to make a material difference.
Anyone else have a hex core trying this?
I'm also running 12 tasks on my 980X and it seemed to proved better than running less.
I'm running 12 threads feeding a 7870. I was running 8 before and GPU useage was around 87%, with 12 threads it went up to 95%. Only an 8% difference but on 500k PPD, that's an extra 40k PPD! It's been running a few days like this without a single error or invalid but the lag is pretty bad.
I have a 990x but it's paired with 2 5970s that I can't do the app_info thing on. I just turned HT off to minimize the amount of CPU crunching time each WU needs.Quote:
Originally Posted by fallwind
I wonder if you might get better results if you turn HT off and tell the card to run 6 threads. You may be able to get a higher overclock on the CPU as well.
Quote:
Originally Posted by =[PULSAR]=
It's good to have some reassurance that I'm not seeing things.Quote:
Originally Posted by fallwind
Maybe I'll try <count>.05</count> to see what happens. :hrhr:
I tried 10 (<count>.10</count) on a 2600K, but only 9 would run.
Little percentages on big numbers = big numbers. :DQuote:
Only an 8% difference but on 500k PPD, that's an extra 40k PPD! It's been running a few days like this without a single error or invalid but the lag is pretty bad.
Although I didn't notice any lag, this one is a dedicated cruncher, so lag doesn't matter.
It seems to me that running more concurrent WUs through is the way to go. The 980X + 7970 w/ 12 concurrent WUs is looking like it might start averaging 1MM PPD as the PVs start to fade.Quote:
I wonder if you might get better results if you turn HT off and tell the card to run 6 threads. You may be able to get a higher overclock on the CPU as well.
Hey guys, I have my 990x and 970 each running a 5970, I am using the app info file set at .133. It is taking about 12 minutes to finish 13 w/u's on each machine, I'm not sure why it is running 13 but does that seem like a reasonable amount of time? Would I be able to contribute considerably more with a pair of of newer 7970's?
I keep trying to tell myself "I'm not buying new hardware, I'm not buying new hardware" and then I read these posts and wonder why I'm not running a 7970 for 12 threads on my 980X !!!
I don't know the physical difference between a 5970 and a 7970, but my 980X @ 4 GHz w/ a cheap SSD and the 7970 running 12 at a time takes around 4 minutes per WU.Quote:
Originally Posted by sRHunt3r
Where did you get .133 for the <count> to get 12 running, I'm using .083?
I would say a pair of 7970s would crush whatever you're doing now. You'd probably want to use <count>.166</count> to keep 6 running on both. I don't know how you could get it to run more than 12 between the cards without all kinds of monkeying around with multiple BOINC instances. Maybe nanoprobe can chime in.
edit: I just reread your post and saw you have 2 separate PCs, not running 2 GPUs on one. Keep the <count> at .083 on both.
Or just get 2 7950s and run 6 threads each. Just imagine the points you'll get! You'll be eating lots of pie.Quote:
Originally Posted by Snow Crash
@ oneshot i think you can even delete your own posts now. so one would want to have 1 cpu thread per gpu thread?
It's 2 separate pc's but the 5970's are dual GPU so the way it is running now is 6 w/u's per GPU so 12 or 13 per machine, I will try changing to .166, thx! Oh yeah and I learned the hard way not to remote desktop into a machine running 12 gpu w/u's ...... Last night I logged into It and when I woke up this morning and checked on it and it had halted work last night and started back up when I logged back in this morning:shakes:Quote:
Originally Posted by pirogue
I had 9 gpu threads running on an 8 cpu thread machine.Quote:
Originally Posted by haschioz
@sRHunt3r - Don't know if you know or not but....I read on here somewhere that windows switches to a generic video driver when using remote desktop....something like that, so you found out.
Yup ... use VNC instead of RDPQuote:
Originally Posted by my87csx481
As my results of proved the CPU has little to no affect on GPU WU times.Quote:
Originally Posted by artemm
My cards in my Q9650 stock produce almost as much as my 2600K's OC'ed to 4.5GHz+
Memory bandwidth though... that affects it.
It seems 6-8 threads per GPU is a good amount for maximizing PPD. I'm going to test the scaling of 6 threads versus 8 threads so I'll find out more in the next few days.Quote:
Originally Posted by haschioz
Weird little quirk I'm having dual 7950's on my 980X...which rock by the way! Had it running great at 6 threads per GPU, but if I change the app_info for more than that it will give 12 threads to 1 GPU and 1 thread to the other.
http://i47.tinypic.com/2ch00b5.jpg
I noticed something similar trying to run 10 threads on my 2600K. By using <count>.10</count>, I got 9 running. I think BOINC includes a GPU in the CPU/thread count calculation.Quote:
Originally Posted by =[PULSAR]=
You *may* be able to get more running by trying <count>.125</count>. You could also try <count>.143</count>.
Just be careful, because if you get too carried away, you'll get instant computation errors. You may want to suspend everything that's not running and let everything finish before you stop BOINC and change the count. When you start BOINC back up, resume them a couple at a time to make sure you don't get computation errors on all of the downloaded WUs.
Can't seem to run anymore tasks at this time. GPU usage maxes out @ 92%, points are all over the place waiting to see the average the slower WU will average 100k a day while others would hit 160k...man that would be nice!
A factor in getting BOINC to run more GPU Wus than you have CPU cores/threads is that in conjunction with your <count> tag changes, you may also need to change the <avg_no_cpu> and <max_no_cpu> (the tags may not be exact but should get you to the right ones as they are just above the <count> in the <app_version> block). Basically if you want to run twice as many WUs as you have GPU cores and you have the <count> correct you will need to change the <cpu> tags to .5
<count> = (1) / (number of WUs you want to run)
<cpu> = (number of cpu threads you have) / (number of WUs you want to run)
^^ Yup, I used this to get 24 threads running on my 980x/7970 crossfire machine. So far so good and GPU useage went up about 10%! :up:
How many WCG PPD does that do?Quote:
Originally Posted by fallwind
A little over 1.4 million running 6 threads for each card, maybe with 12 threads each I can hit 1.6 mil.
Nice. I thought you *have* 12 per running.?? How many watts does that draw?Quote:
Originally Posted by fallwind
I have another machine running a 7870 with 12 threads. The 7970 crossfire rig was running 6 threads for each GPU up until yesterday when I changed it to 12 each, 24 total. So it's too early to tell exactly how many more points the change will yield.
I haven't had a chance to measure power draw. I'm in the middle of some major home renovations so I'm not even sure where the Kill-a watt is...lol!
Gotcha. Thanks.
I'm trying some new things today.
2600K + 7970 #1:
<count>.041</count>
<avg_ncpus>.33</avg_ncpus>
<max_ncpus>.33</max_ncpus>
2600K + 7970 #2:
<count>.0625</count>
<avg_ncpus>.5</avg_ncpus>
<max_ncpus>.5</max_ncpus>
980X + 7970:
<count>.041</count>
<avg_ncpus>.5</avg_ncpus>
<max_ncpus>.5</max_ncpus>
2600K #1 is averaging 24 every 7-8 minutes.
2600K #2 is averaging 16 every 5-6 minutes.
980X is averaging 24 every 6.5-7.5 minutes.
It looks like I've reached my max at around 1 per ~19-20 seconds or 3 per minute, which works out to ~998K WCG PPD each. (3 x 60 x 24 x 33 (BOINC points average over 125K WUs) * 7 (WCG conversion))
I'm hoping that the longer individual times will help with the PV backlog. Only time will tell.
With PVs getting validated, I should be able to sustain ~3MM WCG PPD until I go broke paying the electric bill. :D
Aside: I wish they reported the time in seconds. Trying to keep track of exact times is tough when you have to look at boinc manager at exactly the right instant.
Quote:
Originally Posted by pirogue
Suspend network for a while to get a bunch waiting to go then highlight each wu in turn and use "properties" to see the total and cpu time at your leisure.
Spent a little time playing with the 7950 and things like:
<avg_ncpus>0.8</avg_ncpus>
<max_ncpus>0.8</max_ncpus>
<coproc>
<type>ATI</type>
<count>.1</count>
At first glance The advantage is slim. I will probably need to let it run for a day to be sure.
Thanks. I do remember trying that.Quote:
Originally Posted by OldChap
The last time I tried suspending anything, I got distracted, forgot to set it back, and didn't realize it until the next morning.:hitself:
I'll just stick to eyeballin'. :yepp:
I hate when that happens. Done it a few times myself.
I need to get me a X79 MB + 3930k setup so I can run 2x7950 in it @ 8 WUs per card!!!