Getting the most from GPU at WCG...

[veebee]

I have finally, after a lot of fiddling (note to others, BOINC version 7.0.40 does not allow multiple HCC GPU WU's...they all crash immediately.. using 7.0.31 and all's well) managed to get both my i7 machines crunching 4 HCC WU's concurrently on the ir respective GPU's.

I know for some, like Kashi, this may be a given, but I wanted to work it our by myself...and finally did..  :hbang:.

What I would like to know NOW, if anyone (like Kashi ..heh heh ..)has "experiments", what settings get the most from your cards.

I am running a 6950 in one machine and a 5850 in the other.

Should clock speed be increased ? memory frequency ? can/ should voltages be reduced etc ?? ??

All seems to be running well, I have wound both cards up a bit - the 6950 is doing them in around 17 minutes, the 5850 about a minute longer..not a great difference !

whilst temps are under control, it IS taking a bit of fan noise for that...

So, Im wondering what sort of setting others have finished up with to get the most from their cards.

Also, in the machine with the 6950, I have a 660Ti, crunching GPU grid. I am happy with the science there, but wouldnt mind attacking HCC with it as well - I have heard the nVidia cards arent as good as ATI at HCC, but am wondering how to go about running WCG on both cards, with Multiple WU's on both cards... (would simply be added lines to the app_info.xml file I assume ??)

Thanks in advance !

[LawryB]

G'day Veebs.

On my Q6600 using BOINC 7.0.42 (x64) on Windows 7.
Running an ATI 7850  Core Clock  870 Mem Clock 630,  Fan 61%  GPU temp 61C GPU Usage 98%  4 WUs concurrently with each WU taking 8 mins to 8 mins 30 secs give or take a bit either way.

Ramping the GPU up to max doesn't benefit all that much (about 20 secs per WU) ?  Taking all factors into consideration I am happy with 8 mins per WU for such an old machine.

Oh yes, this version of BOINC doesn't use an app_info.xml file.   Uses app_config.xml instead.   Example

<app_config>
<app>
<name>hcc1</name>
<user_friendly_name>Help Conquer Cancer</user_friendly_name>
<gpu_versions>
<gpu_usage>0.25</gpu_usage>
<cpu_usage>0.40</cpu_usage>
</gpu_versions>
</app>
</app_config>


Hope this helps  ??? ??? ??? ??? ???

[Dingo]

Thanks for the post I have now successfully downloaded some GPU tasks for WCG.  I updated BOINC to version 7.0.42 on my Windows 64 PC (Bundy)

@Lawryb --  I used the  app_config.xml  that you posted and it downloaded work and I am running 8 tasks on my two GPU's on my  i7 CPU 930 @ 2.80GHz .  I have not changed anything on the GPU's (2] AMD ATI Radeon HD 5800 series (Cypress) (1024MB) driver: 1.4.1741) just running as they came out the box.


It seems to want 1 cpu for each wcg gpu task so that doen't leave anything for CPU projects like POGS.   I will have to look around and see how to drop it down so I only run 2 on each GPU. that will give me 4 cores for cpu and 4 cores for GPU.


OK the first 8 processed and 4 finished and 4 had Computation errors ??

[veebee]

I will have to look around and see how to drop it down so I only run 2 on each GPU. that will give me 4 cores for cpu and 4 cores for GPU.

Dingo, if you change the GPU figure, from "0.25" to "0.5", it will run 2 WU's as opposed to 4, per GPU.


as for the errors, not sure... if they ALL errored out, I would say got to an older version of BOINC, butr still 7.x.x....but as half worked, Im not sure...

Maybe cutting down to 2 at a time will help... I hope.

[Dingo]

I have cut it down to 2 and it is working fine now.  Thanks

[tazzduke]

Hey, have been on and off with the WCG GPU, successfully returning results with no computation errors, I have a HD 1GB 5850, and a q8400 quad, so I only do two tasks at a time, I did try three but it really really slowed down the puter and also the times blew out.

Running the lastest build of boinc, and the app_config.xml file, I was using the latest AMD beta drivers, but they were erroring out on Seti@beta, so I am back on 11.8 drivers and Seti Beta, WCG GPU and others still running good.

Bloody hot weather over here has killed off me doing any more GPU work for the time being, its nuts as the temps for the whole box can get up around the high 50's and low 60's, not seen the GPU go above 60, but then again thats when all the fans go into overdrive to cool the thing down. Did a trial a couple of days ago, when it was 40 deg C outside and the computer was cooler on 33 deg C, well I did shut down Boinc and was using the IPAD in the loungeroom with the AIRCON lol.

Okay, enough of my ranting, good to see some success stories with the team members trying the WCG GPU

Regards
Tazzduke

PS. Oh looky at what they have done to the flags

[kashi]

Until POEM came back earlier this morning I was running 3 concurrent HCC GPU tasks on my HD 5870. I allow 1 CPU core to support each GPU task. GPU usage is 97%. HT is on and I run 7 BOINC cores, so that gives me 3 HCC GPU and 4 CPU POGS. I'm still using app_info.xml on BOINC 7.0.28. CPU is at 3.6 GHz. CPU speed is of course very important for HCC GPU because so much of each task is processed on the CPU.

Core clock 900 MHz up a little from the default 875 MHz of this model 5870. I tried 925 MHz, handled it easily with no errors but it made the screen a bit too jerky when I scrolled anything, so back to 900 MHz. Memory I have left at default 1250 MHz. It runs a bit hotter that way but the memory on my 5970 failed after I gave it jolly good thrashing at MilkyWay and DNETC at a low 500 MHz memory speed.

Tasks take between 11 to 13 minutes depending on how they are aligned, so 220-260 seconds each when divided by 3. They keep synchronising and end up with their CPU phases together and this slows down processing. Running 4 concurrent instead of 3 reduces this and I will switch back to 4 when I do some more HCC, will put my POGS down to only 3 cores though which is a bit low.

The more tasks you run concurrently, the higher the claim so you get a little more credit but it adjusts down again after a while. CreditNew was very poor at the best of times but is even more unsuited to calculating credit where varying numbers of concurrent tasks are involved. You could game it a bit if you sought über credit houndness but I can't be bothered other than setting my card to 990 MHz before BOINC starts to boost the card's reported FLOPS value a little, then reducing back to 900 MHz to crunch. Not even sure that does anything much, still get outclaimed by almost all my wingmen, who are probably running 8 or more tasks concurrently.

I noticed a great 13% improvement in POEM speed by updating to Cat 12.11 beta 8 and it also sped up HCC a bit. On hotter nights I often stop all GPU crunching as the fan gets too noisy.

[veebee]

Although I probably should, I dont worry too much about the heat amd temps, as long as nothings above 70 I dont mind...withe the GPU fans running at around 50%, their temps are in the mid 50's so Im happy.

I have 4 WU's running concurrently on both the 5850 and the 6950.

Not sure if its the fact MY machine (with the 6950 AND a 660Ti) has lots more rubbish running in the background, but my partners machine with the 5850 seems to be smooth on all fronts...no jumpiness etc..whereas THIOS one, is having trouble waving all of Dingo's various flags smoothly in the "send message" window.

I have noticed ONE of my machiines installation of CCC has facility for under/ over voltage of GPU, the other doesnt.. not quite sure if its the same version..probably not.

The WU's seem to vaey a bit in time taken..have upped the memory speed on the 6950 to 770 and mem to 1250 , with a temp of 60 degrees C (mind you the 660 Ti is underneath it, crunching GPUgrid)- WU's take anywhere from 11 minutes to 18 (9most I have noticed , withou going into WCG site,) are around 14 minutes.

The 5850 is set at 755MHz clock speed,   1125 Mhz memory frequency, at  49 degrees C and taking around the 15 minute mark.

would like to get the temp down a bit on the 6950..it is a much newer card and want it to last a BIT longer I suppose.. might pop the 660Ti into the spare machine thats running an AYI4770 ATM...

[Dingo]

I have stopped processing on GPU, not that I do much anyway.  It is just too hot here and getting worse the next few days.  Going to be in the 40's till next Tuesday so the forecast says.

http://www.weatherzone.com.au/nsw/riverina/albury

Crunch on  brothers and sisters:

[LawryB]

Startting to get a few "computation errors" on one of my Win 8 boxes with CAT 12.11 and Afterburner.  For some reason Win 8 reboots (on this machine) whenever it gets an error of any sort.  I guess I should try going back to Cat 12.9 and see jpw ot gpes with the older drivers.    :furious: 


@Dingo.  Hit 35 at noon expecting 37 so we look headed for a 40 plus today, tomorrow and through until next Thursday.  Going to have to stop crunching for a few days and that won't hurt my back pocket either.