WCG GPU app for HCC launched

[Jason1478963]

They have launched the gpu app for HCC. I am currently running on a few cards and it seems to run well on the cards I currently have running. I believe you have to have the HCC app selected in your WCG profile and check the box with :
If my computer can process work on my graphics card, then please send me work to run on my graphics card for the projects that I have selected above. it may be possible to run with cpu disabled, but I can't confirm this as I haven't used this option as my other cores run WCG as well.

[AMDave]

Does it suspend & resume ok?
ie If I have my preferences set to suspend GPU work while the computer is in use, will it release the resources and pick the work unit up again where it left off correctly.
Several GPU projects had trouble getting this right.

[Jason1478963]

I have not seen it pick up where it left off, but on a card like the 5850 they only take 4 minutes. 5870 is about 3 and 6570 around 10 with a 6450 in the 20 minute range. I think the HCC work unit on a 1090T CPU core takes an hour and twenty minutes. I have been letting mine run always and haven't tried to suspend them much. It seemed from what I saw in BETA they would restart the tasks if they were suspended.

[Dirk Broer]

The WUs take a full CPU core and a whole GPU each on my Llano systems, anybody dared to write a less resource-eating app_info.xml file yet?

edit: even with one GT430 and two HD 6690D2s -actually two HD 6670s, a HD 6550D/A8-3870K and a HD 6530D/A6-3500- the production is amazing: more than 8k credits, and six pages pending validation in just one day.

[Jason1478963]

The forum at WCG had someone trying to run 2 work units on an nvidia card as this helps to feed the GPU during the parts the CPU does. He thinks it greatly increases the output as one unit is on cpu and one gpu for the most part. I'll keep an eye out for others this weekend.

[AMDave]

suspend and resume works fine.
they run at around 4:50 to 5:50 on a HD5770 Juniper

[Jason1478963]

There has been an ATI appinfo.xml written and posted in the WCG HCC GPU forum. I believe at this time it is setup to just run the HCC GPU work units. This isn't helpful for me as I plan to run other WCG projects along side it. I haven't seen many testers posting the results of this file yet.

[Jason1478963]

I am currently trying to get 3 GPU crunching on this in one XP64 machine. It seems the cards are good as I can run 2 cards at a time. When I try and get the 3rd card to run WCG it seems like the driver keeps giving me the blue screen of death. I had the 3 running collatz for days before this launched and then when testing milkyway things fell apart. I believe Collatz uses more power then WCG for crunching on the GPU according to a 5870 I put on the watt meter. This makes me feel the 750 watt PSU should be fine after running 3 on collatz for days. I tried various different versions of the drivers and end up back at 11.12. I would like to get this machine working with 3 or 4 GPU. Has anybody else had a similar situation that found a solution? Would moving to windows 7 or 8 fix this problem? I spent 6 hours trying drivers and different card configurations to find it doesn't like 12.4, 12.8 with just 2 cards and sometimes 1.

I moved the cards to different slots and cleaned the driver install up and had luck with catalyst 11.12 install

[Dirk Broer]

If you want to run [x,all] applications you have to include [x,all] executables in the app_info.xml.
If you want to run more WUs at a time you have to fiddle with the 'Count' variable/add more GPUs to your system,
and use a cc_config.xml to include all co-processors.

[Jason1478963]

sounds like a pain to do a little bit of tweaking

[Dirk Broer]

It's just the sheer mass of WCG applications. But when you're not into doing more than 1 WU per GPU,
you can stick to just having the cc_config.xml stating 'use all coprocessors' and stick in as many cards as your board can handle.
Once saw a Belgian super-computer, FASTRA II, with seven (?!) video cards (http://www.bit-tech.net/news/bits/2009/12/15/astra-builds-13-gpu-supercomputer/1)
http://fastra2.ua.ac.be/images/fastra_big.jpg

[Jason1478963]

That sure would be nice to have a setup like that right about now.

[Dirk Broer]

My advice: rip out the GTX 295s, replace them with GTX 690s and replace the single GTX 275 with a GTX 660 (to make FASTRA III).

And build a competing machine using six HD 7990s and a single HD7950 for comparison! Name it FASTRA IV

A single GTX 295 has a computational power of 1788 GFlop SP and 149 GFlop DP, a GTX 275 does 1011 vs 84
A single GTX 690 has a computational power of 5622 GFlop SP and 234 GFlop DP, a GTX 660 does 1880 vs 78

A single HD7990 has a computational power of 7000 GFlop SP and 1750 GFlop DP, a HD7950 does 2867 vs 717

FASTRA IV wins!

[Jason1478963]

That 7990 looks pretty hard on the pocket book. We did play with the app_info file that was posted at WCG. It seems I have other bottlenecks that make it difficult to keep 3 GPU busy. when running 2 work units per card It didn't seem to help the times that much. I also got to deal with the driver issue again in xp64.

[Ramjet]

If you want to run [x,all] applications you have to include [x,all] executables in the app_info.xml.
If you want to run more WUs at a time you have to fiddle with the 'Count' variable/add more GPUs to your system,
and use a cc_config.xml to include all co-processors.

Any idea where to get the executables for the various projects? I read somewhere you can find them in the sched_reply_www.worldcommunitygrid.org.xml file, but that doesn't look like the right info to me. I'm not sure exactly what I need though.

[Dirk Broer]

The executables are the least of your worries!
If all is right, they are already in the right directory: C:\ProgramData\BOINC\projects\www.worldcommunitygr id.org

[plonk420]

edit: NEVER MIND ... it's working now :D

what drivers/openCL/compute versions are you guys using? seems while my stuff was new enough to crunch Bitcoin last year, it seems they're not new enough to do HCC...

http://i1.minus.com/icdOo87Ro8gBJ.png

[AMDave]

Using default Catalyst 12.9 driver in Ubuntu 13.04 AMD64 [***], including ia32 libs and OpenCL 1.2 AMD-APP (1084.4)
with BOINC 7.0.52
BONC event log shows:
Mon 25 Feb 2013 14:36:56 EST | | CAL: ATI GPU 0: AMD Radeon HD 7700 series (Cape Verde) (CAL version 1.4.1741, 1024MB, 559MB available, 3360 GFLOPS peak)
Mon 25 Feb 2013 14:36:56 EST | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7700 series (Cape Verde) (driver version 1084.4 (VM), device version OpenCL 1.2 AMD-APP (1084.4), 1024MB, 559MB available, 3360 GFLOPS peak)
Mon 25 Feb 2013 14:36:56 EST | Collatz Conjecture | Found app_info.xml; using anonymous platform
Mon 25 Feb 2013 14:36:56 EST | World Community Grid | Found app_config.xml
Mon 25 Feb 2013 14:36:56 EST | | Config: use all coprocessors

[***] This is alpha for my testing.
Not recommended for normal use.
If you want to use Ubuntu I recommend 12.04 LTS.
I think default driver in 12.04 is up to 12.4 with OpenCl 1.2 IIRC.
yup. http://packages.ubuntu.com/search?keywords=fglrx-amdcccle
Experimental driver is Catalyst 12.11 (see "precise-updates" "restricted")
but if you can get working with the default drivers you are much better off as they are tested and supported.

[AMDave]

edit: NEVER MIND ... it's working now :D
Well done.
...but that is OpenCL 1.1
It looks like you are using Catalyst 10.1
You'd be better using OpenCL 1.2, say Catalyst 12.9 or even the latest and greatest Catalyst 13.1 should work on your Win2K8-R2 server there.
Be sure to read the doco before you install though.

[plonk420]

i'm using an early 2012 driver ... however, wikipedia seems to report that my GPU maxes out at OpenCL 1.1

[Dirk Broer]

Wikipedia reports OpenCL 1.2 capability for HD 5000 series (http://en.wikipedia.org/wiki/Evergreen_(GPU_family))

[AMDave]

On hind-sight I recant my advice now that it is working for you.
If it ain't broke, don't fix it :icon_mrgreen:

[plonk420]

Wikipedia reports OpenCL 1.2 capability for HD 5000 series (http://en.wikipedia.org/wiki/Evergreen_(GPU_family))

oh, you, wikipedia x)

http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units#Evergr een_.28HD_5xxx.29_series
http://en.wikipedia.org/wiki/Radeon_HD_5000_Series


On hind-sight I recant my advice now that it is working for you.
If it ain't broke, don't fix it :icon_mrgreen:

it's sad we have to think like that, what with occasional driver oddities :D

[Jason1478963]

yes... I found if your running them in xp64 you cannot get drivers newer then 12.1 if I remember right. The 5xxx series cards are not very impressive on HCC at WCG , but they do still get a fair amount of work done compared to a CPU. The 7xxx series with the gain in power efficiency really seems to work well at WCG right now. Don't make the same mistake as me and buy 5xxx series with waterblocks for this project.

[Dirk Broer]

The HD 5000 series definitely supports OpenCL 1.2, that's why you can use them in WCG. OpenCL 1.2 support is a prerequisite for WCG.
I have a OpenCL 1.1 capable HD 4770 which cannot be used for WCG, so it is running mainly MilkyWay.
I'll change the Wikipedia mentioning of OpenCL 1.1 above and below the table for the HD 5000 series (mind you: the infobox has the right mentioning)

Here's a complete list from AMD to see which AMD GPUs will operate using OpenCL 1.2 with the latest SDK (HD 5000 family included) (http://developer.amd.com/tools/heterogeneous-computing/amd-accelerated-parallel-processing-app-sdk/system-requirements-driver-compatibility/)

Looks like Plonk's OS is the culprit (Win2K8-R2 server is not on the list of supported OS-es for the last SDK)...

[Jason1478963]

XP 64bit was not on the list, but it is able to run under its current setup with the 5xxx and 6xxx series cards. I believe i'm running 11.12 on my xp64 bit system and that may be one of the last versions they supported OpenCL on xp.

[plonk420]

Don't make the same mistake as me and buy 5xxx series with waterblocks for this project.

if all else fails, you can just do bitcoin if you have cheap power ;)