CUDA

I checked the option to run two threads concurrently to increase the workload on my GTX295.

http://www.primegrid.com/results.php?hostid=95767&offset=20&show_names=0&state=3&appid=9

<app_info>
<app>
<name>pps_sr2sieve</name>
<user_friendly_name>Proth Prime Search (Sieve)</user_friendly_name>
</app>
<file_info>
<name>primegrid_tpsieve_1.35_windows_intelx86__cuda23.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart.dll</name>
</file_info>
<app_version>
<app_name>pps_sr2sieve</app_name>
<version_num>135</version_num>
<file_ref>
<file_name>primegrid_tpsieve_1.35_windows_intelx86__cuda23.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart.dll</file_name>
<open_name>cudart.dll</open_name>
</file_ref>
<plan_class>cuda23</plan_class>
<avg_ncpus>0.050000</avg_ncpus>
<max_ncpus>0.050000</max_ncpus>
<coproc>
<type>CUDA</type>
<count>0.500000</count>
</coproc>
</app_version>
</app_info>


GTX295:
GPU Takt (Geometric Domain) 710 MHz
GPU Takt (Shader Domain) 1512 MHz
Speichertakt 783 MHz (Original: 1025 MHz)

System: Microsoft Windows 7 Professional

One thread / GPU:
GPU load ~86% to ~88%, runtime ~980sec to 1020sec

Two threads per GPU:
GPU load ~95% to ~98%, runtime ~1780sec to 1820sec = ~ 900sec/single thread.

Conclusion: ~ 10% less runtime per WU


The same test on a 9800GT (GPU load of 95 % with one thread/GPU) did neither increase the load nor reduce the runtime - but increase the CPU-time.
____________
Member of Crunching Family
http://crunching-family.at/

where i can get optimized app in order to play with app_info.xml? or I just should create app_info.xml in the project folder?

OK, clear :-)

So, I just should create app_info.xml in the project folder? I'm running cuda on linux

I'm running older then fermi card (GTX275), but I'd like to play aroung app_info. What I do not like in PrimeGrid - my rig's interface on linux (and other guys reporting the same issue on windows) became very slow in terms of user responce (I'm comparing it with GPUGRID, where GPU usage is barely 70%). Ideally - I'd like to have two app_info files and switch between them: one with slower crunching speed, but when I can use my rig and another one - for max speed, when I'm away and do not really care about responce. I think small bash script will work fine to this stuff.

I also tried to set up PrimeGrid as backup project of Milky Way (MW) on my rig with ATI cards, but, as u mentioned, I got no WUs. What should I do in terms of app_info in order to get WUs? May be, let's try to make app_info file for the community. I'm ready to be a tester :-)

Trust me - I know what app_info is about. At least - on MW and Collatz :-) So, I'm ready to play :-)
____________

and one nore thing: as far as I understand nvidia is way better then ATI in terms of speed, right? Although GFLOPs are higher for ATI... For instance, I saw that here
____________

that's a good idea to click "do not use GPU" in BOINC manager :-)

But also I'd like PrimeGrid app run 24/7 and only when I need my rig it will allow somewhat 10% of GPU for the operation system (OS). It's literally pain in @$$ ever during web surfing to wait some time to see tha page even with 15MB/s channel :-)

I'll try to find that topic a bit later and play with different parameters. But any clue from project development team are very much welcome and appreciated :-) I do not think this should be "CIA top secret" :-)
____________

Saying frankly, my plan was to buy two ATI 6970 cards for MW rig (it's for crunching only, so I don't mind to run windows there) and one 6970 to run my main rig on linux. While MW do not have OpenCL app yet, my plan was to run Collatz while it will be available. But now I wanna crunch for PrimeGrid and setup PrimeGrid as backup project on my crunching rig.

So, I wonder if there are chances that OpenCL app will be as fast as CUDA one and what I should buy for PrimeGrid - 6970 or 580/570?
____________

I quickely estimated thousands credits per day based on link above. Here what I've got:
GTX 580 - 322
GTX 275 - 183 (my card)
GTX 460 - 153
58Ñ…0 - 95

Pretty much clear.
____________

Well, GTX 460 is slower then GTX 275 (907.2 GFLOPs vs 1010.880 GFLOPs - by wiki) and more or less the same amount it slower in that comparison.

so looks my comparison it pretty acurate OR .Steve got room for improvement :-)

But I really should consider 580/570. 570 price looks to be $349, what looks attractive :-)
____________

1.2 Mil for a Pair of 580's & about 600k for a pair of 460's >>> http://stats.free-dc.org/stats.php?page=user&proj=pgrid&name=103

thanks a lot, Steve :-) I believe it's to be put somewhere in FAQ

BTW, did u OC'd your cards or they running on stock clocks?
____________

Trying to make sense of all the posts dealing with ATI cards, app_info.xml and different platforms. I've noticed that there are several versions of the app_info file being shared and from what I've been able to decipher the app_info file takes control of the PG jobs being sent. However, with so many different posts it was hard to see which one I needed and exactly what folder I needed to put it in. I'm running Windows XP with an ATI Radeon HD 2400XT.

My request is to ask one of you to post the app_info file for the 32bit XP and the commands to get PPS Sieve running along with PPS LLR and SG LLR and what folder to place it in. I think once I get a file that works with these I can figure out how to add or remove jobs. The most important part is getting the GPU up and running to test the PPS Sieve jobs.

Any help would be greatly appreciated - Rick

Great :(

Thanks for the info

My 4 core Vista64 box w/ hd-4890 is crunching away

w/ no app_info.xml file

ppsieve GPU (running (.31CPUs + 1.00 ATI GPUs)
4x Sophie CPU WU's

Thanks Rytis

Steve

My 4 core Vista64 box w/ hd-4890 is crunching away

w/ no app_info.xml file

ppsieve GPU (running (.31CPUs + 1.00 ATI GPUs)
4x Sophie CPU WU's

Thanks Rytis

Steve

How do you get this to work with linux without an app_info?

I get the error below since one of the requirements of ATI Stream SDK is setting a couple of environmental variables (LD_LIBRARY_PATH, ATISTREAMSDKROOT and ATISTREAMSDKSAMPLESROOT). I know the standalone app works.

So where should these variables be set so BOINC sets these variables correctly?

I believe it is now possible to run ATI cards without app_info.

I believe it is now possible to run ATI cards without app_info.

Would it be possible to set up lower values for "<avg_ncpus>0.55</avg_ncpus>" and "<max_ncpus>1</max_ncpus>" ?
On a host with a single gpu is this value no problem but on a dual gpu host causes this a blocking from one cpu-core. An idle cpu-core would be useless because the work is done by the gpu only.

I changed something, is it better now?

How do you get this to work with linux without an app_info?

I get the error below since one of the requirements of ATI Stream SDK is setting a couple of environmental variables (LD_LIBRARY_PATH, ATISTREAMSDKROOT and ATISTREAMSDKSAMPLESROOT). I know the standalone app works.

So where should these variables be set so BOINC sets these variables correctly?

....
<flops>1.0e11</flops>
....

what's mean flops?
cpu's? or gpu's? or other means?
who know that means?

in my case, increase processing time at two thread.. :(

ah....
i'm a stupid...
the problem is SLI :(

I get a computation error on startup for Proth Prime Search ( Sieve ) V1.35 the WU's seem to stop after a few seconds.

Not too much of a problem I just abort the WU in question.

The WU's for Proth Prime Search ( Sieve ) V1.36 appear to run OK.

Is there anything I can do to prevent downloads of Proth Prime Search ( Sieve ) V1.35 WU's.

I am running a Windows 7, 64 Bit Intel Dual Core machine with a Radeon HD4570 GPU.

Any suggestions and help would be most welcome.

Thanks
____________