Running SETI@home on an nVidia Fermi GPU

Message boards : Number crunching : Running SETI@home on an nVidia Fermi GPU
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 . . . 16 · Next

AuthorMessage
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 13190
Credit: 154,595,211
RAC: 197,980
United Kingdom
Message 997947 - Posted: 22 May 2010, 17:12:15 UTC

I'm starting this thread as a place where people who are successfully running SETI on Fermi can share experiences and pass on suggestions. Please keep at least the first few messages free for factual information which people may want to refer to in the future - keep any questions or problems for later, or post them in one of the other threads. Thank you.
ID: 997947 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 13190
Credit: 154,595,211
RAC: 197,980
United Kingdom
Message 997948 - Posted: 22 May 2010, 17:13:00 UTC

Re-posting, by popular request, from CUDA MB V12b rebuild supposed to work with Fermi GPUs.

This note is aimed at people running optimised apps already. That's essential: you will need to modify an existing app_info.xml file.

This isn't the place to go into the whole science of installing optimised applications by hand, but in brief: stop BOINC first, keep backups, use a simple ANSI-mode text editor. If in doubt, read the BOINC FAQ version - fuller, and can be updated in the future.

You will need to add a whole new <app_version> section:

    <app_version>
        <app_name>setiathome_enhanced</app_name>
    	<version_num>610</version_num>
    	<avg_ncpus>0.200000</avg_ncpus>
    	<max_ncpus>0.200000</max_ncpus>
    	<plan_class>cuda_fermi</plan_class>
	<file_ref>
            <file_name>setiathome_6.10_windows_intelx86__cuda_fermi.exe</file_name>
            <main_program/>
	</file_ref>
	<file_ref>
            <file_name>cudart32_30_14.dll</file_name>
	</file_ref>
   	<file_ref>
            <file_name>cufft32_30_14.dll</file_name>
   	</file_ref>
	<file_ref>
            <file_name>libfftw3f-3-1-1a_upx.dll</file_name>
   	</file_ref>
	<coproc>
            <type>CUDA</type>
            <count>1</count>
	</coproc>
    </app_version>

You'll also have to add <file_info> sections for any file you're not already using. You may have libfftw3f-3-1-1a_upx.dll already, in which case you don't need to add it again: but the full set is

    <file_info>
        <name>setiathome_6.10_windows_intelx86__cuda_fermi.exe</name>
        <executable/>
    </file_info>
    <file_info>
        <name>cudart32_30_14.dll</name>
        <executable/>
    </file_info>
    <file_info>
        <name>cufft32_30_14.dll</name>
        <executable/>
    </file_info>
    <file_info>
        <name>libfftw3f-3-1-1a_upx.dll</name>
        <executable/>
    </file_info>

Finally, you'll need the files themselves: download from

http://boinc2.ssl.berkeley.edu/beta/download/setiathome_6.10_windows_intelx86__cuda_fermi.exe (2,476 KB)
http://boinc2.ssl.berkeley.edu/beta/download/cudart32_30_14.dll (280 KB)
http://boinc2.ssl.berkeley.edu/beta/download/cufft32_30_14.dll (12,188 KB)
http://boinc2.ssl.berkeley.edu/beta/download/libfftw3f-3-1-1a_upx.dll (439 KB)

Make sure the file sizes are correct - sometimes these manual downloads crash out early.
ID: 997948 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 13190
Credit: 154,595,211
RAC: 197,980
United Kingdom
Message 997954 - Posted: 22 May 2010, 17:33:41 UTC

I'm running Fermi - a GTX 470 - on my dual Xeon host 2901600. I had to reorganise some cabling this afternoon, and took the opportunity to hook it up to a power meter (Killawatt style) for interest.

Base load: 235 - 240 watts.
It's quite an old host, a Dell Precision 490 workstation based on an OEM version Intel 5000 server motherboard. This quite high baseload figure was taken at the Windows 7 desktop, with the Fermi driving dual 1600 x 1200 monitors (only the GPU power is included in the measurement, the monitors are on a separate circuit).

Fermi running OpenCL: 325 - 330 watts.
Fermi running CUDA: 315 - 320 watts.
These readings come from an offline benchmark test of two AQUA applications - BOINC project work, but without BOINC itself actually running. Although the CUDA version used less power, it was actually almost 10% faster - the benchmark took 31:24 under OpenCL, but only 28:45 under CUDA. One of the reasons for the extra power draw is that the OpenCL application used almost a full CPU core, and the CUDA application used hardly any CPU.

BOINC running on CPU only: 325 - 330 watts.
That's actually seven tasks, four CPDN (high memory usage) and three QuantumFIRE Alpha (low memory usage). I think using memory is energy-expensive on this machine, because of the server-class FB-DIMMs.

BOINC running fully, including Fermi: 375 - 390 watts.
The SETI app jumps around a lot while running. GPUGrid is steadier, but the numbers still kept flickering.

So, very roughly, it appears that the GTX 470 has added an extra 60 watts to my average crunching bill. Not green, but better than I expected.

There are suggestions that the forthcoming CUDA 3.1, still under restricted developer NDA, will increase performance significantly: no doubt that will also mean increased power usage too.
ID: 997954 · Report as offensive
Profile SciManStev Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Jun 99
Posts: 6507
Credit: 120,996,918
RAC: 1,510
United States
Message 997965 - Posted: 22 May 2010, 18:21:43 UTC

With my two GTX 480's installed, with the 1.8 amp fans removed and replaced with water blocks I am pulling about 231 watts with the system at idle. 523 watts running BOINC with GPU's snoozing, and 671 watts with everything running. The CPU is an i7 980 OC'd to 4.18 GHz.

Steve
Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website
ID: 997965 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7104
Credit: 147,374,587
RAC: 1,456
Germany
Message 997990 - Posted: 22 May 2010, 19:29:46 UTC
Last modified: 22 May 2010, 19:39:20 UTC

~ 60 W (GTX470), ~ 74 W (GTX480) for 'full loaded' Fermi GPU, look like very small.
A well sign that the GPU isn't used complete with the current available CUDA app/version.

My OCed GTX260-216 use ~ 40 W @ idle and ~ 140 W @ full load with CUDA (~ 100 W).
ID: 997990 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7104
Credit: 147,374,587
RAC: 1,456
Germany
Message 998436 - Posted: 24 May 2010, 10:20:57 UTC

Is there also at least an other person out there, which was/is curious about if the new 6.10_Fermi_app will run also on non Fermi GPUs?

This person made maybe already a test if the app work for example on GTX2xx series GPUs?
How it worked?

So maybe I don't need to crash also my WUs.. :o)

ID: 998436 · Report as offensive
Profile -= Vyper =-
Volunteer tester
Avatar

Send message
Joined: 5 Sep 99
Posts: 1584
Credit: 864,566,488
RAC: 1,161,058
Sweden
Message 998457 - Posted: 24 May 2010, 12:17:45 UTC

Yup i've tried it running Cuda3 paired with a GTX470 and GTX295's.
It worked flawlessly for me so give it a go if you want..
But i think it will reduce productivity for you Sutaru.
Those times are very nice indeed that you have.

Kind regards Vyper

_________________________________________________________________________
Addicted to SETI crunching!
Founder of GPU Users Group
ID: 998457 · Report as offensive
Profile SciManStev Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Jun 99
Posts: 6507
Credit: 120,996,918
RAC: 1,510
United States
Message 998462 - Posted: 24 May 2010, 12:46:12 UTC - in response to Message 998457.  

Yup i've tried it running Cuda3 paired with a GTX470 and GTX295's.
It worked flawlessly for me so give it a go if you want..
But i think it will reduce productivity for you Sutaru.
Those times are very nice indeed that you have.

Kind regards Vyper


Darn! If my 295 was worth anything, and I had waited to get the 480's with nozzles on the end facing out, that would have been perfect!

Steve
Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website
ID: 998462 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7104
Credit: 147,374,587
RAC: 1,456
Germany
Message 998469 - Posted: 24 May 2010, 13:32:19 UTC - in response to Message 998457.  

Yup i've tried it running Cuda3 paired with a GTX470 and GTX295's.
It worked flawlessly for me so give it a go if you want..
But i think it will reduce productivity for you Sutaru.
Those times are very nice indeed that you have.

Kind regards Vyper


Thanks for reply!

Why you think it's maybe not worth for a try (for me)?
Maybe I'll see a slowdown in calculation speed with my old GTX2xx graphic cards?

What you saw on your GTX295 graphic cards?

ID: 998469 · Report as offensive
Profile -= Vyper =-
Volunteer tester
Avatar

Send message
Joined: 5 Sep 99
Posts: 1584
Credit: 864,566,488
RAC: 1,161,058
Sweden
Message 998532 - Posted: 24 May 2010, 16:59:26 UTC - in response to Message 998469.  
Last modified: 24 May 2010, 16:59:50 UTC



Thanks for reply!

Why you think it's maybe not worth for a try (for me)?
Maybe I'll see a slowdown in calculation speed with my old GTX2xx graphic cards?

What you saw on your GTX295 graphic cards?


Well it slowed down for me it seemed like, maybe faster on VHARs but not with the regular 0.3 - 0.6 AR wu's , i got higher completion times in my case.
And to that you need to run the Rescheduler so the machine won't stall.
It's not worth the hassle for what i concern but that's entirely up to you :)

Regards Vyper

_________________________________________________________________________
Addicted to SETI crunching!
Founder of GPU Users Group
ID: 998532 · Report as offensive
Profile MadMaC
Volunteer tester
Avatar

Send message
Joined: 4 Apr 01
Posts: 201
Credit: 47,158,217
RAC: 0
United Kingdom
Message 999164 - Posted: 27 May 2010, 17:39:32 UTC

Ok, I need some advice on how to spend £300!

All I am interested in is crunching - not games.
Typical me, I was looking at a machine filled with 9800 GX2's, but I got excited and bought a 295 of ebay. It had a fault so I returned it, but I loved the output it gave me :-)
So I have to buy another card
I know there is no hard and fast data, but what will give the best ppd?
A GTX295 or a GTX470 as there isn't that much price difference now
I know Fermi has a different design and the 256 drivers and CUDA 3.1 are meant to improve things further so if you had the money, what would you do???
Help a cruncher in need :-)
ID: 999164 · Report as offensive
ftpd
Volunteer tester

Send message
Joined: 1 Mar 07
Posts: 11
Credit: 4,117,544
RAC: 0
Netherlands
Message 999688 - Posted: 31 May 2010, 12:22:05 UTC - in response to Message 999164.  

@MadMac
I use both types (295 and 470/480). Go for the 470!
This one is fast, no noise and no heat.

You have to wait for seti 6.10 (fermi) or go to the beta-site.

Ton (ftpd) Netherlands
ID: 999688 · Report as offensive
Profile MadMaC
Volunteer tester
Avatar

Send message
Joined: 4 Apr 01
Posts: 201
Credit: 47,158,217
RAC: 0
United Kingdom
Message 999711 - Posted: 31 May 2010, 14:51:21 UTC - in response to Message 999688.  

Already ordered, arrives tommorrow :-) I went for a 470 and will overclock it :-)
ID: 999711 · Report as offensive
Speedy
Volunteer tester
Avatar

Send message
Joined: 26 Jun 04
Posts: 1348
Credit: 10,600,614
RAC: 3,259
New Zealand
Message 1000755 - Posted: 4 Jun 2010, 23:48:12 UTC - in response to Message 999688.  
Last modified: 4 Jun 2010, 23:50:34 UTC

@MadMac
I use both types (295 and 470/480). Go for the 470!
This one is fast, no noise and no heat.

You have to wait for seti 6.10 (fermi) or go to the beta-site.

Ton (ftpd) Netherlands

Has been in testing sine 19 May. As yet haven't tested it.

I have a NVIDIA GeForce GTX 470 (1248MB) with driver: 19775 I ran GPU Grid for a bit it was sitting round the 71 degree c mark.
ID: 1000755 · Report as offensive
John

Send message
Joined: 21 May 99
Posts: 51
Credit: 5,667,907
RAC: 0
United States
Message 1000943 - Posted: 5 Jun 2010, 15:11:31 UTC

I have 6.10.58 installed with a 470 card. Can i now run seti@home or do I need
the below listed files installed?

Aqua@Home and Collatz and GPUGrid appear to be running fine. Milkway@Home and Seti@Home appear to be doing alot of shorties 17sec and some 1.5hrs.
I have not checked to see if they Validated yet.
Some running cuda 2.0 2.3 and 3.0


ID: 1000943 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1000949 - Posted: 5 Jun 2010, 15:34:52 UTC - in response to Message 1000943.  

Something very strange going on John. You seem to be completing the VLARs ok on your GPU but not the shorties. Looks like all your shorties are giving -9 overflows.I didn't see any in between. That's not good. I'm sure someone more experienced with Fermi will be along soon to give you a hand on that.


PROUD MEMBER OF Team Starfire World BOINC
ID: 1000949 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 13190
Credit: 154,595,211
RAC: 197,980
United Kingdom
Message 1000950 - Posted: 5 Jun 2010, 15:38:37 UTC - in response to Message 1000943.  

I have 6.10.58 installed with a 470 card. Can i now run seti@home or do I need the below listed files installed?

If you want to run a Fermi card at SETI, you MUST install the files listed below (or an equivalent set from another supplier).

The main SETI project does not yet have a compatible program installed for general use - Beta testing has not yet been declared complete.

If you do not wish to use the 'Anonymous Platform' facility to install the compatible program manually, set 'No New Tasks' for SETI and crunch for another project in the meantime. Keep an eye on the SETI Application page: that's the first (and probably only) place any Fermi-compatible application will be announced. And like the two previous CUDA applications, it will probably be Windows-only.

In the meantime, I'm afraid that all those 17-second tasks are not shorties: they are errors, and a waste of the project's resources. Worse, some of them are starting to pollute the project's scientific database with false data. Please consider aborting any you still have left in your cache.
ID: 1000950 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 13190
Credit: 154,595,211
RAC: 197,980
United Kingdom
Message 1000952 - Posted: 5 Jun 2010, 15:40:01 UTC - in response to Message 1000943.  

I have 6.10.58 installed....

I would be surprised if that were true: it doesn't exist yet :-o
ID: 1000952 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1000964 - Posted: 5 Jun 2010, 16:01:41 UTC - in response to Message 1000950.  

Sorry Richard, I didn't look at enough of his work. The ones I picked to check happened to finish as shorties for his wingmen. Still strange though that he can complete VLARs ok.


PROUD MEMBER OF Team Starfire World BOINC
ID: 1000964 · Report as offensive
John

Send message
Joined: 21 May 99
Posts: 51
Credit: 5,667,907
RAC: 0
United States
Message 1000969 - Posted: 5 Jun 2010, 16:11:21 UTC

I had installed only the CUDA Toolkit as i use for other apps. Other than that I just updated to 6.10.56 (Excuss may typo from last post)

I have set seti to no new tasks for now and have downloaded the files listed below.

In what order do Install them and where to they go. Sorry for the newbie question.

Thanks in advance

ID: 1000969 · Report as offensive
1 · 2 · 3 · 4 . . . 16 · Next

Message boards : Number crunching : Running SETI@home on an nVidia Fermi GPU


 
©2019 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.