My CUDA Story

Questions and Answers : GPU applications : My CUDA Story
Message board moderation

To post messages, you must log in.

AuthorMessage
Peter Farrow

Send message
Joined: 15 Jun 99
Posts: 6
Credit: 5,974,136
RAC: 0
United Kingdom
Message 917455 - Posted: 13 Jul 2009, 21:09:22 UTC

I thought I would add my experience to the CUDA forum, just in case there are others considering an upgrade.

I am runnig boinc on a 3 year old HP Pavilion T3070uk. recently the screen suffered quite a lot of flickering, despite increasing the refresh rate and so I thought it must be time for a new graphics card.

Having seen the detail regarding CUDA, I decided to add this capability, and replaced my old Nvidia Geforce 6200 with a Geforce 9500 GT.

The installation was easy, by following the instructions with the new card. Drivers were uninstalled prior to the changeover, and the latest driver installed from the Nvidia site after reboot.

Boinc saw the card straight away and downloaded the CUDA work units.

BUT, then my system crashed after about 25 minutes, and continued to do so after anything from 10 to 30 minute intervals.

My first port of call to resolve this issue was this forum, and was lucky enough to find a post relating to this problem.

Having followed all instructions regarding removing all graphics drivers using Driver Sweeper, and then a reload from Nvidia, hey presto, problem solved.

The card has run several units so far over the past 2 days, and the units seem to be completed in 45-50 minutes, even thought the card only has 512Mb of ram.

The moral of this story.....check the forums, there is always someone out there willing to help, it is as they say a community!

As I mentioned in other posts I am somewhat of an amateur when it comes to computer architecture and operating systems so here are a couple of questions:

I also downloaded Atitools, which as you will be aware has an overclocking feature, it says my card, currently at 550Mhz could run at 750Mhz, so:

1. By overclocking the card, I suppose this would make the units run even faster?

2. I also suppose that overclocking will increase its temperature (at present Atitools states it runs at an average of 50-52 degrees)

3. If I tried to overclock, up to the maximum, should I do this in stages to review the resulting effect (ie temperature increase?)

4. If question 2 is true, what should I aim at as a maximum temperture when overclocked?

5. will overclocking shorten the life of the card or affect other parts of the system?

Thanks guys,

Regards

Peter


ID: 917455 · Report as offensive
agnawt

Send message
Joined: 25 Jun 07
Posts: 15
Credit: 4,838,223
RAC: 0
Israel
Message 917460 - Posted: 13 Jul 2009, 21:35:22 UTC

Overclocking beyond spec always decreases the lifetime of the chip and other components part of the same system. See http://en.wikipedia.org/wiki/Overclocking and the section "Disadvantages" for a short overview of some issues. Given that CUDA is especially sensitive to card stability (it processing function implies needing much greater accuracy than graphics processing which can get away with some noise), any factor that pushes the card to its edge cant make things better.

Overclocking might help finish WU's faster, but u might have more that fail to complete or cant be verified.

If you do decide to go the overclocking route, I would not rely on temperature alone as the guide for how high u can go. The higher the frequency, the more power the GPU will need and that equation is definitely not linear.

If 550Mhz is stock speed and 750Mhz is the limit, keep to 650mhz tops. Move in small increments from stock speed. Also, the GPU's RAM frequency should not be touched - the chips used usually barely fit the stock spec and are often already overclocked - no headroom there mostly.

Keep trying to run CUDA samples at every step from the CUDA SDK. As long as the more complex ones (especially those that verify their datasets with the CPU) can run u should be ok; for how long is another story entirely ;-)
ID: 917460 · Report as offensive
Profile BigDaddyDave
Avatar

Send message
Joined: 8 Oct 00
Posts: 67
Credit: 20,337,995
RAC: 15
United States
Message 917655 - Posted: 14 Jul 2009, 17:10:35 UTC - in response to Message 917455.  
Last modified: 14 Jul 2009, 17:14:09 UTC

Hi Peter Farrow,

I would suggest that you download NVIDIA’s free GPU over clocking tool, it has a lot of the answers to your questions in an easy to use GUI (I've bumped up the original message in this forum for you to click on) search for "EVGA Precision Over clocking Tool v1.7.1" or just click http://setiathome.berkeley.edu/forum_thread.php?id=53202. Also, per some of the members, you don't even need an NVIDIA card to use it on some GPU.

Regarding the quest to over clock, this tool will let you test it first to prevent OS crashes. As for MHz increments, I found bumping it up in 10 MHz increments at a time and then waiting about 10 seconds for the system to catch up works best. For the temps, you can use the build in fan adjustments. I'm at 80% but may increase since it is summer.

And last down load NVIDIA CUDA 2.0, it will crunch your SETI CUDAs faster! (agnawt added the link for you)


Get crunchin!

BDDave
ID: 917655 · Report as offensive

Questions and Answers : GPU applications : My CUDA Story


 
©2024 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.