Advanced search

Message boards : Number crunching : N00B question = can u run CPU projects + GPU at same time?

Author Message
far
Send message
Joined: 5 Jan 09
Posts: 32
Credit: 1,412,042,305
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5290 - Posted: 5 Jan 2009 | 3:10:04 UTC

Hi, Was wondering - Does this work by running GPU grid on your GPU, and then you can also run different nonGPUgrid CPU based projects on the CPU's spare cycles?
Thanks,
Far

Profile K1atOdessa
Send message
Joined: 25 Feb 08
Posts: 249
Credit: 387,028,788
RAC: 1,197,795
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5297 - Posted: 5 Jan 2009 | 4:52:58 UTC - in response to Message 5290.

Hi, Was wondering - Does this work by running GPU grid on your GPU, and then you can also run different nonGPUgrid CPU based projects on the CPU's spare cycles?
Thanks,
Far


GPUGrid does run on your GPU (see forums for CUDA-enabled graphics cards), using some CPU cycles as necessary.

GPUGrid (current Windows application 6.55) using about 60-80% of a single core for processing of the GPU data. So, this equates to about 30-40% of a dual core (both cores included) or about 15-20% of a quad core (all 4 cores included).

So, say you had a dual core. GPUGrid will use about 30-40% of the total processing power of both cores combined (60-80% of a single core). That will leave an entire other core for a cpu-bound project, plus about 20-40% of the 2nd core for another cpu-bound task. This is called 2+1 in the forums, which is 2 cpu-bound processes + 1 GPU-bound process running at once.

If you currently run 2 cpu-bound projects on a dual core, you will see a slight dip in the processing of one cpu-bound task. However, the power of the gpu results in more credits per time period when compared to a cpu. Linux process scheduling is better than Windows and uses < 5% of a single core. An updated Windows application (hopefully) utilizing much less cpu is expected soon. This will work to minimize the impact on running cpu-bound projects.

Hopefully, this helps some.

far
Send message
Joined: 5 Jan 09
Posts: 32
Credit: 1,412,042,305
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5299 - Posted: 5 Jan 2009 | 5:59:31 UTC - in response to Message 5297.

Thanks for the detail, now I understand what the +1 meant :-)

Does the overhead on windows translate 1:1 for CPU's monitoring GPU's? By that I mean if I had 2 GPU's, then it would take 60-80% of each CPU to monitor the 2 GPU's (and be called 2+2)? Which would imply that to run 4 GPU's under windows you would essentially require a quad core for now (but not on linux) ?

Thanks,
Far

Profile K1atOdessa
Send message
Joined: 25 Feb 08
Posts: 249
Credit: 387,028,788
RAC: 1,197,795
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5307 - Posted: 5 Jan 2009 | 14:19:46 UTC - in response to Message 5299.

Thanks for the detail, now I understand what the +1 meant :-)

Does the overhead on windows translate 1:1 for CPU's monitoring GPU's? By that I mean if I had 2 GPU's, then it would take 60-80% of each CPU to monitor the 2 GPU's (and be called 2+2)? Which would imply that to run 4 GPU's under windows you would essentially require a quad core for now (but not on linux) ?

Thanks,
Far


I have 3 GPU's running, and they take a combined 45-60% of my 4 processor cores. So, I generally have about 2 cores that get full use for a cpu-bound project with a little left over for a 3rd (generally). I guess this would be called 3+3??

Each instance of GPUGrid requires 60-80% of a single core on Windows and that does scale linearly. If you want to run four GPU's, you would need a quad (60-80% of the quads power would be utilized by GPUGrid, leaving a little left over for cpu-bound projects).

Also, GPUGrid restricts the # of WU's sent out to a computer by the # of CPU's (not GPU's) in the system. If you had more than 2 GPU's in a dual core system, 1 GPU would always sit idle because only 2 tasks would ever be sent at once. This might change in the future as the technology and methods to utilize it become more advanced, but right now that is what we are working with.

far
Send message
Joined: 5 Jan 09
Posts: 32
Credit: 1,412,042,305
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5317 - Posted: 5 Jan 2009 | 21:20:59 UTC - in response to Message 5307.

Thanks, I've gotta much better understanding now.

And will discuss finances with the "boss" re updating some graphics cards to give this a go :-)

Far

Profile K1atOdessa
Send message
Joined: 25 Feb 08
Posts: 249
Credit: 387,028,788
RAC: 1,197,795
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5327 - Posted: 6 Jan 2009 | 2:09:18 UTC - in response to Message 5317.

And will discuss finances with the "boss" re updating some graphics cards to give this a go :-)

Far


You're not you're own boss? :-)

If you're adding a new graphics card, you can get a low-end CUDA for $50-60 (9500GT), up through mid-range for $120 (9800GT), up to high-end for $280 (260 series). The prices are just estimates from newegg.com, and of course the prices change a lot and rebates / deals are always there.

Obviously, the more you can afford, the more work will be done in the same amount of time.

However, take into account the heat and size of cards as you go up in power. The low end cards with be not much longer than a pci-e slot, the mid-range will go a few inches past the pci-e slot, and the high end cards will take up most of your case room. Also, be sure to check what psu connections are required for the cards. Again, low end cards don't necessarily require addition power, mid-range probably need a 6-pin, and the high end probably need a 6-pin and an 8-pin. Your psu might limit what you can install as well.

There are a variety of options to suite any budget, case, and overclocking desire. Plenty of research to do now.

Scott Brown
Send message
Joined: 21 Oct 08
Posts: 144
Credit: 2,973,555
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwat
Message 5329 - Posted: 6 Jan 2009 | 4:18:47 UTC

Also be careful not to buy the 8800GTS (320mb/640mb), 8800GTX, or 8800 Ultra which are being reduced by online sellers to rock-bottom prices to clear out old stock. These cards are based on the G80 chip and can only do certain CUDA applications...They will NOT work with the GPUGRID applications. The 8800 GTS (512mb) is based on the G92 chip, however, and will work well.

Profile K1atOdessa
Send message
Joined: 25 Feb 08
Posts: 249
Credit: 387,028,788
RAC: 1,197,795
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 5330 - Posted: 6 Jan 2009 | 4:52:20 UTC

Scott is definitely correct. See the instructions page. There is a thread on this precise information: GPUGrid-capable cards.

Post to thread

Message boards : Number crunching : N00B question = can u run CPU projects + GPU at same time?

//