100MDPP
 

Uncwilly, next version of mfactc should be running on Windows (or Cygwin) and include my routines to parse Prime95 worktodo.txt and worktodo.ini files. :smile:

Luigi

Maybe i will have to break down and by a box. At some point within the next 18 months I am expecting a sort of settlement check. It should be in the range of 7 reasonably well outfitted PC's, plus interest and penalties, less taxes. So, when that ship comes in, I will likely build a couple rockin' boxen.

[QUOTE=ET_;214062]Note that checking one OBD exponent from 1 to 69 requires about 90 [COLOR="Red"]minutes[/COLOR] with Factor5 (2 threads) and less than 85 [COLOR="Red"]seconds[/COLOR] with mfaktc.[/QUOTE]Holy crap!!! :shock:

Edit: Uncwilly, if it really is so bad to use the PC while crunching on the graphics card as ET says, perhaps you should consider a motherboard with onboard graphics, use that to drive the display and then have three or four nVidia cards in SLI for crunching. Pretty much your own factoring supercomputer.

I should point out, that with code that fast on the graphics card, it would take but 5 days 6 hours from now, not to complete the current level, but the one after that.

(Also it would be cool if you could arrange it so you complete both levels on the same day. :wink:)

[QUOTE=lavalamp;214096]Holy crap!!! :shock:

Edit: Uncwilly, if it really is so bad to use the PC while crunching on the graphics card as ET says, perhaps you should consider a motherboard with onboard graphics, use that to drive the display and then have three or four nVidia cards in SLI for crunching. Pretty much your own factoring supercomputer.[/QUOTE]

A small GPU for the display: good idea!

Three or four GPUs for crunching: not so good idea... ;)
Lets asume you've a Core i7 750 (2.66GHz + turbomode) than I recommend to
- dedicate [B]1-2 cores[/B] to feed [B]one GPU[/B] like GTX 275 / GTX 285
- dedicate [B]2-3 cores[/B] to feed [B]one GPU[/B] like GTX 470 / GTX 480

Oliver

Hm, I didn't realise the CPU usage would be so high when the number crunching is done on the graphics card. Can anyone post some CPU usage figures on various systems (on link me to somewhere they are already posted)?

What work does the CPU actually need to do?

Hi,

CPU does sieving of candidates while the GPU does the calculation of the remaining candidates (those who haven't been sieved out).

Want a link: [url]http://www.mersenneforum.org/showthread.php?t=12827[/url] :razz:

The CPU usage will be 100% of one core (per instance of mfaktc). Even on a superfast CPU with a slow GPU because I was to lazy to fix this. On faster GPUs this is not an issue!

Oliver

[QUOTE=TheJudger;214307]Hi,

CPU does sieving of candidates while the GPU does the calculation of the remaining candidates (those who haven't been sieved out).

Want a link: [url]http://www.mersenneforum.org/showthread.php?t=12827[/url] :razz:

The CPU usage will be 100% of one core (per instance of mfaktc). Even on a superfast CPU with a slow GPU because I was to lazy to fix this. On faster GPUs this is not an issue!

Oliver[/QUOTE]

If lavalamp needs to get low CPU usage, he can use a lower bound of SievePrimes, so that the CPU sits idle while the GPU finishes its work, I guess.

Luigi

Sadly I can't use the code. I have an extremely powerful ATI card that I can't use for crunching.

I was under the impression that ATI cards had more grunt than nVidia cards, yet most work is being done with CUDA over Firestream, or best yet, OpenCL.

Perhaps I will have a little dabble with Firestream/OpenCL, but I have extreme scepticism that I'll be able to produce any efficient, or even workable, code.

[QUOTE=ET_;214359]If lavalamp needs to get low CPU usage, he can use a lower bound of SievePrimes, so that the CPU sits idle while the GPU finishes its work, I guess.

Luigi[/QUOTE]

No, this won't work because I was lazy. I'm using cudaStreamSynchronize() to wait for the GPU. This function is very easy to use but causes 100% CPU load (on one core).

Oliver

[QUOTE=TheJudger;214363]No, this won't work because I was lazy. I'm using cudaStreamSynchronize() to wait for the GPU. This function is very easy to use but causes 100% CPU load (on one core).

Oliver[/QUOTE]

I guess that's not a semaphore...

Luigi