[QUOTE]This will be another xubuntu box, this time with intel sandy bridge i7 on it, and I'm wondering if the 32 Gigs of Ram option is worth considering.[/QUOTE]What does the i7 bring to the table for GPU work that the i5 does not? (We have the i5 at 3.3GHz.) The i7 is $100 more.

We have 16GiB per box. Since we are trial factoring, what is the need beyond 2 × 1GiB modules? (Altough 2 × 8GiB is under $90 right now.)

Sorry for the dumb questions!

:max:

For mfaktc, cpu minimum requirements go a little like this:
460/560: 3.5-4.5GHz
470/570: 8-9GHz
480/580: 9-10GHz

GHz is based on corei5/i7. If you have earlier generation then probably scale up a little bit on the GHz. I haven't seen any figures for AMD generation CPUs.

This is a *rough* guide.

Based on my experiences and other posts here.

This is to max out the card. Yes 10GHz CPUs don't exist, but add the cores up. i.e. 3x instances on separate real 3.5GHz cores would be enough to max out a 580. A hyper threaded core is not enough to drive an instance.

For expected results:
460/560: 100-130GHz.days/day
470/570: 200-240GHz.days/day
480/580: 250-280GHz.days/day

Again, rough figures guide only. These figures are for high bit depth too. If you aren't getting anywhere near these figures - check you have enough cores/instances running (check min CPU spec above)

As for power consumption check nvidia's web page. They do list recommended PSU specs.

As for headless.....on windows PCs it doesn't seem to work for me. I can't RDP to a windows box and kick off the process. It fails on cudadevicecheck(). I'm guessing this is more driver related. *But* I can pull the monitor out, and RDP to it at a later time to check on it.

As for GUI corruption, generally the models listed in this post haven't reported screen corruption. There have been some posts with screen corruption. In my experience the GUI has been quite workable (I'm even able to play high end games like Starcraft2 while it's running). I run Win7 though.

-- Craig

As for memory requirements. It hasn't been mentioned much as it's minuscule. I'm running 2x instances and they use 100MiB. That's nothing on the machines you're talking about.

The only gotcha I'd list - I wouldn't run it off a flash key or an SSD, unless you turn off checkpoint saves. Looking at the save rate - it's too high for my liking for flash based devices.

I run it off a ramdisk.

I don't expect anyone to agree with me - but that's just me. :)

-- Craig

[QUOTE=Xyzzy;258655]What does the i7 bring to the table for GPU work that the i5 does not? (We have the i5 at 3.3GHz.) The i7 is $100 more.

We have 16GiB per box. Since we are trial factoring, what is the need beyond 2 × 1GiB modules? (Altough 2 × 8GiB is under $90 right now.)

Sorry for the dumb questions!

:max:[/QUOTE]

I might be pushing things beyond optimal...but the judger said I'd be able to get some standard Prime95 work out of a box alongside mfaktc on the next release, so I'm thinking of a really nice mobo and CPU.

There's also that attack on block Lanczos I'm making, and it's going to require a lot of memory to work. (Basically, it's at the vague idea stage -- I'm wondering if it would be possible to do the post-processing of semi-sparse sieve data for factoring incrementally -- and no, i haven't read the literature yet, so RDS be quiet, that's what I'm doing right now; I'm trying to understand the MP part of MPQS).

Finally, this is my toy....and I can afford to spend a bit more than the minimum, so here goes...as soon as i do my homework.

[QUOTE=Prime95;258648]Alright, I have a working theory. Does this sound possible:

1) You submitted the results using the manual form but forgot to login to the web pages.
2) Maybe because you noticed the exponents were still assigned to you, you resubmitted them after logging in.
3) The server didn't record the duplicate results, but since you had a valid assignment ID, it gave you the CPU credit. This whole credit for duplicate results mess was added to avoid outrage about losing CPU credit because someone poached an assignment.

If this scenario is correct, then the "bug" is that you didn't get a "success" for finding an already known factor with your valid assignment ID.[/QUOTE]It may well be possible --- I don't remember exactly what I did in detail.

Although the performance figures are interesting, that wasn't the primary reason for asking about the incident. I wanted to ensure that the factors were registered so that the LL tests wouldn't be attempted. Each factor took less than 30 minutes to find but a LL test would take [b]much[/b] longer than that.

The second batch of 100 exponents has just finished. I''ll upload them and get some more.

A genuine error I made was to fail to check the TF box on the manual registration page and was assigned a couple of 1st time LLs instead. Can I unreserve them through the web page or should I just let them expire naturally? I couldn't find an obvious "unreserve these numbers" facility.


Paul

[QUOTE=xilman;258674]A genuine error I made was to fail to check the TF box on the manual registration page and was assigned a couple of 1st time LLs instead. Can I unreserve them through the web page or should I just let them expire naturally? I couldn't find an obvious "unreserve these numbers" facility.


Paul[/QUOTE]
It should be possible to unreserve your assignments via My account/Assignments:

[URL]http://www.mersenne.org/workload/[/URL]

Just check the assignments you want to unreserve (the LL tasks) and the checkbox near the "Unreserve checked exponents" button at the bottom of the list and then click the button. This should unreserve the selected exponents.

Hi!

[QUOTE=nucleon;258659]For mfaktc, cpu minimum requirements go a little like this:
460/560: 3.5-4.5GHz
470/570: 8-9GHz
480/580: 9-10GHz[/QUOTE]

For mfaktc: 460 < 560 < 470 < [B]480 < 570[/B] < 580
If the GPUs have the same compute capability you can just multiply the number of cores with their clockrate and compare those numbers, bigger is better. :wink:

[QUOTE=nucleon;258659]GHz is based on corei5/i7. If you have earlier generation then probably scale up a little bit on the GHz. I haven't seen any figures for AMD generation CPUs. [/QUOTE]

Last time I've checked the sieve code it runs [B]minimal faster per clock[/B] on Core 2 Duo/Quad compared to the first i3/5/7 generation. I don't have exact numbers for the 2nd generation i3/5/7 (2xxx series) or AMD. :sad:


[QUOTE=Christenson;258670]I might be pushing things beyond optimal...but the judger said I'd be able to get some standard Prime95 work out of a box alongside mfaktc on the next release, so I'm thinking of a really nice mobo and CPU.[/QUOTE]

mfaktc and prime95 on the same core will only work on slow GPUs when one core is more than enough to feed the GPU. Sieving is still done on CPU but I've replaced the busywaits with some sleeps.

Oliver

[QUOTE=Ralf Recker;258676]It should be possible to unreserve your assignments via My account/Assignments:

[URL]http://www.mersenne.org/workload/[/URL]

Just check the assignments you want to unreserve (the LL tasks) and the checkbox near the "Unreserve checked exponents" button at the bottom of the list and then click the button. This should unreserve the selected exponents.[/QUOTE]Thanks, that did the job nicely.

Slow GPU is relative to CPU speed...and the other types of work that might be done on the GPU might need less CPU than mfaktc...so I still want my toy i7 sandy bridge screamer....

[QUOTE=TheJudger;258677]
mfaktc and prime95 on the same core will only work on slow GPUs when one core is more than enough to feed the GPU. Sieving is still done on CPU but I've replaced the busywaits with some sleeps.
[/QUOTE]

As I own this type of slow GPU, is there already some pre-version I could test? Can you share some details how you calculate the sleeps?

I currently bind mfactc to a specific CPU and lower its prio to match the (increased) prio of prime95, so that each get around 50% of this one CPU.

[QUOTE=Bdot;258757]As I own this type of slow GPU, is there already some pre-version I could test? Can you share some details how you calculate the sleeps?

I currently bind mfactc to a specific CPU and lower its prio to match the (increased) prio of prime95, so that each get around 50% of this one CPU.[/QUOTE]

Bdot isn't alone; my Galaxy GEForce210 is in that class....I could run it during the work day if I could get adequate performance out of the rest of the system.