mfaktc: a CUDA program for Mersenne prefactoring

[James Heinrich]

Originally Posted by TheJudger

Do we need this as an option or should we enable this allways?

My preference would be that it's always there.

[chalsall]

Since we're asking for new features / abilities...

It would be nice (if it's not already there) to have a function like Prime95's worktodo.add feature. This would allow for automated reservation and processing of work from PrimeNet and/or G72 by way of another spider until Christenson has completed his integrated functionality.

I understand from another thread that the mfakt* programs allow for editing live worktodo.txt files (except for the first line), but I personally would prefer not to have a 'bot touch it.

[sonjohan]

Isn't GPU trial factoring unfair for GHzday-statistics?
I'm running it on my GTX 570M, for less than a month; and I already have doubled my GHz-days.

It's also the reason I shut down my old desktop. (Which only did TF.)
If that desktop needs 3 days to TF to 2^69, where the GPU does up to 2^72 in 12h for the same exponent, I don't see the point in wasting more electricity on an old PC. (Never mind the "you might be addicted to GIMPS if..." thread.)

[kladner]

Originally Posted by sonjohan

Isn't GPU trial factoring unfair for GHzday-statistics?
It's also the reason I shut down my old desktop. (Which only did TF.)
If that desktop needs 3 days to TF to 2^69, where the GPU does up to 2^72 in 12h for the same exponent, I don't see the point in wasting more electricity on an old PC. (Never mind the "you might be addicted to GIMPS if..." thread.)

That old machine would probably still make a fine contribution doing DC or P-1 work.

Don't knock GPU TF. Hardware advances will always seem unfair to the practitioners of previous technology. If there were similar comparisons available, I'm sure that 486's seemed unfairly fast to those with 386's.

[flashjh]

Originally Posted by kladner

That old machine would probably still make a fine contribution doing DC or P-1 work.

Don't knock GPU TF. Hardware advances will always seem unfair to the practitioners of previous technology. If there were similar comparisons available, I'm sure that 486's seemed unfairly fast to those with 386's.

Agreed! I have an older socket 478 P4 3.4GHz 2Gb system that I've had for many years now (upgraded to that slowly over the years). Civ V played ok on it, but when you get too far along it would lockup sometimes (my Son requested a better system ). So, I finally upgraded and then dedicated it to P-1 and to mfakto. It has an HD 4670 AGP card that gives about 12M/s (200,000 SievePrimes) (I'm pretty sure the card would do better in a faster system). It does ~four DC TF checks 68^69 per day and each P-1 takes ~3-4 days.

All in all nothing special, but it's more than nothing.

[Dubslow]

Originally Posted by sonjohan

Isn't GPU trial factoring unfair for GHzday-statistics?

Many (including myself) have made that argument, however figuring out what to actually do about it is what's given us trouble. I think we've all just kind of given up on it. (The main problem is in figuring out what the 'average' CPU and GPU working for GIMPS can do.)



As for that old PC, you can do DC or P-1 as others have mentioned, or you could even do TF-LMH. That's something that still yields a factor every couple of days, even on that slow a machine, and each factor is anywhere from 500 to many thousands of GHz-Days of LL work (admittedly far off in the future, but still.) In fact, at a minimum, the chalsall (GPU to 72 creator and admin) has his old PC's doing TF-LMH (obviously he knows better than I, but that's my understanding).

[bcp19]

Originally Posted by kladner

That old machine would probably still make a fine contribution doing DC or P-1 work.

Don't knock GPU TF. Hardware advances will always seem unfair to the practitioners of previous technology. If there were similar comparisons available, I'm sure that 486's seemed unfairly fast to those with 386's.

I had an 8088 laptop and had a program running to calculate how many entries would be needed to get every possible 4 number combination in a 51 number lottery. Program ran for somewhere around 6 months. Put it on my first Pentium and it took about 12 hours.

[Uncwilly]

Originally Posted by bcp19

I had an 8088 laptop and had a program running to calculate how many entries would be needed to get every possible 4 number combination in a 51 number lottery. Program ran for somewhere around 6 months. Put it on my first Pentium and it took about 12 hours.

(51 * 50 * 49 * 48) / 4! to 1
Your pocket 8 digit calculator can do that one. That is a super trivial calculation.

[LaurV]

Originally Posted by bcp19

I had an 8088 laptop and had a program running to calculate how many entries would be needed to get every possible 4 number combination in a 51 number lottery. Program ran for somewhere around 6 months. Put it on my first Pentium and it took about 12 hours.

This post is good for xyzzy's top. Sure that program was not doing something else too? There are only 249900 of them, and even a 8-bit Sinclair could spit them out in a blink, or use a matrix (needle) printer to print them in less then 12 hours...

[ckdo]

Both of you are, of course, completely wrong in case bcp19 didn't talk about a "4 out of 51" lottery in the first place.

[kjaget]

Originally Posted by James Heinrich

My preference would be that it's always there.

I'd like it as an option. I've run into cases where I have to break up results.txt when pasting it into the manual results page, presumably because it's too much data for one shot. Adding time stamps will make this happen quicker, meaning more chance for error when submitting results.

Once we get automated Primenet interaction it won't be a big deal.