mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   GPU Computing (https://www.mersenneforum.org/forumdisplay.php?f=92)
-   -   GPU questions (https://www.mersenneforum.org/showthread.php?t=16789)

c10ck3r 2012-05-08 01:48

GPU questions
 
Howdy!
So, I have a GTX 460 that is currently running mfaktc 0.18 (2 instances), which is being fed by a 3.2Ghz single-core Pent D 940 w/ 2G DDR2 RAM.
My first question is: for GPU72 exponents (currently 56M 71-72 bits), will there be a noticeable difference between 420 and 4620 classes?
My second question is: for GPU72 exponents (and similarly larger exponents), would P-1 be feasible to implement on a GPU, even if only on a partial scale. I guess I'll break this down into two parts: for B1/B2 at the normal bounds (or a new multiplier based on GPU performance), or for B1=B2 (ie no stage 2).
My third question (sorry, they just keep coming :P ) is whether, if (2) is feasible, is anyone currently trying/successfully implementing this?
Thanks!
Johann

Dubslow 2012-05-08 02:48

[QUOTE=c10ck3r;298738]Howdy!
So, I have a GTX 460 that is currently running mfaktc 0.18 (2 instances), which is being fed by a 3.2Ghz single-core Pent D 940 w/ 2G DDR2 RAM.
My first question is: for GPU72 exponents (currently 56M 71-72 bits), will there be a noticeable difference between 420 and 4620 classes?
My second question is: for GPU72 exponents (and similarly larger exponents), would P-1 be feasible to implement on a GPU, even if only on a partial scale. I guess I'll break this down into two parts: for B1/B2 at the normal bounds (or a new multiplier based on GPU performance), or for B1=B2 (ie no stage 2).
My third question (sorry, they just keep coming :P ) is whether, if (2) is feasible, is anyone currently trying/successfully implementing this?
Thanks!
Johann[/QUOTE]
1) No, and in fact running multiple instances on just one core probably won't help anything. You should definitely be using more classes though, while not a major difference, it is still enough of a difference to choose more classes over less.

2) I'm not entirely sure, but it would require similar FFT multiplication-of-numbers to the LL test, just the actual calculations are a bit different.

3) Again, not sure, but I don't think anybody's working on it. LaurV had a small experiment, but I don't think he's gotten anywhere with that.


All times are UTC. The time now is 11:29.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.