Cuda Cores
 

[SIZE=2]My current video adapter is an nVidia GT-610. It has 48 Cuda cores. I just ordered a GTX-750 Ti. It has 640 Cuda cores. I will go out on a limb and say that this will make some difference in calculation speed. I've been running mfaktc. With the one I have now, a 72 to 73 factorization takes about 5.5 hours Comments anyone.[/SIZE]

[QUOTE=storm5510;445444][SIZE=2]My current video adapter is an nVidia GT-610. It has 48 Cuda cores. I just ordered a GTX-750 Ti. It has 640 Cuda cores. I will go out on a limb and say that this will make some difference in calculation speed. I've been running mfaktc. With the one I have now, a 72 to 73 factorization takes about 5.5 hours Comments anyone.[/SIZE][/QUOTE]

Probably, as long as mfaktc doesn't use FP64 math extensively.
The 750Ti is a Maxwell, and Nvidia cut back drastically on FP64 capabilities
of those boards.

[QUOTE=tServo;445445]Probably, as long as mfaktc doesn't use FP64 math extensively.
The 750Ti is a Maxwell, and Nvidia cut back drastically on FP64 capabilities
of those boards.[/QUOTE]
As far as I understand, mfaktc makes little or no use of FP64. There is a different story with CUDALucas.

[QUOTE=kladner;445447]As far as I understand, mfaktc makes little or no use of FP64. There is a different story with CUDALucas.[/QUOTE]

mfaktc uses a lot of 32x32 multiplies. IIRC, some (all?) nVidia use the FP64 unit to do these multiplies.

From [url]http://www.mersenne.ca/mfaktc.php[/url]

GT-610 produces 29 GHz day/day consuming 29 W for an efficiency score of 1 GHzd/d/w

750 Ti - 139.7 GHd/d, 60W, 2.3GHzd/d/w

So, about 5x performance, and 2x power consumption

EDIT:- FWIW, the low end GPU sweetspot for mfaktc/o is probably Radeon RX 460

My desktop is not new by any means so there is probably a limit as to what I can put in it. Replacing it has not been an option, as of yet. Perhaps soon. Thank you all for your replies. :smile:

[QUOTE=Prime95;445462]mfaktc uses a lot of 32x32 multiplies. IIRC, some (all?) nVidia use the FP64 unit to do these multiplies.[/QUOTE]
Thanks for the correction, George. I guess my comment was just an inference from TF work being less affected by the reduced FP capabilities of nVidia consumer cards above the 500 series.

[QUOTE]There is a different story with CUDALucas.[/QUOTE]Speaking of CUDALucas, This is something I would like to try at some point. I've only seen references to a Mac build. Nothing for Windows yet.

[QUOTE=Prime95;445462]mfaktc uses a lot of 32x32 multiplies. IIRC, some (all?) nVidia use the FP64 unit to do these multiplies.[/QUOTE]

You mean 32 bit INTs, correct?

[QUOTE=storm5510;445503]Speaking of CUDALucas, This is something I would like to try at some point. I've only seen references to a Mac build. Nothing for Windows yet.[/QUOTE]

Here are the binaries for many different versions of CUDA:
[url]http://mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.win.cuda80.zip[/url]

and here are the extra Wagstaff factoring binary and the Less Classes versions:
[url]http://mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.win.cuda80.extra-versions.zip[/url]

You need to download the CUDA library files for the version you selected in the "CUDA Libs" directory here:
[url]https://sourceforge.net/projects/cudalucas/files/[/url]

[QUOTE=tServo;445504]You mean 32 bit INTs, correct?[/QUOTE]

Yes, 32-bit integer multiplies yielding either the high or low 32 bits of the result.