mfaktc: a CUDA program for Mersenne prefactoring

[Mark Rose]

Got a 1% boost on a GT 430, too.

I tried tweaking the GPU* configuration parameters to see if their optimal values changed but they seem to be the same.

Also, I noticed this is still in the mfaktc.ini file at least in the Linux source download:

# GPU sieving is supported on GPUs with compute capability 2.0 or higher.
# (e.g. Geforce 400 series or newer)

[MatWur-S530113]

Thank you!
Installed on win7-64 bit with GTX760
both selftest were passed
~2% more output (from 266 -> 271)
(I changed 2 settings in the ini-file: GPUSieveSize to max 128
and GPUSieveProcessSize to min 8,
~10% more performance with my GTX760)

greetings

Matthias

[flashjh]

Originally Posted by TheJudger

Special thanks goes to Jerry Hallett who spent a lot time on building and testing Windows binaries as well as some hints for Windows compability. Thank you Jerry!

You are welcome, my pleasure!

Awesome work Oliver!

[Mark Rose]

Originally Posted by MatWur-S530113

Thank you!
Installed on win7-64 bit with GTX760
both selftest were passed
~2% more output (from 266 -> 271)
(I changed 2 settings in the ini-file: GPUSieveSize to max 128
and GPUSieveProcessSize to min 8,
~10% more performance with my GTX760)

greetings

Matthias

I've found that my GTX 760 also benefits from a slightly higher GPUSievePrimes value than default. I have mine set to 100000.

[MatWur-S530113]

Originally Posted by Mark Rose

I've found that my GTX 760 also benefits from a slightly higher GPUSievePrimes value than default. I have mine set to 100000.

I quickly testet some values for GPUSievePrimes (k*10000 for k=5..13, only a few classes per test),
but it seems that for my card the default-value of 82486 is indeed the best,
Between 70000 and 100000 no measurable difference, larger differences to the default-value
are resulting in a worse performance (but still not remarkable, maybe 1%)
Very large and very low values for GPUSievePrimes slow down my card up to 50%.

[Mark Rose]

Originally Posted by MatWur-S530113

I quickly testet some values for GPUSievePrimes (k*10000 for k=5..13, only a few classes per test),
but it seems that for my card the default-value of 82486 is indeed the best,
Between 70000 and 100000 no measurable difference, larger differences to the default-value
are resulting in a worse performance (but still not remarkable, maybe 1%)
Very large and very low values for GPUSievePrimes slow down my card up to 50%.

I'll fiddle with mine again to verify it's better than the default.

[TheMawn]

Originally Posted by Mark Rose

I've found that my GTX 760 also benefits from a slightly higher GPUSievePrimes value than default. I have mine set to 100000.

I've never really fiddled with the options but I believe you've done a bit more.

Which options do I want to be messing around with to try to increase throughput?

[Mark Rose]

Originally Posted by TheMawn

I've never really fiddled with the options but I believe you've done a bit more.

Which options do I want to be messing around with to try to increase throughput?

GPUSieveProcessSize
GPUSieveSize
GPUSievePrimes

I would do them in that order.

[TheMawn]

Originally Posted by Mark Rose

GPUSieveProcessSize
GPUSieveSize
GPUSievePrimes

I would do them in that order.

Do they need to be bigger or smaller, or does it tend to be a bit of both?

[Mark Rose]

Originally Posted by TheMawn

Do they need to be bigger or smaller, or does it tend to be a bit of both?

Depends on the card.

[TimSorbet]

Originally Posted by TheMawn

Do they need to be bigger or smaller, or does it tend to be a bit of both?

(for maximum performance, in my experience)

GPUSieveProcessSize: (probably) bigger
GPUSieveSize: (probably) bigger
GPUSievePrimes: just right

Experiment to find the best values for your GPU, Mersenne number size, and bit depth. If you use your GPU to drive a display, responsiveness becomes a concern, too, and dialing back the first two values may be necessary. I use a tool I wrote to automatically switch the GPUSieveSize between a few values to trade off responsiveness and speed, based on what I'm doing (when idle, 128; common usage, 32; most games, 8; some games, stopped).