New PC build

[philbo0042]

Thank you for all of the responses. You all are quick to respond and awesome with your details. I had considered a high-end video card but I am so unfamiliar with using them for a purpose like this that it scared me off, however I will now reconsider this option.

I will take the time to check the threads referenced above and go from there.

Thanks for all of the advice!

[philbo0042]

Based on the advice given here, I went ahead and purchased a Radeon VII linked to on the previous page. It was $130 off on Amazon.

https://www.amazon.com/XFX-Radeon-16...323943&sr=8-2#

[nomead]

Originally Posted by M344587487

That's true, in some markets the 1660 could be competitive. Looks like the 1660 Super is the latest refresh that sits between the 580 and Vega 56 in my list, at least here the non-super 1660 is not worth getting: https://youtu.be/Y1Rr1e99eBQ?t=803

It's still good "bang for the buck" in trial factoring, I thought I specifically stated that. See https://www.mersenne.ca/mfaktc.php but I belive you already know about that benchmark list. See the JVR column values for a rough comparison with what I believe are US market prices. The relative prices are wildly different here in the northern reaches of Europe, so... 240 euros for GTX 1660 (non Super) and about 1300 GHz-d/day, or 200 euros for RX 580 and about 580 GHz-d/day, or, well, the Vega 56 isn't even sold as new here anymore.

But in general gaming use your mileage will vary, a lot.

[philbo0042]

Thanks Nomead. I missed that link on my first glance. I will re-evaluate my decision. I can always return the Radeon at no charge.

I really appreciate all of the advice I have received here!

[nomead]

Originally Posted by philbo0042

Thanks Nomead. I missed that link on my first glance. I will re-evaluate my decision. I can always return the Radeon at no charge.

I really appreciate all of the advice I have received here!

It is not a problem. The Radeon VII is the greatest value (in fact, currently, the greatest card) for primality testing. That second link is for factoring only (mfaktc / mfakto). Again, for complicated reasons, different hardware is better for different purposes. There is really no affordable solution for "all-round" performance, you need to choose what you want to do and plan accordingly

[Prime95]

Originally Posted by philbo0042

Based on the advice given here, I went ahead and purchased a Radeon VII linked to on the previous page. It was $130 off on Amazon.

https://www.amazon.com/XFX-Radeon-16...323943&sr=8-2#

Congratulations!

This thread is your friend: https://www.mersenneforum.org/showthread.php?t=23982
It is mostly Linux instructions/observations but much applies to Windows as well.

Best throughput comes from running 2 gpuowl instances on the 1 GPU. For a 5M FFT (exponents in the low 90 millions) each gpuowl is doing ~1.7 ms./iter. consuming about 185 watts. That comes to about 2 PRP tests every 43 hours.

[philbo0042]

Originally Posted by Prime95

Congratulations!

This thread is your friend: https://www.mersenneforum.org/showthread.php?t=23982
It is mostly Linux instructions/observations but much applies to Windows as well.

Best throughput comes from running 2 gpuowl instances on the 1 GPU. For a 5M FFT (exponents in the low 90 millions) each gpuowl is doing ~1.7 ms./iter. consuming about 185 watts. That comes to about 2 PRP tests every 43 hours.

Thank you! That answers two questions I was researching, and I was going to ask which thread to read.

Could you provide an estimate for a 18M FFT?

[kriesel]

Originally Posted by philbo0042

Thank you! That answers two questions I was researching, and I was going to ask which thread to read.

Could you provide an estimate for a 18M FFT?

In the absence of other data, a rough estimate based on typical scaling observed is time is proportional to exponent to the 2.1 power, so around 2 exponents requiring 18M fft in 633 hours (3.8 weeks) on the Radeon VII.
43 * (18/5)^2.1. https://www.mersenneforum.org/showpo...6&postcount=10 This scaling emerges from the fft method of doing huge multiplications scaling as n log n log log n, and the number of iterations being essentially proportional to exponent, independent of whether it's PRP, LL, or P-1 being performed, or a gpu or cpu doing the computation. Of course the multiplier in front of that varies a lot with the hardware and computation type. This estimate isn't quite right, because fft length grows a little faster than exponent, but it will get you a number in the rough neighborhood, and is a bit conservative. I think the scaling is fft^1.9; that gives ~490 hours; 20.4 days.

[philbo0042]

Thank you Kriesel! It is awesome that you took the time to write all of that out for me.

[kriesel]

Originally Posted by philbo0042

Thank you Kriesel! It is awesome that you took the time to write all of that out for me.

You're welcome. I often learn from the questions of others. Writing the message you reacted to was some of the process of thinking it through.

A lot of what's at https://www.mersenneforum.org/showth...922#post521922 came from doing that. Often when someone asks or comments, I realize a piece of it is missing and there's another identified gap in my understanding.