Max FFT Size?

[davieddy]

Originally Posted by Xyzzy

Too many notes!

Too many posts to cross

[aketilander]

Originally Posted by Prime95

In v27.6 max exponent is down to 595.8M. However, I think if you specify the FFT length to use in worktodo.txt, then the larger exponent might be accepted.

OK I see what you mean.

Source code Prime95 version 27.6 from gwnum.h:

#define MAX_PRIME 79300000L /* Maximum number of x87 bits */
#define MAX_PRIME_SSE2 595800000L /* SSE2 bit limit */
#define MAX_PRIME_AVX 595800000L /* AVX bit limit */
#define MAX_FFTLEN 4194304L /* 4M FFT max for x87 */
#define MAX_FFTLEN_SSE2 33554432L /* 32M FFT max for SSE2 */
#define MAX_FFTLEN_AVX 33554432L /* 32M FFT max for AVX */

So the largest possible candidate to LL with this version will be M595799947. So I think I'll keep version 27.4 until I'm finished with M595999993 if I decide to test it.

[Madpoo]

Originally Posted by aketilander

...
So the largest possible candidate to LL with this version will be M595799947. So I think I'll keep version 27.4 until I'm finished with M595999993 if I decide to test it.

I'd like a larger FFT size option in the next 28.x build, just so I can do a DC on Ake's result of M595999993 (when he finally turned it in, it had some errors that marked it suspect). It's currently the largest result but I have no way to run a verification.

[0PolarBearsHere]

Originally Posted by Madpoo

I'd like a larger FFT size option in the next 28.x build, just so I can do a DC on Ake's result of M595999993 (when he finally turned it in, it had some errors that marked it suspect). It's currently the largest result but I have no way to run a verification.

I just started it on 27.4 to see how long it would take. The estimate was May 2018 on my 3930k hahaha (I stopped it). Why can't I get access to a meteorological super computer when I need one :P

[Madpoo]

Originally Posted by 0PolarBearsHere

I just started it on 27.4 to see how long it would take. The estimate was May 2018 on my 3930k hahaha (I stopped it). Why can't I get access to a meteorological super computer when I need one :P

I was tempted to do the same, but then I don't think 27.4 had the AVX support and that would just be dumb to forgo that nicety.

[0PolarBearsHere]

Originally Posted by Madpoo

I was tempted to do the same, but then I don't think 27.4 had the AVX support and that would just be dumb to forgo that nicety.

Here's the starting message
"Resuming primality test of 595999993 using AVX Core2 type-3 FFT length 32M, Pass1=2K, Pass2=16K, 10 threads"

[Madpoo]

Originally Posted by 0PolarBearsHere

Here's the starting message
"Resuming primality test of 595999993 using AVX Core2 type-3 FFT length 32M, Pass1=2K, Pass2=16K, 10 threads"

Oh, I probably meant the AVX2/FMA code. Doh!

Then again, I had issues with 28.7 and the 32M FFT size on the largest possible exponent for that version... throws instant round-off errors on all the systems I attempted it with.

Of course that could have been specifically related to the AVX2 code... perhaps AVX would do better? May have to do some experimenting with disabling FMA to see if it behaves, at the expense of worse performance.

[Prime95]

Originally Posted by Madpoo

I'd like a larger FFT size option in the next 28.x build, just so I can do a DC on Ake's result of M595999993 (when he finally turned it in, it had some errors that marked it suspect). It's currently the largest result but I have no way to run a verification.

I think if you put FFT2=32M as part of the worktodo.txt entry then v28.x will test exponents as high as 595999993.

[aketilander]

Oh, since this is the day of whishes, what about a FFT for LL of the ("smallest") billion digit Mersennes? It would be so nice if it could be included in the next version of Prime95. It is of course not necessary to have FFTs for all LLs in between.

[Madpoo]

Well, my mission of making sure no exponents are self-verified has hit a roadblock today.

M601248421

I can't test an exponent that size using v28.x with the max FFT of 32M

I'm not sure what NOoE had for this particular exponent but he sure seemed to go after it with gusto... doing a lot of TF and P-1 on it before throwing a pair of simultaneous LL tests at it.

Only one assignment for it (back in 2014-04-18) but running on two different systems with different shift-counts.

Looks like I'll be leaving that one as-is for now. Even if a future version of Prime95 supported larger FFT, I might leave that one alone anyway because there are better things to be doing with a system for however many months it would take to do a triple-check of that guy.

I'm not sure how this was run... the reporting app version (for both) was:
Windows64,Prime95,v28.5,build 1

And the CPU it ran on was (also for both runs):
Intel(R) Core(TM) i7-4810MQ CPU @ 2.80GHz

The CPU may have been (almost definitely) upgraded since it was first assigned and started nearly 2 years ago.

One problem I have with them is that even though both reported the same 64-bit residue and had different shift counts, both of them reported an identical error code of "7F00FF00"

What are the odds that two runs on two CPUs using an app version that I wasn't aware could even handle an exponent that large would report the exact same error code?

I know there are some who are suspicious of NOoE's results and this one in particular is eyebrow raising for the reasons mentioned above.

I'm sure the reason for running two at the same time was to compare residues along the way, and they were turned in within 12 hours of each other.

If I could, I'd ask NOoE how often he had to start the runs over again from a previous matching iteration because one or the other mismatched. The error codes involved sure make it seem like they ran into significant problems along the way. And if it someone managed to run that at the max 32M FFT size, I'd say it was well past the upper limit of what that FFT size could realistically handle... is that safe to say?

[Batalov]

Originally Posted by Madpoo

I know there are some who are suspicious of NOoE's results and this one in particular is eyebrow raising for the reasons mentioned above.

I've shown a white-hat demonstration years ago how to submit two results (for the price of one calculation), always. It is nearly trivial. With different shifts and with different error codes (if one wants ) ...or without errors. With or without actual work done even once. Using completely unmodified Prime95/mprpime, so the result is de facto trusted. The entire hack is done with savefiles.