[QUOTE=CRGreathouse;463046]Some quotes:

and

Clarifications in brackets are mine.[/QUOTE]

I don't think bilirubin* (er, I mean Billy Rubin) has quite gotten over the criticism he faced last time.

He also seems like it's a perfectly normal thing for him to blindly download and run things found on the internet, trusting in the almighty AV software to protect him... good luck with all that. Trust is earned, and from his attitude, you can judge for yourselves how much trust he's earned. :smile:

[QUOTE]So why the flame-fest? I guess Mersenne Research, Inc. must really be worried that CEMPLLA really is a threat to what it has always imagined to be it's monopoly on the EFF awards.

Well gee, sorry, not a big fan of monopolistic corporations. I guess they'll just have to make do with their "Mlucas" or whatever, for the next, what is it? Oh yeah. "850 years". Yeah. That's the consensus right? "850 years" to eliminate a single billion decimal digit Mersenne Prime candidate? Yeah, I can read..[/QUOTE]

The evil Mersenne Research corporation and it's horrible monopoly... sorry Ernst and others who have non-Prime95 software that also competes for the prize.

I kind of feel like he's either really naive or a troll (or both?)

* "Bilirubin is a brownish yellow substance found in bile"

Addendum: I know he feels persecuted because he failed to demonstrate any timings of his software or proved that it generated valid results... by way of comparison, I remember just recently when gpuOwl came here with improvements ... everything was above board, open source, he freely discussed what was going on and came across as very knowledgeable. Contrast that with CEMPLLA and it's locked down, no-scientific-method-allowed approach. Even if he just released a few benchmarks like "using this graphics card I can check M74xxxxxx using a 4M FFT in xx hours" and also have that residue match what any other software does so it can be verified. He can't even give us that much so of course we're skeptical. It'd be the height of stupidity to accept his claims at face value. When he asks us to accept it at face value with no proof whatsoever, I feel like he's calling us all stupid. :smile: His only answer is "trust me... run it and see for yourself". That's right up there with strangers in windowless vans trying to pass out candy to kids... "trust me..." LOL

[QUOTE=Madpoo;463099]The evil Mersenne Research corporation and it's horrible monopoly... sorry Ernst and others who have non-Prime95 software that also competes for the prize.[/QUOTE]

I might be wounded, were I not the evil monopolist he claims me to be, and thus immune to such lèse-majesté from the deplorable hoi polloi.

Based on my timings on David Stanfill's 32-core Xeon which I am using to complete one of the two independent-data runs of F29, a gigadigit - not to be confused with a gigabit - Mersenne would need ~20 years, so he's only off my a factor of 40-50. Even more laughable because he provides no timings of his own code, if it even deserves to be called that, based on the amount of other-people's-work which he appears to have shamelessly appropriated.

[QUOTE=ewmayer;463106]...he provides no timings of his own code, if it even deserves to be called that, based on the amount of other-people's-work which he appears to have shamelessly appropriated.[/QUOTE]

Therein is probably the reason he doesn't provide source, because we'd see bits and pieces of other programs in there (assuming it's a legit program at all).

Which begs the question of whether someone filing a claim for the EFF prize would need to disclose their source code... I don't known if their rules specify or not.

[QUOTE=Madpoo;463108]Therein is probably the reason he doesn't provide source, because we'd see bits and pieces of other programs in there (assuming it's a legit program at all).

Which begs the question of whether someone filing a claim for the EFF prize would need to disclose their source code... I don't known if their rules specify or not.[/QUOTE]

Thought experiment:

What if the method of finding the prime didn't involve a computer with source code? (not saying that's going to happen anytime soon)

Let's say you found some alien script that listed the first 1000 mersenne primes. All the known ones match exactly what's on the list, but of course the list continues on...

You use Prime95 to verify the next two exponents which are unknown to the rest of the world. Now you know you're aren't getting trolled by someone (from Earth).

You go and announce the smallest of them that isn't already known. The first thing that will happen is that depending on what credibility you have you're gonna be called an idiot. But [I]someone[/I] on this forum is gonna try to TF it and fail. And since the exponent is small, someone will also try to LL it. Once it turns up prime, you're gonna get blamed for using GIMPS under the cover to find it without proper attribution.

Then you announce a 50 million digit prime - rinse and repeat. Then you announce a 100 million digit prime. (a few months to LL?) Then people will start to take you seriously.

[QUOTE=Madpoo;463108]Which begs the question of whether someone filing a claim for the EFF prize would need to disclose their source code... I don't known if their rules specify or not.[/QUOTE] they need to allow for EFF to do that it seems see sections 3 of the following:

[QUOTE="https://www.eff.org/awards/coop/rules"]those receiving EFF Computation Awards must allow EFF to freely publish their methods, algorithms, [U][I][B]source code[/B][/I][/U]...[/QUOTE]

edit: which of course, means the OP didn't read the rules to begin with. otherwise they wouldn't state the source code would only be published once both were found ( quoted from the site, by chalsall, in post 6) ...

[QUOTE=Madpoo;463108]Therein is probably the reason he doesn't provide source, because we'd see bits and pieces of other programs in there (assuming it's a legit program at all).

Which begs the question of whether someone filing a claim for the EFF prize would need to disclose their source code... I don't known if their rules specify or not.[/QUOTE]

The [url=https://www.eff.org/awards/coop/rules]EFF prize rules[/url] are very clear on the subject of "Full Disclosure is Required".

[QUOTE=Mysticial;463109]Thought experiment:

What if the method of finding the prime didn't involve a computer with source code? (not saying that's going to happen anytime soon)

Let's say you found some alien script that listed the first 1000 mersenne primes. All the known ones match exactly what's on the list, but of course the list continues on...

You use Prime95 to verify the next two exponents which are unknown to the rest of the world. Now you know you're aren't getting trolled by someone (from Earth).[/QUOTE]

I don't see a problem with that, since if it was testable by you using prize-eligible software, it is by definition independently verifiable by others using same. Your wild story of how-I-got-the-exponents is immaterial as far as prize eligibility goes, and would of course seem rather more plausible were you to produce an actual new prime from your list.

Update: I also just tried a few sample timings of the smallest billion-digit M-number on an 8-core Ryzen system - here is the 1000-iter timing:
[code]NTHREADS = 8
M3321928097: using FFT length 196608K = 201326592 8-byte floats.
this gives an average 16.500195349256199 bits per digit
Using complex FFT radices 768 16 16 16 32
mers_mod_square: Complex-roots arrays have 8192, 12288 elements.
Mers_mod_square: Init threadpool of 8 threads
radix16_dif_dit_pass pfetch_dist = 4096
radix16_wrapper_square: pfetch_dist = 4096
Using 8 threads in carry step

1000 iterations of M3321928097 with FFT length 201326592 = 196608 K
Res64: 39E25DF480D46237. AvgMaxErr = 0.158501615. MaxErr = 0.187500000. Program: E17.0
Res mod 2^36 = 19341271607
Res mod 2^35 - 1 = 10199050979
Res mod 2^36 - 1 = 66182425439
Clocks = 00:09:17.186[/code]
The 100-iter timing clocks in at almost exactly 60 sec, so subtracting and dividing by 900 we get a per-iteration time free of startup-overhead of 0.55 sec, thus even on a modestly priced system like this we need 'only' 58 years to test such a behemoth.

From the Nvidia thread:

[QUOTE]But any timings that I've performed would be meaningless to publish, because my GPU hardware is not what CEMPLLA requires. In fact, the very last time I ran CEMPLLA on my machine, it very correctly erred out after ten minutes with the message "At this rate, you'll never get a result, so proceeding is pointless." (which is by design)..[/QUOTE]

He forcefully avoids ever mentioning any real numbers. It's to the point of absurdity.

Does anyone recall if he ever mentioned, even in generic terms, what algorithm he's using? I assume it's LLR unless he's discovered some new totally new method?

What he has also never made clear is how he's solved the bandwidth problem that we're all too familiar with here. According to him, adding more GPUs increases the parallel computing power, which would technically be true, but all those threads working in parallel at some point have to collect the data from each iteration and finish the aggregate result before it can continue. For even a smallish FFT (like 4M) that's a lot of data, and he's talking about billion-digit numbers with a...what... 200M FFT size (did I estimate that right)?

What are the memory requirements of the GPU when attempting such a feat anyway?

All I know is, there are *so* many unanswered questions, and his attitude when anyone asks him something specific is to immediately become defensive, and that raises all kinds of alarm bells.

I installed and ran his last program (in a silo'd virtual machine...didn't have a GPU) just so I could read the documentation, such as it was. I didn't see it doing anything particularly scary, but then again as mentioned I didn't have a CPU so...

Best I could guess, this guy is running some mining software and duping people into running it on their GPUs and he collects the profits. Since I wasn't running it on a system with a GPU it simply didn't kick into it's stealth mode? Who knows.

[QUOTE=Madpoo;463133]From the Nvidia thread:[/QUOTE]

From the first nvidia thread:

Question was asked:
[QUOTE]What is your estimate for finding the numbers you're looking for using e.g. a single higher end nVidia card? (say, a GTX 980 Ti) Did you make a rough computation how many compute hours are expected to hit the jackpot?[/QUOTE]

Answer was (not) given:
[QUOTE]As for computing how long it will take to "hit the jackpot", that depends entirely on how many people participate, how many NVidia GPU boards each of them have, how often they run it, how fast the various and sundry NVidia GPU boards really are, how many participants own or have access to a CUDA 'supercomputer', etc., etc.. I've run lots of numbers, but it's still impossible to predict exactly when either of the primes will be discovered, or how long it might take. For all I know, someone with a fast CUDA supercomputer might discover both primes in as many weeks. That's entirely possible.[/QUOTE]

Giving non-answers is his MO.

[QUOTE=axn;463134]Giving non-answers is his MO.[/QUOTE]

Yeah... I mean, it wouldn't take a programming genius to phony up a display that shows random "residues" and counts how many "iterations" it has gone through. Even magically getting [B]faster[/B] as the size of the exponent grows (which is another of his more remarkable claims).

I mean, we all know that the larger the FFT, the slower LLR will be per iteration. And the software out there walks that fine line of finding the smallest possible FFT size without getting into the roundoff errors.

This is all common knowledge for anyone who has spent a little bit of time here on the forums, so either this guy is on crack when he thinks we'd all fall for his snake oil... or... he's a visitor from the future and really does have some future tech or undiscovered algorithms that will blow us all away, but if he revealed it to us it would have a ripple effect and blow up the future world he came from.

I personally put more stock in the "visitor from the future" theory than the theory that he got the next few primes from an alien civilization, but only just barely. :smile:

EDIT: In another thread in the Nvidia accel computing forum, the same guy says:
[QUOTE]If so, that's a relief to me, because I just released software that uses it (both mad.wide.u32 & mul.wide.u32) quite extensively, and it is computationally sensitive to even the smallest of inefficiencies..[/QUOTE]
[URL]https://devtalk.nvidia.com/default/topic/1017754/long-integer-multiplication-mul-wide-u64-and-mul-wide-u128/#5183575[/URL]

I don't know if that gives anyone more hints about how he may be accomplishing such amazing feats of mathyness. :smile:

[QUOTE=axn;463134]Giving non-answers is his MO.[/QUOTE]

How about the rather more narrow question "what is the shortest time your code needs to test a 1-billion-digit Mersenne number on any hardware likely to be available to your intended user base?"

I got no problem answering that:

8-core AMD Ryzen: 58 years
64-core Intel KNL: 20 years

The KNL (running avx-512 assembly) is virtually identical to the 32-core Xeon (faster clock speed, but running 'just' avx2) at 192M FFT for my code, but at < $5000 is the far more cost-effective of the two 20-year options. (Assuming one were so keen to get cracking on such a test that 20-years-one0current hardware were not a deterrent.) The Ryzen is more cost-effective, still, but with a correspondingly longer ETA. :)

I suspect Jerry Hallett's CUDA code running on a high-end GPU can beat my KNL times - if it doesn't currently permit such large exponents perhaps he can fiddle it to do so, just for the grins.

[QUOTE=Madpoo;463135]I don't know if that gives anyone more hints about how he may be accomplishing such amazing feats of mathyness. :smile:[/QUOTE]
Those are integer instructions, so possibly he could be doing integer transforms.

[QUOTE=ewmayer;463136]I suspect Jerry Hallett's CUDA code running on a high-end GPU can beat my KNL times - if it doesn't currently permit such large exponents perhaps he can fiddle it to do so, just for the grins.[/QUOTE]

IIRC, cudaLucas uses nvidia's cuFFT library where max FFT size is "128 million elements" ([url]https://developer.nvidia.com/cufft[/url])
I think clLucas also has same kind of limitation (dependent on clFFT library from AMD?)
gpuOWL, OTOH, uses hand-rolled FFT, but currently only supports p-o-2 2M & 4M FFTs. But maybe the author can write a 256M FFT for s&g.
George should write a 192M one for s&g as well (if not already done).