GPU to 72 status...

[chalsall]

Originally Posted by lycorn

How old is your GPU?
That increased amount of noise, together with rising temperatures, may indicate some fan overhaul is needed. You may wish to check whether the fans need to be lubricated. Check the fan rpm, if you have the means, that is a good indicator.

Indeed.

We sometimes find ourselves in interesting places, don't we?

[davieddy]

Originally Posted by chalsall

The stated agenda of GPU to 72 73 74 is to help find the next Mersenne Prime as quickly as possible.

The best way to do this, in my opinion, is to TF as deeply as Makes Sense according to James' analysis of the cross over point, balanced with our available firepower.

Well that agenda is certainly what I am talking about. Funnily enough, that's the first time I've heard or read you say that.

James' crossover point is rendered irrelevant because we are firepower limited (especially since Pete went off in a huff of smoke).

However, there is a link between the two, dictated by these simple facts:

1) LL time goes as expo²
2) TF goes as expo^-1
3) Incrementing bit level doubles the time.

It follows that the interval from one bit level to the next occurs when the exponent increases by 2^1/3.

Now if (as I suggested) we could go to 74 bits for exponents>70M, what on earth is the slightest bit ridiculous in the suggestion that 75 bits will be appropriate above 90M?

THIS AIN'T ROCKET SCIENCE.

Try to adjust your intuition to the appropriate time scale, namely 4 years not 4 months.

David

[TheMawn]

Here is the breakdown.

Do we go bit-first, which means getting one exponent to the optimal bit before even starting the next exponent? Or, do we go breadth-first which means we factor a whole bunch of exponents to one bit, then the next, then the next, etc until we again reach the optimal point? I tend to agree with Chris on this one.

Let's assume we want to factor to 74 bits in the following explanation. Breadth-first SOUNDS like a great idea because if all of a sudden the LL wavefront catches up and the TF work is not sufficiently getting 74-bits work done, we have an entire batch of 73-bits work done, and it really isn't the end of the world if some 73-bits gets done instead of 74-bits. It looks like a very conservative approach.

Bit-first is actually a better idea because if we factor 10,000 exponents from 72 to 73 before then taking all of them to 74, we have spent enough work to instead take 3,333 exponents from 72 to 74. Every single amount of factoring work on an exponent we don't care about NOW is work that isn't useful to us NOW. It's useful to us later, for sure, but we will have more firepower later.

I was interested in working on taking all exponents below 1,000,000,000 to 66 from 65 for a while, until I realized that the LL front will effectively never reach that point and that all my work would be "wasted" as far as the present interests of the project are concerned.

If we put all our efforts into taking 60M-70M to 74 bits instead of doing all this stuff taking 80M to 72 bits or whatever work we're doing on future exponents, we would be leaps and bounds ahead of the LL-wave. It's just that we're doing TF work now that saves LL work in two or three years.

[sonjohan]

Originally Posted by davieddy

The supply of these is approaching exhaustion, and TF from 71 to 74 will not keep pace.

Whether or not it is true, your incessant nagging has chased away a few of the biggest contributors to the TF effort.
I estimate that we would've kept up the TF from 71 to 74 if the big guns were still there.

Could you estimate the pace we would've had if you had <expletive deleted> and left people alone, choosing to do whatever they want to do?
And yes, we can decide on what we work. Not you, not George, not Chris can decide: I have the possibility to shut down Prime95 or mfaktc whenever I want. Of course, it could be the case that you wish to chase people away, which is the reason you come trolling in this forum.

[chalsall]

Originally Posted by davieddy

In a month or so's time it will dawn on you that we need to get expos to 73 bits as a matter of urgency. I have been pointing this out via my broken record since the days before GPUto72 was created.

If I remember correctly, you make the same prediction a couple of months ago...

Originally Posted by davieddy

If only Chris hadn't refused to listen to me "because I didn't run a GPU".

I refused to listen to you because you didn't, and continue to not, make sense.

[chalsall]

Originally Posted by sonjohan

I estimate that we would've kept up the TF from 71 to 74 if the big guns were still there.

I appreciate your message, but please let me be clear...

GPU to 74 IS currently keeping up. All the way from 71. Even with the ebb and flow of participants.

And, there are automatic "safety valves" in place to release candidates already at 73 for LL assignment if we ever fall behind. Again, automatic; while I monitor things closely, I'm not needed in the loop for this to happen iff needed.

To cut to the chase, David has admitted in the past that his agenda in all this nagging is for him, personally, to be assigned low candidates because they take less time, and there's a (slightly) greater probably that the lower candidates are prime.

David could, of course, simply sign up to GPU72 and have access to the (small number of) low candidates we hold for LL assignments. Ironically, David was actually the very first beta tester of this service...

But, at the end of the day, the truth is the next Mersenne Prime is probably somewhere up in the high 60Ms or even low 70Ms range; all the work below has to be done by someone(s). Those of us who TF are simply helping those who LL to not waste their valuable time if it can be avoided.

[TheMawn]

Originally Posted by davieddy

You still don't get it, do you?

I can't say I exactly appreciate this as a rebuttal to what I'm trying to say. To reiterate:

We can get every exponent to 100M up to 73 before ever even looking at getting anything to 74. In the meantime, exponents in the upper 60M with only 73 bits done on them are being LL'ed because the TF effort is too focused on 2018's work instead of 2013's work.

Alternatively, we can focus on having a smaller buffer of more deeply factored exponents, which means that no work today is being done without the optimal amount of trial factoring.


I've been getting 30 GHz-Days of credit for every trial factoring job I do from 73 to 74 via my GPUs. Also, I get roughly 150 GHz-Days on a LL test in the same range via my CPU.

Trial factoring TO 73 would also take 30 GHz days. By the looks of things, it takes about two-fifths the effort to TF to 74 as it does to LL. Chris seems to believe that for every 150 GHz-Days that LL is getting, there IS 60 GHz-Days of TF. If there is, then we CAN keep up TF to 74 before releasing for further work.


David: How much of a 73-bits buffer do you even want? I found 384 composite results in the report recent cleared in the last 24 hours. If you're asking for a buffer of 50,000 then you've just gone completely <censored> mad. FOUR MONTHS? Explain to me how even a week's worth is going to suddenly evaporate.

The 50,000 exponents to 73 buffer you're asking for requires the same amount of work as a 16,666 exponent to 74 buffer. How is THAT not enough? It's like you think we're choosing between doing a few hundred to 74 versus doing tens of thousands to 73. If we HAD the capacity to make this buffer for ourselves then we would HAVE the capacity to be so far ahead of the LL curve with 74 bits that we'd be looking at doing 75 bits a few millions early.

[VictordeHolland]

I don't see the relevance of this discussion, since we are ahead of the LL and DC wavefronts. If you don't like factoring to 74 bit, nobody is forcing you to do so. You can take expo's to 73 and somebody else might do 73 to 74. Besides, the TF GPUto72 is doing is 'extra', in other words, higher than CPUs efficiently can do.

[chalsall]

Originally Posted by VictordeHolland

I don't see the relevance of this discussion.

Please understand this.

It is ill-advised to step in-between giants when they're being stupid.

[kladner]

Originally Posted by VictordeHolland

I don't see the relevance of this discussion, since we are ahead of the LL and DC wavefronts. If you don't like factoring to 74 bit, nobody is forcing you to do so. You can take expo's to 73 and somebody else might do 73 to 74. Besides, the TF GPUto72 is doing is 'extra', in other words, higher than CPUs efficiently can do.

This discussion mostly seems relevant to David's need to feel that he has some form of influence. If he cannot obtain this in the form of others following his advice, he gets it by provoking turmoil. He revels in gaining the attention of others, whether that attention is negative or positive.

He has been predicting a disastrous shortfall in factoring for as long as I have hung out here. It seems that he has probably done so for longer, but I feel no real need to review the history. In any case, it does not seem that these predictions are accurate.

Since the justifications, rationales, and proposed remedies for these predicted disasters have run a wide, and sometimes contradictory gamut, I am inclined to see the whole thing as a displacement of hostility and resentment arising from other causes. It seems that the community respect, which Chalsall has earned through much work, is probably one of these causes.

However, I rather think that there may well be multiple burrs under David's blanket, or bees in his bonnet if you prefer.

EDIT:
Chalsall said:

Please understand this.

It is ill-advised to step in-between giants when they're being stupid.

I suppose that I understand the peril, and can take the heat.

[chalsall]

Originally Posted by kladner

I suppose that I understand the peril, and can take the heat.

Good on you.

Some think they can cause you discomfort with no cost.

Some might be wrong....