Should there exists a "Moore's Law" version of the GHz-Days calculated?

[Neutron3529]

Although now I'm using GPU as an expensive room heater, I am very doubt for how much I've actually done with GPU TF.


the default measurement is "GHz-Days", I found it is acceptable since it measure the calculation for us. It is fair.


But recently I think "fair" is not enough for finding a new mersenne prime.
So I want to know if there exists a better way to estimate how much I've done with my Room Heater


For example, suggest that GIMPS will do a LL test for M1,000,000,000,039 1000 years later. and I have a SUPER-FAST computer so I decided to start a TF routine for primes larger than 1,000,000,000,000. Maybe I will found a factor per year. 10 years pasts and I may gain a huge amount of GHz-Days, which is not helpful for GIMPS project at all.


That is not the worst thing. Since Moore's Law may keep in the future (new AMD CPU/GPU, Quantum computer, etc.), a personal computer might be capable to do the same TF routine in 500 years and finish TF in 1 year.


If I start a LL routine rather than a TF routine, the SUPER-FAST computer may generate meanful results for GIMPS and GIMPS may reach larger Mersenne primes earlier.


I want to know, how much time it actually save to do a TF for large mersenne numbers and find a factor.





Since the propose for GPU TF is to elimated the proposal candidates.
If I elimated a number (for example, M211092829 has a factor: 1774657187806827090727) the contribute should be equals to someone do a TF test and shows that M211092829 is not prime.


The question is quite same to what I supposed above. I elimated this number now, but the result is not needed for GIMPS project until we goes to M200,000,000 or even larger. This might happened several years later, and with Moore's Law, we should have more computation resources in the future so that It may take less LL time for a person in the future to show M211092829 is not prime.


It is not difficult calculating the first time LL speed for GIMPS project, so we could have a acceptable result for both how much GHz-Days will be spend during a LL test and how soon will M211092829 enter GIMPS's LL list. If we further assume Moore's Law holds in the future, we could convert GHz-Days saved in the future to the time generate such many GHz-Days, and know how much time we saved.





Will GIMPS support such Moore's Law GHz-Days in the future?

[kriesel]

Originally Posted by Neutron3529

For example, suggest that GIMPS will do a LL test for M1,000,000,000,039 1000 years later. and I have a SUPER-FAST computer so I decided to start a TF routine for primes larger than 1,000,000,000,000. Maybe I will found a factor per year. 10 years pasts and I may gain a huge amount of GHz-Days, which is not helpful for GIMPS project at all.

Planning on smashing some longevity records? https://en.wikipedia.org/wiki/List_o..._oldest_people
Also handily beating the recent and current rate of GIMPS progress? https://www.mersenneforum.org/showpo...5&postcount=11
The 87M milestone for 2019 is in doubt.

Since Moore's Law may keep in the future (new AMD CPU/GPU, Quantum computer, etc.), a personal computer might be capable to do the same TF routine in 500 years and finish TF in 1 year.

Not only will Moore's law fail, inevitably, since it is reliant on ever shrinking feature size in things made out of atoms with fixed size, it appears to be failing now. Clock rates have stopped climbing. Increased parallelism, chip size, and larger power budget are being relied on more and more. Moore himself expects it to stop (in a decade from 2015). https://en.wikipedia.org/wiki/Moore's_law
Note also that as the features become exponentially smaller, the cleanliness must become exponentially better, and the fab plant capital costs exponentially grow.

Will GIMPS support such Moore's Law GHz-Days in the future?

I hope not. Units that change in meaning with time are close to having meaningless units.

[LaurV]

Originally Posted by kriesel

larger power budget are being relied on more and more

That's wrong, you have to say "Mo-ore and Mo-ore"...
We totally agree about not supporting "volatile" units. However, GIMPS will adapt to any new situation (thinking about changing from P90-years to GHzDays in the past).
(edit: and why everybody write this unit always wrong? It is neither GHz-Days, nor GHz/Days, correctly is GHzDays/Day, and it can be abbreviated to GHzD/D, only!)

[axn]

Originally Posted by LaurV

It is neither GHz-Days, nor GHz/Days, correctly is GHzDays/Day, and it can be abbreviated to GHzD/D, only!)

GhzD/D is the rate of credit. Actual accumulated credit is GHzD (however you want to write it, including GHz-Days)

[kriesel]

Originally Posted by LaurV

That's wrong, you have to say "Mo-ore and Mo-ore"...
We totally agree about not supporting "volatile" units. However, GIMPS will adapt to any new situation (thinking about changing from P90-years to GHzDays in the past).
(edit: and why everybody write this unit always wrong? It is neither GHz-Days, nor GHz/Days, correctly is GHzDays/Day, and it can be abbreviated to GHzD/D, only!)

Even GHz-Days/Day is an abbreviation, for the throughput rate of one core of a theoretical 1 GHz Core 2 cpu. I maintain that GHz-Days/Day is a very clumsy unit for throughput, needlessly verbose, and incorrect. My engineering professors would never have allowed the uncanceled units.
They might have tolerated GHz-equivalent, abbreviated GHe. That would make the accumulated effort of a computation some number of GHeD, and avoid possible trademark issues with Ge.

[R.D. Silverman]

Originally Posted by kriesel

Even GHz-Days/Day is an abbreviation, for the throughput rate of one core of a theoretical 1 GHz Core 2 cpu. I maintain that GHz-Days/Day is a very clumsy unit for throughput, needlessly verbose, and incorrect. My engineering professors would never have allowed the uncanceled units.
They might have tolerated GHz-equivalent, abbreviated GHe. That would make the accumulated effort of a computation some number of GHeD, and avoid possible trademark issues with Ge.

see my paper Exposing the Mythical MIPS-Year

https://www.researchgate.net/publica...ical_MIPS_year

[storm5510]

Originally Posted by kriesel

...I hope not. Units that change in meaning with time are close to having meaningless units.

This has been a hard concept to wrap my brain around even though I understand its meaning: One gigahertz day is a device running at one gigahertz for 24 hours. If this is incorrect, somebody step-up. It is the translation to higher speeds i struggle with.

Bottom line: Leave it as is because it is consistent. Changing it would throw well over a decade's worth of data on the junk heap.

[kriesel]

Originally Posted by R.D. Silverman

see my paper Exposing the Mythical MIPS-Year

https://www.researchgate.net/publica...ical_MIPS_year

And Digital Equipment Corp defended the VAX instruction set as getting more done per instruction than a typical MIPS-measured instruction mix, creating a new measure called a VUP. https://en.wikipedia.org/wiki/VAX_Unit_of_Performance
There was a proliferation of measures, int and float and double, and eventually with year suffixes, such as SpecInt92.

[ATH]

Originally Posted by storm5510

This has been a hard concept to wrap my brain around even though I understand its meaning: One gigahertz day is a device running at one gigahertz for 24 hours. If this is incorrect, somebody step-up. It is the translation to higher speeds i struggle with.

In GIMPS 1 Ghz-days is defined as 2 GFLOP/sec for 1 day (86400sec) so 172,800 GFLOP = 172.8 TFLOP in total.

I guess it was determined that a 1 Ghz processor could do roughly 2 GFLOP/sec.

[Neutron3529]

Originally Posted by kriesel

Planning on smashing some longevity records? https://en.wikipedia.org/wiki/List_o..._oldest_people
Also handily beating the recent and current rate of GIMPS progress? https://www.mersenneforum.org/showpo...5&postcount=11
The 87M milestone for 2019 is in doubt.

I'm sorry for both my poor math and incorrect assumptions.
I just believe that the cost of TF a mersenne number now may be larger than TF it several years later. And in my opinion the most common assumptions is that the cost obey the Moore's law.
In your milestone, the first time of LL test for exponent over 214M will later than the year 2040. No one knows what will happen in the future, but I just believe that assuming the cost of TF in the future will equals to TF it now is not the best choice.


Now, I am TF-ing a larger mersenne number since TF it is faster than TF a smaller number, and could generate positive results very fast.(TF over 200M, found about 4-7 factor per day with my laptop)
And with your milestone, before 2040 my results are useless and actually save no time for first time LL test.
How to tell me that TF something will be useful after 2040 is not worth than TF what will enter GIMPS's LL list now?
I think a Moore's Law version of GHz-Days may help.


That's my idea.

[Neutron3529]

Originally Posted by LaurV

(edit: and why everybody write this unit always wrong? It is neither GHz-Days, nor GHz/Days, correctly is GHzDays/Day, and it can be abbreviated to GHzD/D, only!)

I got the unit GHz-Days in https://www.mersenne.org/tmembers/
GHz-Days Lifetime, GHz-Days last year, etc.


TF will save no GHz-Days/Days since we could not save a ratio, that's why I use GHz-Days and want a Moore's Law version of GHz-Days