Predict discovery date of the 1st 100M digit prime

[Flatlander]

Originally Posted by davieddy

You've been reading
http://primes.utm.edu/notes/by_year.html#3
haven't you?

Nope. I haven't looked at that for a year or so.
Tony's guess of 2023 looked too far ahead so I went about half way.
I was also figuring on a couple of hardware breakthroughs and everyone having 16 or 32 cores.
It's all my other posts that are pure plagiarism.

edit:
I was also thinking that there might be a reward for the first 100m digit prime, so people would be searching artificially high.

[T.Rex]

Originally Posted by Flatlander

... everyone having 16 or 32 cores.

The main problem with many cores is the difficulty/cost to have a good hardware and software scalability !!
HW scalability is VERY expensive now. FFT scalability is difficult. Preliminary results (to be confirmed) are showing that, on a 8 cores, Mlucas is running 5.4 times faster than with only 1, though Prime95 has a 3.3 factor only. It would be worst with 16 or 32 cores. IBM/Bull Power machines are VERY scalable ! But a 64xPower6 machine costs more than 1MEuros...
Tony

[davieddy]

Originally Posted by Flatlander

Nope. I haven't looked at that for a year or so.
Tony's guess of 2023 looked too far ahead so I went about half way.
I was also figuring on a couple of hardware breakthroughs and everyone having 16 or 32 cores.
It's all my other posts that are pure plagiarism.

edit:
I was also thinking that there might be a reward for the first 100m digit prime, so people would be searching artificially high.

viz a viz the GIMPS wavefront, a doubling of the exponent involves
a multiplying by ten in the power needed to keep up:
iterations x numbers to test x time per iteration.

People tried leaping to 10M digits a long time ago without success.

David

[retina]

Originally Posted by Flatlander

I was also thinking that there might be a reward for the first 100m digit prime, so people would be searching artificially high.

There is, currently US$150K from EFF.

[Flatlander]

Originally Posted by T.Rex

The main problem with many cores is the difficulty/cost to have a good hardware and software scalability !!
HW scalability is VERY expensive now. FFT scalability is difficult. Preliminary results (to be confirmed) are showing that, on a 8 cores, Mlucas is running 5.4 times faster than with only 1, though Prime95 has a 3.3 factor only. It would be worst with 16 or 32 cores. IBM/Bull Power machines are VERY scalable ! But a 64xPower6 machine costs more than 1MEuros...
Tony

I was assuming people would run 16 or 32 concurrent tests and CPUs would have huge on-die caches. (But why would anyone apart from prime hunters etc. want that??? And I've used up one of my "couple of hardware breakthroughs". Just one left; I'd better shut up.)


edit (Will I never learn?):
How big would an on-die cache have to be to hold a 100m digit test?

[T.Rex]

Originally Posted by Flatlander

I was assuming people would run 16 or 32 concurrent tests and CPUs would have huge on-die caches.

Yes, you're right. But, since the speed of CPUs will stay about 3GHz, one core will take several months to check 1 bigger exponent. So people will use 2, then 4, ... cores for one exponent. In order to see a result every 1 or 2 months, I think. But I may be wrong...
T.

[jinydu]

Originally Posted by davieddy

People tried leaping to 10M digits a long time ago without success.

As is discussed here (http://www.mersenneforum.org/showthr...114#post142114), several exponents above 100M digits have already been reserved for LL testing. But I'm not sure ANONYMOUS seriously intends to carry through the LL tests to the end.

[ixfd64]

Originally Posted by T.Rex

The main problem with many cores is the difficulty/cost to have a good hardware and software scalability !!
HW scalability is VERY expensive now. FFT scalability is difficult. Preliminary results (to be confirmed) are showing that, on a 8 cores, Mlucas is running 5.4 times faster than with only 1, though Prime95 has a 3.3 factor only. It would be worst with 16 or 32 cores. IBM/Bull Power machines are VERY scalable ! But a 64xPower6 machine costs more than 1MEuros...
Tony

I guess we really need a new version of Prime95!

[henryzz]

25.6 is a new version of Prime95

[Nelson]

Originally Posted by henryzz

25.6 is a new version of Prime95

I think ixfd64 means newer than Prime95 v.25.6. That's what Tony is using to arrive at his scaling benchmark. Prime95 v.24.14 uses only one thread and cannot achieve any type of scaling.

nelson

[fivemack]

Originally Posted by Flatlander

edit (Will I never learn?):
How big would an on-die cache have to be to hold a 100m digit test?

100M digits is 330M bits; so the FFT size is about 2^24 complex-doubles of 16 bytes each, so 256MB. Which is not ridiculous for on-die cache for an ultra-high-end product in 22nm, a high-end infinitely-rich-gamer one in 16nm or a normal-cost-unconscious-user one in 11nm - Intel's 32nm process has a cell size of 0.182um^2, so 256MB in that process would occupy 390mm^2 and leave little space for cores on a maximum-size die.

(to calibrate, current ultra-high-end is Dunnington, high-end infinitely-rich-gamer is QX9772, normal-user is Q6600)

Tulsa ('Xeon 7100 series') has a 16MB L3 cache made with 0.624um^2 cells in 65nm; the Itanium 9100 has a 24MB L3 cache made with 1um^2 cells in 90nm but that's still only about 190mm^2. So it's probably unrealistic to expect more than 100mm^2 of L3 cache on any remotely affordable processor.