When will GIMPS's first prime...

[axn]

Originally Posted by CRGreathouse

Assuming (for simplicity) that the work needed to test an exponent is proportional to the square of the exponent and that all prime exponents are tested, that would require about 10,000 times the total effort of GIMPS to date. If half of that effort happened in the past two years, and GIMPS' computing power doubles every two years (by some combination of Moore's law and recruitment), this would require 2 lg(ln(2) * 20,001) or 27.5 years. Assuming instead that it doubles every two years for a decade and then holds constant, it would take about 2 * 20000/2^5 + 10 = 1260 years. So the timeline depends strongly on the assumptions made.

GPUs are the game changers. They will, in fact, keep pace with Moore's law, and bump the constants. I expect 2 decades (i.e before the end of 2030) for the necessary 8 primes to pop out.

[CRGreathouse]

Originally Posted by axn

GPUs are the game changers. They will, in fact, keep pace with Moore's law, and bump the constants. I expect 2 decades (i.e before the end of 2030) for the necessary 8 primes to pop out.

I'll bet you a (post-inflation) quarter* that we won't have the 8 by 2030.

I happily cede that they're game-changing, but I'd be surprised if their performance doubled even every three years after, say, 2025.

* With 2-3% inflation it would only be worth 13 to 17 cents in 2010 dollars by the time I'd be able to collect.

[axn]

Originally Posted by CRGreathouse

I'll bet you a (post-inflation) quarter* that we won't have the 8 by 2030.

I happily cede that they're game-changing, but I'd be surprised if their performance doubled even every three years after, say, 2025.

* With 2-3% inflation it would only be worth 13 to 17 cents in 2010 dollars by the time I'd be able to collect.

You're on

[axn]

Another data point: Today, M35 is at 274. That's a drop of 104 positions in 244 days. Linear projection gives me 2014 days to fall off top 5000 - that is 5.5 years! That means in practice, it'll be like 3-4 years!

[mart_r]

I've been keeping track of the rank of numbers as well... lemme see...
100000-digit-mark: What was at # 170 by the end of 2002 dropped below # 5000 in July 2007.
120000-digit-mark: # 170 @ 1st quarter of 2003, # 5000 @ about early May 2009.
150000-digit-mark: # 170 @ 1st or early 2nd quarter of 2004, # 5000 @ mid April 2010.
--> Prediction: late 2016.

But right now it's 274th, so...
100000-digit-mark: # 274 @ mid 2003 --> about 4 years
120000-digit-mark: # 274 @ 1st quarter 2004 --> a bit more than 5 years
150000-digit-mark: # 274 @ early 2005 (?, rather roughly interpolated) --> about 5 years.
200000-digit-mark: # 274 @ early/mid 2006, # 5000 @ abt. April 2011 --> again about 5 years.

So my educated guess remains: late 2016.

[jasong]

The real problem, in my opinion, is that cpus(now referred to as cores, since what we now call cpus actually tend to contain more than one actual cpu) haven't increased in speed all that much, so while people can do multiple tests, each test is getting longer. So, I think the biggest enemy to GIMPS is cruncher boredom.

I know this has probably been suggested before, but sometime in the future I think GIMPS should consider keeping intermediate residues(or whatever they're called) for Mersenne tests. Like every 5 million bits or so.

[aketilander]

Originally Posted by jasong

I know this has probably been suggested before, but sometime in the future I think GIMPS should consider keeping intermediate residues(or whatever they're called) for Mersenne tests. Like every 5 million bits or so.

Yes, I think jasong is right, especially when large LL-tests are concerned (above 332.000.000). You see a lot of LL-work being done in this region but almost no results because the participants get tired, the computers break, and so on, before the LL-tests finish. Wouldn't it be a good idea to upload the backup-files (intermediary results file output) to the primeserver lets say for every 25.000.000 iteration when this region (above 332.000.000) is concerned. In that case someone else could continue the test from that point if the first one gets tired and drops out. If you also calculate a residue for every 25.000.000 iteration the LL-D test could be interrupted or a second LL-D test could be continued at the last point where we had a residue match between the LL-test and first LL-D-test.

I think that such a procedure would save quite a lot of energy and GHz-days.

[Brain]

Originally Posted by jasong

I know this has probably been suggested before, but sometime in the future I think GIMPS should consider keeping intermediate residues(or whatever they're called) for Mersenne tests. Like every 5 million bits or so.

I love this idea. LL resuming could really speed PrimeNet up.
Preconditions: Exponent is assigned to myself or abandoned and error-free.
Saving 1 million LL tests in 4M range would take 32MB * 1.000.000 = 32TB... Derived from CUDALucas save files. Or am I wrong?

I suggest enhancing to P-1: One could also resume P-1 runs that hadn't stage 2 done.
I have to admit that PrimeNet's throughput would not increase by factoring numbers that have already been LL tested. But it's fun.

[aketilander]

Originally Posted by Brain

Saving 1 million LL tests in 4M range would take 32MB * 1.000.000 = 32TB... Derived from CUDALucas save files. Or am I wrong?

That sounds a lot! But I suppose if the stored tests are always the first ones to go when someone is requesting a new test, then I suppose you don't need to store 1 million of them, maybe a few thousand will be sufficient.

[Christenson]

Originally Posted by aketilander

That sounds a lot! But I suppose if the stored tests are always the first ones to go when someone is requesting a new test, then I suppose you don't need to store 1 million of them, maybe a few thousand will be sufficient.

We have < 100K participants, so that's only 3.2TB...assuming ALL of us abandoned an exponent mid-way, which isn't going to happen....

3.2TB (or double that, for a hot backup) isn't out of reach of Primenet, if P95 and Scott ask for the donations -- look what MET did to the forum budget, and I think you'd only need about $200, not $500. But I'd start out at M(100M), not M(332M).
[So says the man that's working on billion digit primes. P95? Scott?]

[Brain]

Originally Posted by aketilander

That sounds a lot! But I suppose if the stored tests are always the first ones to go when someone is requesting a new test, then I suppose you don't need to store 1 million of them, maybe a few thousand will be sufficient.

Prime95 save file size is about 6.5MB for 3M exponent. Maybe already compressed?