Thinking out loud about getting DC's "finished" by 2030

[ATH]

http://hoegge.dk/mersenne/GIMPSstats.html

DC: (increase in LL-D column)
<100M: 1510 last 7 days ~ 216 per day, 6558 last 30 days ~ 219 per day

DC+Factoring: (decrease in LL column)
<100M: 1773 last 7 days ~ 253 per day, 7641 last 30 days ~ 255 per day


DC + Cert: (increase in LL-D column)
<1000M: 2687 last 7 days ~ 384 per day, 11563 last 30 days ~ 385 per day


Noticed that new LL tests >100M have almost stopped or are being double checked at the same pace:
100M-332M: -24 last 7 days, +18 last 30 days

[Jurzal]

I am doing my part

[Uncwilly]

Every bit helps!

[SethTro]

Originally Posted by ATH

Created a graph over the last 5 years development since Jan 1st 2018 of DCs and factors below 100M. Each point is the average number of DC / Factored (or sum of those) per day over the last 30 days from that point (below 100M).

These are the changes in the total DC and Factored columns (and NOT actual turned in DCs / factors from recent results), so in Sep 2021 the "DC rate" is actually negative per day for a few days because more factors were found for exponents in the DC column than new DC's completed.

There is a clear trend upwards in DC recently starting around Sep 5th 2022.

I'd love to see an updated version of this, watching DCstats it seems that 30 day DC processing power has increased from ~1.0 to ~1.2M Ghz-days/30 days = 40,000 Ghz of processing power over the last several months. At the current rate (1.2M Ghz-days / 30 days) we have 11 years of work left.

If we integrate effort by processing power, finishing by 31st Dec 2029 (6.7 years) requires a constant 6.7% processing power growth, which seems reasonable against Moore's law but maybe isn't realistic if we consider Ryan's significant contribution is likely transient.

[ATH]

Originally Posted by SethTro

I'd love to see an updated version of this, watching DCstats it seems that 30 day DC processing power has increased from ~1.0 to ~1.2M Ghz-days/30 days = 40,000 Ghz of processing power over the last several months. At the current rate (1.2M Ghz-days / 30 days) we have 11 years of work left.

If we integrate effort by processing power, finishing by 31st Dec 2029 (6.7 years) requires a constant 6.7% processing power growth, which seems reasonable against Moore's law but maybe isn't realistic if we consider Ryan's significant contribution is likely transient.

Here is an updated chart attached.

[Mark Rose]

Originally Posted by SethTro

I'd love to see an updated version of this, watching DCstats it seems that 30 day DC processing power has increased from ~1.0 to ~1.2M Ghz-days/30 days = 40,000 Ghz of processing power over the last several months. At the current rate (1.2M Ghz-days / 30 days) we have 11 years of work left.

If we integrate effort by processing power, finishing by 31st Dec 2029 (6.7 years) requires a constant 6.7% processing power growth, which seems reasonable against Moore's law but maybe isn't realistic if we consider Ryan's significant contribution is likely transient.

Does this also take into account the recent effort to redo P-1 on first time checked exponents?

[kriesel]

Originally Posted by SethTro

At the current rate (1.2M Ghz-days / 30 days) we have 11 years of work left.

Roughly equivalent to 100 Radeon VIIs. (At 400GHD/day/RadeonVII; varies depending on power settings, relative fft efficiency, and driver efficiency).

[ATH]

Made a graph of "GHz-days / day" of DC's instead of number of "exponents / day".

Using the FFT bounds below and https://www.mersenne.ca/credit.php to get the credit for each FFT size.
In each 1M section from 42M to 100M (DC edge was at 42M on 2018-01-01) I'm using the average exponent size times average credit in that 1M section times the number of DC's finished.


Again the graph at each point shows the average "GHz-days / day" during the last 30 days from that point.
First screenshot is the full graph, and 2nd screenshot is ignoring the big spike and zooming in on the lower part below 45000 GHz-days / day.


This graph shows that the last 4-6 months have higher throughput per day than any earlier times (except the big spike), which is not shown on the "exponents / day" graph because the smaller exponent sizes ~3 years ago made the throughput look larger back then.

2304K FFT	39.7M - 44.5M
2400K FFT	44.5M - 46.3M
2560K FFT	46.3M - 49.4M
2688K FFT	49.4M - 51.8M
2880K FFT	51.8M - 55.3M
3072K FFT (3M)	55.3M - 59.0M
3200K FFT	59.0M - 61.4M
3360K FFT	61.4M - 64.4M
3456K FFT	64.4M - 66.0M
3584K FFT	66.0M - 68.7M
3840K FFT	68.7M - 73.4M
4096K FFT (4M)	73.4M - 78.2M
4480K FFT	78.2M - 85.4M
4608K FFT	85.4M - 87.7M
4800K FFT	87.7M - 91.3M
5120K FFT (5M)	91.3M - 97.3M
5376K FFT	97.3M - 101.9M

[Andrew Usher]

These estimates still seem pretty on-target. Is the recent rise in DC throughput the cause of the fall in FTCs that I have noticed? That is, could the sum of the two be graphed the same way? Also what was that big spike - genuine?

I would really like to know where you got the FFT size breaks (for what version and processor type?) - I've looked in the source for them, and haven't found them. I've been trying to track the sizes used on my own machines, but that's still incomplete. The ability to calculate the exact FFT size that will be used has value for planning.

Those you present match almost exactly the theoretical bounds for 0.4 average roundoff (non-powers of 2 will always be slightly below that) while my results match a 0.36 roundoff better. They seem to additionally fall off at the lowest end and maybe at the highest, also, as the 1169M limit should be (rounded down) 1175 at 0.4 or 1170 at 0.36 roundoff.

[Andrew Usher]

Never mind one point - I just saw why it has to fall off at the smallest sizes (obvious really, but as usual overlooked in theoretical papers). Exactly how much might be interesting, though.