[QUOTE=TheMawn;398015]EDIT: Nevermind! I figured it out. It's for factoring in the 100M+ digits range. By the way, scroll down to the bottom of [url]https://www.gpu72.com/reports/worker/factors/d8a75f85f90457298bd3c366a8de2410/[/url] and can you explain to me why the DC cost is 10,000 GHz-days flat whereas the LL cost is not?[/QUOTE]

Yeah... LaurV is correct -- don't trust my "Work Saved" metric! But, in my defence, it wasn't an actual intended cheat, simply a SPE when I was experimenting with having large LMH work being done through GPU72 and I forgot to filter out the results.

WRT the "DC Saved" on the report you linked to showing 10,000.00 while the "LL Saved" shows large (but different) values, again, a SPE. The "GHzDaysLL" field is defined as "float(20,10) unsigned" in that table, while the "GHzDaysDC" field is defined as "float(14,10) unsigned".

I'll look at filtering out my high "saved" values from the database sometime; until then, simply assume I'm cheating.... :smile:

P.S. Actually, this brings up a memory... When this was raised before (years ago) James proposed a very elegant equation which would solve the problem. James, please forgive me for this, but can you find the post where this was defined (or repost)? It would probably take as much work for me to implement your suggestion as to filter out "edge" cases.

[QUOTE=chalsall;398086]James proposed a very elegant equation which would solve the problem. James, please forgive me for this, but can you find the post where this was defined (or repost)? It would probably take as much work for me to implement your suggestion as to filter out "edge" cases.[/QUOTE]I remember that. It was elegant. I'll have to think about what I proposed to try and figure out where I posted it... :unsure:
Google is impressively good sometimes, but not quite to the point of "remember that clever idea I had a while ago?..." :smile:

[QUOTE=LaurV;398063]I tell you but don't tell to these guys here: if I drill a hole in my garden straight down, it goes through the Earth core and pops directly into Chris garden... :w00t:[/QUOTE]

Yeah, don't tell anyone, but LaurV and I have an arrangement. I send him authentic Bajan food through the tunnel (arriving 42.2 minutes later), and he reciprocates with authentic Thai (I think I'm getting the better part of the deal).

The tunnel was relatively easy; the stasis field technology (patent pending) to keep the food at the intended temperature was the difficult part....

After some searching I was unable to find the post where I first suggested it. But it was something along the lines of a scaled number that isn't really how much work was saved, but "rewards" finding larger/harder factors equally on any exponent, rather than trivial factors on large exponents. For example I'm finding 1 factor per second across my 4 GPUs looking in the ~3500M range up to 2[sup]64[/sup], and each factor found saves about 100,000GHz-days of LL effort (200k if you count DC). But I only put 0.5s of GPU time into finding it.

Some examples[code]mysql> SELECT `exponent`, `factor`, `factorbits`, POW(2, `factorbits`) / (`exponent` * `exponent`) AS `ratio` FROM `known_factors_000` WHERE CEIL(`factorbits`) = 32 LIMIT 1;
+------------+--------------------------+------------+-------------------+
| exponent | factor | factorbits | ratio |
+------------+--------------------------+------------+-------------------+
| 183329 | 2610604961 | 31.2817 | 0.0776725 |
| 180503 | 3617977583593 | 41.7183 | 111.0428958 |
| 9255461 | 2586129499585367 | 51.1997 | 30.1890313 |
| 180799 | 3519866543349568537 | 61.6102 | 107677727.6770670 |
| 50000017 | 3615901229407 | 41.7175 | 0.0014463 |
| 50003377 | 2629357074783431 | 51.2236 | 1.0515767 |
| 50008481 | 3034732720866794417 | 61.3963 | 1213.5033475 |
| 50005399 | 3363174795243180938039 | 71.5103 | 1344966.3214317 |
| 150000047 | 2244000703121 | 41.0292 | 0.0000997 |
| 150000667 | 3119505371338871 | 51.4702 | 0.1386396 |
| 150001763 | 4550635694009568991 | 61.9808 | 202.2494818 |
| 159919951 | 2545635882015204168857 | 71.1085 | 99537.2032619 |
| 500001259 | 2397006035647 | 41.1244 | 0.0000095 |
| 500000467 | 4503064205858041 | 51.9998 | 0.0180118 |
| 500001763 | 2741363214012610249 | 61.2496 | 10.9653775 |
| 500000701 | 4705811278909163282143 | 71.9949 | 18822.8023273 |
| 1508205187 | 2569981638649 | 41.2249 | 0.0000011 |
| 1500000107 | 3765639268615583 | 51.7418 | 0.0016735 |
| 1504000241 | 4011556747001189569 | 61.7989 | 1.7734827 |
| 4205000179 | 2287520097377 | 41.0569 | 0.0000001 |
| 4203000167 | 4451052830856007 | 51.9831 | 0.0002519 |
| 4209000109 | 2642363834740997513 | 61.1965 | 0.1491502 |
| 4200003253 | 4559069828180196281983 | 71.9492 | 258.4456111 |
+------------+--------------------------+------------+-------------------+[/code]I don't think this is quite the same as what I had before, but something along the general idea. Those who understand math can tweak what I'm using to something that makes more sense, I just played with numbers until I saw something that looked reasonable (monkeys and typewriters, you know...)

[QUOTE=chalsall;398089]...arriving 42.2 minutes later[/QUOTE]For an average speed of 5km/s. Now that's some fast food! :w00t:

[STRIKE]Do you have a rough date? Like 2011-2013 or better?[/STRIKE]

[STRIKE]Was it in this thread?[/STRIKE] Do you remember which thread it was in?

Edit: Working on the assumption that it was in this thread, I started searching. Is this it?

[QUOTE=James Heinrich;323022]You need some kind of self-balancing metric, perhaps something along the lines of[code]worth = GHd_saved * (GHd_factor / GHd_LL)

// examples:
// 72-bit TF factor on 60M (TF to 2[sup]73[/sup])
value = (133.292 + 133.292 + 15.94) * (11.956 / 133.292) = 89.7

// 72-bit TF factor on 900M (TF to 2[sup]84[/sup])
value = (24825 + 24825 + 4352) * (0.5314 / 24825) = 1.2

// 83-bit TF factor on 900M (TF to 2[sup]84[/sup])
value = (24825 + 24825 + 2176) * (1088 / 24825) = 2271

// 93-bit P-1 factor on 900M (TF to 2[sup]84[/sup])
value = (24825 + 24825 + 0) * (684 / 24825) = 1368[/code]This correctly shows that a 72-bit factor is worth a lot less on larger exponents than on smaller, despite "saving" a lot more LL effort. As can be seen above, it also works well with P-1 factors -- large factors can be found with relatively less effort than TF factors, but the above automatically scales it in what I think is an appropriate manner.[/QUOTE]

[QUOTE=Dubslow;398096]Edit: Working on the assumption that it was in this thread, I started searching. Is this it?[/QUOTE]

Yup! Thanks!!!

[QUOTE=James Heinrich;398095]For an average speed of 5km/s. Now that's some fast food! :w00t:[/QUOTE]

We're considering offering franchises.

[QUOTE=chalsall;398097]Yup! Thanks!!![/QUOTE]

That formula is ... problematic. Because, after simplification, it comes out as (2 + p-1 bonus) * GHd_factor, where p-1 bonus is earned if an exponent is yet to have a P-1, and is of the order of roughly 0.1. So basically, a small number (2-2.1) times GHd_factor. In other words, you are still measuring the work expended in finding the factor, not the work saved!

[QUOTE=axn;398137]That formula is ... problematic.[/QUOTE]I'd forgotten what my original proposal was until [i]Dubslow[/i] found it, but my revised proposal [url=http://www.mersenneforum.org/showpost.php?p=398092&postcount=3513]above[/url] depends on bits and not GHz-days. Feel free to critique it also, I have no sentimental attachment to it, nor am I skilled in math so I'm sure someone else can propose a similar concept with numbers that make more sense.

[QUOTE=James Heinrich;398178]I'd forgotten what my original proposal was until [i]Dubslow[/i] found it, but my revised proposal [url=http://www.mersenneforum.org/showpost.php?p=398092&postcount=3513]above[/url] depends on bits and not GHz-days. Feel free to critique it also, I have no sentimental attachment to it, nor am I skilled in math so I'm sure someone else can propose a similar concept with numbers that make more sense.[/QUOTE]

Well, IMO, work saved is not (should not be) related to work expended. Straightforward calculation of work saved is good enough. However, in order to avoid people gaming the system by TF-ing too far ahead of the LL wavefront, you can discount the work saved by how far away the exponent is from the current LL (@ 4m/year). This means that a factor found ahead of the wavefront starts out with a small value of "work saved", but will appreciate over time, eventually reaching the full value (after decades?!).