RIP DCTF.

petrw1 · 2014-12-12, 18:32

Disclaimers: GhzDays ToGo for now is a rough estimate ... ideas of how to make it somewhat accurate?

Code:

GhzDays	Member
550	TheMawn
500	petrw1
300	manfred4
460	kladner
1800	LaurV
1200	Mark Rose
4810	Total
33,670	Per Week
3,000,000	 ToGo
	
89.10	Est Weeks
1.71	Est Years
26-Aug-2016	Est Done

petrw1 · 2014-12-12, 18:39

Quote:

Originally Posted by chalsall

Take a look at the second table on this report. ~240 days to complete as the TF goal is currently configured (based on our thirty day rolling average of DCTF'ing).

However, as LaurV (correctly) pointed out, the "optimal TF level" for these ranges are actually higher for the higher ranges than shown, so we have some more work to do. I will adjust the report to be "correct" over this next weekend.

Also, keep in mind that you guys only seriously "surged" two weeks ago. The estimates given on the above report is based on the middle of the linear regression line on this graph. Read: If you guys keep this up, the estimate will continue to drop.

Fair enough....but we are trying to estimate the remaining time for only those who commit to stay on this "Project"...
That will change quite a bit shortly as many are moved back to LLTF with LetGPU72Decide.

Granted once this settles the number will gradually "auto correct" itself again.

Any chance we could get one more column which is the actual GhzDays to complete all the work in that row of the chart?

manfred4 · 2014-12-12, 18:47

I calculated my Estimate by calculating the Credit for each Million range (e.g. 38,500,000 for averaging) and added them together, so it should be pretty accurate ;)
If you say, from where to go to 72 bits or even 73 bits (I think there are some, that need to go there) I can add it up to give a new and better number.

VictordeHolland · 2014-12-12, 19:31

Quote:

Originally Posted by manfred4

If you say, from where to go to 72 bits or even 73 bits (I think there are some, that need to go there) I can add it up to give a new and better number.

Those are different for the different GPU architectures and whether you prefer TF or CudaLucas.

To 72bit
GTX980 @38M 71.92bits http://www.mersenne.ca/cudalucas.php?model=566
GTX580 @40M 71.96bits http://www.mersenne.ca/cudalucas.php?model=12
GTX680 @43M 72.04bits http://www.mersenne.ca/cudalucas.php?model=467
GTX780 @49M 71.93bits http://www.mersenne.ca/cudalucas.php?model=492

To 73bit
GTX980 @48M 73.15bits
GTX580 @50M 72.99bits
GTX680 @54M 72.95bits
GTX780 @65M 73.04bits

So I would personally say from 40M on to 72bits and @50M to73bit, as the GTX580 is still a very popular card for TF. People with a GTX780 and who do mind their efficiency could return to LLTF or run CudaLucas instead.

manfred4 · 2014-12-12, 20:22

Okay, but I think that the value in that graph is the maximum bit depth you should go, so if it says 72.9 bits, then 73 bits is just too much and you waste your effort. For my GPU that would be at 43M/55M for the next bit.
For an average 42M/53M looks reasonable.

Mark Rose · 2014-12-12, 20:24

Given that all the DCTF have already had at least one LL check, we're only saving one LL check not two. Should this be factored into the bit level, too?

manfred4 · 2014-12-12, 20:27

That and the fact, that P-1 is done already is factored in, as you can see there are two lines in those charts, one for DCTF and one for LLTF

VictordeHolland · 2014-12-12, 20:41

Quote:

Originally Posted by manfred4

That and the fact, that P-1 is done already is factored in, as you can see there are two lines in those charts, one for DCTF and one for LLTF

But isn't the number listed (for instance 71.95) at the start of the range? In that case we should increase the bitlevel halfway, for instance between 40,000,000 and 41,000,000 ?

Mark Rose · 2014-12-12, 20:42

Quote:

Originally Posted by manfred4

That and the fact, that P-1 is done already is factored in, as you can see there are two lines in those charts, one for DCTF and one for LLTF

Brainfart on my part

petrw1 · 2014-12-15, 01:58

From here http://www.gpu72.com/reports/workers/dctf/week/
115,508 GhzDays DCTF in the last week.
59 different contributors
238 Factors found
13,123 P1/LL/DC work saved.

From here: http://www.gpu72.com/reports/workers/dctf/
31 contributors currently have assignments
8,856 Assignments out.

From here: http://www.gpu72.com/reports/estimat...tion/primenet/
238.1908 estimated days to completion
August 6, 2015

Mark Rose · 2014-12-15, 03:27

LG72D has been switched from DCTF to LLTF. That's probably what dropped the participants.

It also looks like the throughput has dropped from 18 THz-d/d to 13 THz-d/d. The time to estimated completion is now 240 days and still dropping several days a day, so we should be done in less than 8 months from now.

2014-12-12, 18:32	#1
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 10711₈ Posts	RIP DCTF. Disclaimers: GhzDays ToGo for now is a rough estimate ... ideas of how to make it somewhat accurate? Code: GhzDays Member 550 TheMawn 500 petrw1 300 manfred4 460 kladner 1800 LaurV 1200 Mark Rose 4810 Total 33,670 Per Week 3,000,000 ToGo 89.10 Est Weeks 1.71 Est Years 26-Aug-2016 Est Done Last fiddled with by petrw1 on 2014-12-12 at 18:34

2014-12-15, 01:58	#10
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 10711₈ Posts	Sunday Night Wrapup From here http://www.gpu72.com/reports/workers/dctf/week/ 115,508 GhzDays DCTF in the last week. 59 different contributors 238 Factors found 13,123 P1/LL/DC work saved. From here: http://www.gpu72.com/reports/workers/dctf/ 31 contributors currently have assignments 8,856 Assignments out. From here: http://www.gpu72.com/reports/estimat...tion/primenet/ 238.1908 estimated days to completion August 6, 2015

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Getting unneed DCTF work	Mark Rose	GPU to 72	4	2018-01-01 06:14

2014-12-12, 18:47	#3
manfred4 Mar 2014 Germany 2³×3×5 Posts	I calculated my Estimate by calculating the Credit for each Million range (e.g. 38,500,000 for averaging) and added them together, so it should be pretty accurate ;) If you say, from where to go to 72 bits or even 73 bits (I think there are some, that need to go there) I can add it up to give a new and better number.

2014-12-12, 20:22	#5
manfred4 Mar 2014 Germany 2³·3·5 Posts	Okay, but I think that the value in that graph is the maximum bit depth you should go, so if it says 72.9 bits, then 73 bits is just too much and you waste your effort. For my GPU that would be at 43M/55M for the next bit. For an average 42M/53M looks reasonable.

2014-12-12, 20:24	#6
Mark Rose "/X\(‘-‘)/X\" Jan 2013 29·101 Posts	Given that all the DCTF have already had at least one LL check, we're only saving one LL check not two. Should this be factored into the bit level, too?

2014-12-12, 20:27	#7
manfred4 Mar 2014 Germany 2³×3×5 Posts	That and the fact, that P-1 is done already is factored in, as you can see there are two lines in those charts, one for DCTF and one for LLTF

2014-12-15, 03:27	#11
Mark Rose "/X\(‘-‘)/X\" Jan 2013 29·101 Posts	LG72D has been switched from DCTF to LLTF. That's probably what dropped the participants. It also looks like the throughput has dropped from 18 THz-d/d to 13 THz-d/d. The time to estimated completion is now 240 days and still dropping several days a day, so we should be done in less than 8 months from now.