RIP DCTF.

[Mark Rose]

Another little update:

Exponents between 58M and 105M with a completed LL test are DCTF'ed to an appropriate level for a GTX 580. I found 6 factors.

Exponents between 58M and 91M with an assigned LL test are TF'ed to an appropriate level for a GTX 580. I found a few dozen that were assigned for P-1/LL in September and October with no P-1/low TF, probably when we ran out of exponents when a couple users requested big batches. I should have everything up to 105M done by the end of the week. So far I found one factor. I am doing this work to avoid more of the other category of work in the coming months.

Overall, about 20 THz-days of work. I'm pleased there was so little as that shows the system with GPU72 is working very well.

[Mark Rose]

Originally Posted by chalsall

Excellent! Thanks! The way things are going, we could have DCTF completed in about two months or so.

The great thing about this is when the next Mersenne Prime is found and we get a huge influx of new users, most will be first assigned DC'ing work until they've proven themselves. Very *very* few will actually complete, but those who do will have worked candidates already appropriately TF'ed.

I think that's how I heard of the project, when M6972593 was discovered. I downloaded the project, and I do believe I completed an anonymous DC assignment.

It was boring though.

I think what the project needs is graphing or charts showing the completion to the next milestone. Kind of like what GPU72 has for exponents. GPU72 shows how much is done, and gives mini-milestones by the subcategorisation of work into individual table cells. It's easier to see progress with GPU72. Perhaps instead of just giving a count milestone, plot the daily values so people can "see" progress?

[LaurV]

A post for the wish thread:

I wish that DCTF project would never die...
(i.e. we would never be able to finish it)

(this is serious, and before throwing the stone, think about the meaning of that)

[chalsall]

Originally Posted by LaurV

I wish that DCTF project would never die... (i.e. we would never be able to finish it)

I don't get this... It would mean that the LLTF project isn't working correctly.

I know this was a joke; I'm probably just being slow....

[chalsall]

Originally Posted by Mark Rose

I found a few dozen that were assigned for P-1/LL in September and October with no P-1/low TF, probably when we ran out of exponents when a couple users requested big batches. I should have everything up to 105M done by the end of the week.

Thanks for doing this.

And yes, a few candidates "slip through"; I'm not entirely sure why. Certainly there have been cases where users request a huge number of candidates (often manually, and then often never completing them). Thankfully we're slowly building up enough of a buffer in all the different categories such that this is becoming less of an issue.

In other cases it seems like people request specific candidates not yet appropriately TF'ed / P-1'ed. Again, I have no idea why.

[Mark Rose]

Originally Posted by chalsall

Thanks for doing this.

And yes, a few candidates "slip through"; I'm not entirely sure why. Certainly there have been cases where users request a huge number of candidates (often manually, and then often never completing them). Thankfully we're slowly building up enough of a buffer in all the different categories such that this is becoming less of an issue.

In other cases it seems like people request specific candidates not yet appropriately TF'ed / P-1'ed. Again, I have no idea why.

Yeah, there were a bunch in the 76.9M and range where it appears the exponents ran out.

The exponents above 80M are all look like people doing work ahead of the waves for whatever reason. There's actually very little of it.

I think it would be neat if we could collapse the LLTF work into a single front. I reckon if all the people doing DCTF were to work on LLTF for a few months we could accomplish that.

[airsquirrels]

Well, give us another few weeks to remove the DCTF front and we can do that :)

Maybe the people responsible for the mfakto kernel tuning will squeeze some more out of the 74,75 range performance. My GhzDay/Day output on those is definitely not as nice as 71->72.

[chalsall]

Originally Posted by airsquirrels

Well, give us another few weeks to remove the DCTF front and we can do that :)

Sweet.

Originally Posted by airsquirrels

Maybe the people responsible for the mfakto kernel tuning will squeeze some more out of the 74,75 range performance. My GhzDay/Day output on those is definitely not as nice as 71->72.

That might be problematic...

Nvidia's GPUs are very different from AMD's GPUs.

Both are very good, but they do things in different ways.

Edit: I finally understand LaurV's wish.

[LaurV]

Originally Posted by chalsall

I know this was a joke; I'm probably just being slow....

You are not, but you didn't read the white text...

Yes, indeed, you understood the joke/wish at the end: what "re-launched" the DCTF project was bringing the GPUs into equation. Otherwise, they were "dead", the right bitlevel was factorized with P95/CPU before attempting the first LL test. Only apparition of the hardware able to TF 200-300 times faster made reviving TF work profitable, for exponents already TF+LL-ed with the CPU. So, assuming we get in the next 2-3 months (before DCTF finishes) some hardware able to TF few more bits deeper, i.e. 4-8 times faster at TF work than the hardware we currently have, then we will have a new wave of DCTF for the exponents currently LL-ed (in 75M and below). My wish that "DC never dies" was just wishing for such hardware. There is no other way to "keep DCTF alive", unless we continuously get better hardware.

[chalsall]

Originally Posted by LaurV

You are not, but you didn't read the white text...

I did actually read the white text. It was included in the quoted text in the HTML editer.

Are you suggesting we go deeper into DCTF?

You've previously argued that we shouldn't, but I would be happy to facilitate that happening.

Please advise.

[airsquirrels]

Well, I guess we should get around to developing TFing ASICs for dedicated bit depth ranges :)

Let's convince the various crypto-currency folks to mine for primes...