Double checking, quail factoring, trick question

[Mark Rose]

I understand the the idea with TF is to do it before running LL. Makes sense to me.

I'm a little confused as to when DCTF is taking place. Is it only taking place for exponents that have never had TF done, before running it through LL a second time? So at some point it would be possible to completely finish the DCTF work?

[kladner]

DCTF is run before LLDC to trial factor to higher levels than before. I think GPUs made these levels much easier to reach than with the CPU TF which was likely used the first time around. Finding a factor saves a DC.

[manfred4]

In the time when the now DC candidates were at their first time tests, there were no GPU programs for helping GIMPS, so the TF was done to a level, that finding a factor in the highest possible bitlevel and finishing a LL test needed about the same time.
But now with the GPUs and their much higher performance in TF compared to LL, their sweetspot is about two to three bitlevel higher, so that is what GPUs should do at the time.

When the DC wave reaches about 5xM everything there already got TF'ed by the GPUs for first time tests already, so then the DCTF will no longer be there.

But until we are there it will be a few more years...

[Mark Rose]

Thanks for clearing that up :)

[TheMawn]

Is there not also the logic that where we run TF to maybe save the LL test, we can also run TF to save the DC? Or are the optimal limits for LLTF already considering that two tests would be saved?

[Mini-Geek]

Originally Posted by TheMawn

Is there not also the logic that where we run TF to maybe save the LL test, we can also run TF to save the DC? Or are the optimal limits for LLTF already considering that two tests would be saved?

The optimal limits for LLTF do consider that two tests will be saved. This means that only large changes to the optimal TF depth, e.g. because of the advent of TF on GPUs, can make DCTF worth it: what was once sufficient TF to save 2 LLs is no longer sufficient for just 1 LL.

[chalsall]

Originally Posted by Mini-Geek

The optimal limits for LLTF do consider that two tests will be saved. This means that only large changes to the optimal TF depth, e.g. because of the advent of TF on GPUs, can make DCTF worth it: what was once sufficient TF to save 2 LLs is no longer sufficient for just 1 LL.

Yes, James' analysis takes this into account. Basically, we TF for DC candidates appoximately one bit level less than for LL candidates, as only one test will be saved rather than two.

And to answer the implicit question, yes, at some point in the future there will no longer be any DCTF'ing to be done because it would have been covered by the LLTF'ing; but that's a LONG way off (approximately four (4) years).

[lycorn]

Unless... MUCH faster GPUs (or whatever devices suitable for TFing) appear in the meantime.

[chalsall]

Originally Posted by lycorn

Unless... MUCH faster GPUs (or whatever devices suitable for TFing) appear in the meantime.

Except that, most likely, a GPU faster at TF'ing will be correspondingly faster for LL'ing as well, and thus the curves will cross at about the same point.

[LaurV]

... or unless LaurV gets angry with the second table...

(BTW, Chris, I said once but you didn't pay attention, the last 4 cells in the "72" column, of the second table, they have to be white, and not yellow, i.e. 56M to 59M, they have to be DC-ed to 73, and not 72 - the cut point is somewhere at 56.5 even for the lousiest cards)

[chalsall]

Originally Posted by LaurV

(BTW, Chris, I said once but you didn't pay attention, the last 4 cells in the "72" column, of the second table, they have to be white, and not yellow, i.e. 56M to 59M, they have to be DC-ed to 73, and not 72 - the cut point is somewhere at 56.5 even for the lousiest cards)

BTW, LaurV, I don't pay much attention to everything you (or anyone) says.

56M and above are about three years out. I think I have enough time to correct the rendering by that time.

Deal with it. ()