GPU72 plans post-announcement

[LaurV]

Originally Posted by ET_

Would P-1 with just B1=1,500,000 (and no B2) be useful for exponents with no P-1 at all?

Of course! Remember that the B1 we currently do in this range is about 600k [edit: about 2.5% chances to find a factor for a 65M exponent], and only with lots of memory or cheating the worktodo lines (to say there are 5 or 10 saved LL tests, to force P95 to use a higher B1 limit) one can go to 700k or 750k [edit: about 2.8% chances to find a factor for a 65M exponent]. Doing P-1 with B1=1M5 (double than the default [edit: about 3.6% chances to find a factor for a 65M exponent]) it is OF COURSE! VERY! helpful. I consider that you are talking about doing P-1 to that limit, using the GPU (the new trending story in the parallel thread). With the CPU it will take too much time.

[ET_]

Originally Posted by LaurV

I consider that you are talking about doing P-1 to that limit, using the GPU (the new trending story in the parallel thread). With the CPU it will take too much time.

I was kind of thinking about that, yes... but not yet:

- There is no API/WEB communication ready for the program
- There is still no understood file-format
- George has more important things to do than manually update results
- There is a choice to have a CUDA P-1 Stage2 in a few weeks

I was just checking around how a Stage1-only could be useful.

Luigi

[axn]

Originally Posted by ET_

- There is no API/WEB communication ready for the program
- There is still no understood file-format
- George has more important things to do than manually update results

All three can be addressed if the CUDA Stage 1 output can be saved in a P95-compatible save file. Then the stage 2 can be done and results reported by P95.

Originally Posted by ET_

- There is a choice to have a CUDA P-1 Stage2 in a few weeks

Which may or may not run on most cards. Most cards can't spare enough memory for decent P-1 stage 2 performance. Alternately, you can still do stage 2 with severely limited performance (almost 4x slower, relatively speaking, than P95).

Originally Posted by ET_

I was just checking around how a Stage1-only could be useful.

I don't have hard numbers supporting this, but gut feel tells me that a minimal P-1 run with minimal memory and both stages run would be better than Stage-1 only runs (in terms of project-time-saved as well as factors-found-per-unit-time).

[ET_]

Originally Posted by axn

All three can be addressed if the CUDA Stage 1 output can be saved in a P95-compatible save file. Then the stage 2 can be done and results reported by P95.


Which may or may not run on most cards. Most cards can't spare enough memory for decent P-1 stage 2 performance. Alternately, you can still do stage 2 with severely limited performance (almost 4x slower, relatively speaking, than P95).


I don't have hard numbers supporting this, but gut feel tells me that a minimal P-1 run with minimal memory and both stages run would be better than Stage-1 only runs (in terms of project-time-saved as well as factors-found-per-unit-time).

Thank you axn.

I'm sure those issues will be competently addressed in time, I just didn't want to press the developers to complete their work, nor George to manually update new P-1 results.

Luigi