mersenneforum.org > Data Strategic P-1 for DC (requires 24GB or more memory for prime95)
 Register FAQ Search Today's Posts Mark Forums Read

2023-03-26, 16:16   #34
James Heinrich

"James Heinrich"
May 2004
ex-Northern Ontario

2×2,083 Posts

Quote:
 Originally Posted by Mark Rose I'm currently processing with a B1 just under 5M and a B2 between 62M and 78M depending on the machine. I'd aim for a higher B1 than 1.5M.
Quote:
 Originally Posted by kriesel Is that what mprime would optimize the bounds to on your hardware? Those are very different from what I'm seeing
I'm using 1.5 tests saved to ensure that my new B1 is generally similar to the previous B1-only run and I'm also not getting even 1M for B1, and yet my B2 is 287M:
Quote:
 Optimal P-1 factoring of M107773069 using up to 51200MB of memory. Assuming no factors below 2^76 and 1.5 primality tests saved if a factor is found. Optimal bounds are B1=903 000, B2=287 732 000 Chance of finding a factor is an estimated 6.05%
edit: looking at recent results (e.g.), it seems that Mark confused "just under 5M" with "just under 500k"

2023-03-26, 16:27   #35
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

7,537 Posts

Quote:
 Originally Posted by James Heinrich No reason you can't do a higher B1 in Prime95 too. Feel free to adjust the tests-saved upwards from 1 to something a bit higher (e.g. 1.3) to get the B1 you want.
Why not let prime95 do its optimization undistorted by misleading inputs? The most likely scenario is the DC is a PRP with proof generation, so only one more primality test will occur after these retries of P-1 to better bounds with a more efficient algorithm for exponents preselected with really poor initial P-1 bounds.

2023-03-26, 21:09   #36
Mark Rose

"/X\(‘-‘)/X\"
Jan 2013

23·137 Posts

Quote:
 Originally Posted by kriesel Is that what mprime would optimize the bounds to on your hardware? Those are very different from what I'm seeing, and your reported limits don't seem to me to capitalize much on the increased stage 2 efficiency; I'm seeing optimization to sub-1M stage 1 and >150M stage 2 (both stages in prime95, ~60G allowed ram): UID: Kriesel/emu, M106970327 completed P-1, B1=563000, B2=176294580, Wi4: ... UID: Kriesel/emu, M117360923 completed P-1, B1=828000, B2=266878920, Wi4: ...
I miscounted digits: I'm closer to half a million, not 5 million

Some recent results. The first machine has 23 GB allocated and the rest 30 GB.

Code:
worktop		101907739	NF-PM1	2023-03-26 20:44:51	B1=473000, B2=62802600
leonardo	106941547	NF-PM1	2023-03-26 20:44:47	B1=488000, B2=77781900
raphael		106958389	NF-PM1	2023-03-26 20:41:41	B1=488000, B2=77781900
michelangelo	106952711	NF-PM1	2023-03-26 20:04:03	B1=488000, B2=77781900
splinter	106964681	NF-PM1	2023-03-26 20:02:26	B1=488000, B2=77781900

 2023-03-27, 14:49 #37 Mark Rose     "/X\(‘-‘)/X\" Jan 2013 C4F16 Posts Took 229: Pfactor=1,2,100400879,-1,77,1 through Pfactor=1,2,100978307,-1,78,1 Leaving 790. Last fiddled with by Prime95 on 2023-03-27 at 19:43 Reason: List updated
 2023-03-27, 17:37 #38 storm5510 Random Account     Aug 2009 Not U. + S.A. A3E16 Posts As a general rule, I run P-1's on a GPU using gpuOwl. A guess is that your B1 and B2 are too large to run on a GPU. Off hand, I don't know of a GPU with 32 GB of VRAM. Perhaps down the road sometime...
2023-03-27, 17:49   #39
James Heinrich

"James Heinrich"
May 2004
ex-Northern Ontario

104616 Posts

Quote:
 Originally Posted by storm5510 Off hand, I don't know of a GPU with 32 GB of VRAM.
The relatively-common RTX 4090 has 24GB, but higher-end cards can certainly have more, e.g. Radeon PRO W6800 = 32GB, A40 = 48GB, A100 = 80GB.

2023-03-27, 21:25   #40
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

11101011100012 Posts

Quote:
 Originally Posted by storm5510 As a general rule, I run P-1's on a GPU using gpuOwl. A guess is that your B1 and B2 are too large to run on a GPU. Off hand, I don't know of a GPU with 32 GB of VRAM. Perhaps down the road sometime...
Not sure which user "your" refers to. In my experience, gpuowl is limited on exponent by available fft length in stage 1, and more so by available GPU ram in stage 2 (also true of CUDAPm1). Limits in stage 2 vary depending on gpuowl version, because the minimum number of buffers required for stage 2 varies between versions. Mihai has not coded the polynomial multiply in gpuowl, because of the usual GPU ram capacity being insufficient to show much advantage. But there are GPUs with larger than 16GiB; AMD's MI60 is another example with 32 GiB.
On modest exponents, large bounds could be run in gpuowl, and even up to 1G exponents, on 16GiB GPUs, but run times would be long.
See https://www.mersenneforum.org/showpo...5&postcount=17

 Similar Threads Thread Thread Starter Forum Replies Last Post Uncwilly Marin's Mersenne-aries 25 2023-03-24 01:37 jocelynl1204 Data 11 2021-03-02 16:32 Madpoo Marin's Mersenne-aries 1841 2019-07-16 03:30 graeme Puzzles 7 2003-08-19 20:40 NickGlover Puzzles 18 2003-07-26 01:10

All times are UTC. The time now is 23:36.

Mon Mar 27 23:36:01 UTC 2023 up 221 days, 21:04, 0 users, load averages: 1.05, 0.85, 0.90