Factoring to 87 bits

[petrw1]

Originally Posted by lavalamp

so probably 2 months or so to hit 91 bits.

25 days for a 2080Ti.
Maybe 15 if you get a 3090???

[clowns789]

I went ahead and took the liberty of reserving that exponent to 91 bits. It's a comparatively slow 100 days on my GTX 1080, but it will be good to have an exponent fully factored. I will check my local electronics store for a 3060 Ti tomorrow morning, which should give a faster ETA at an affordable price.

[lavalamp]

Originally Posted by clowns789

it will be good to have an exponent fully factored

Perhaps ... but I still hope you find a factor.

[clowns789]

Originally Posted by lavalamp

Perhaps ... but I still hope you find a factor.

Perhaps I should say fully factored...to the optimal 91 bit depth! A factor would at least avoid the arduous P-1 and PRP tests, but those will also be an interesting challenge.

I was lucky enough to get a 3060 Ti this morning, despite there being over 100 people in a line that started forming the day before! I will install that and see what the new ETA is.

[clowns789]

It isn't reflecting on mersenne.ca yet, but I factored 3321928171 to 91 bits with no factor. Anyone with a Radeon VII Pro who is willing to crunch a P-1 and PRP test is good to go!

[lavalamp]

Now you can get to mining ETH with that thing.

[clowns789]

Originally Posted by lavalamp

Now you can get to mining ETH with that thing.

Perhaps once the lousy Texas power grid gets its act together!

[kriesel]

Originally Posted by clowns789

I factored 3321928171 to 91 bits with no factor. Anyone with a Radeon VII Pro who is willing to crunch a P-1 and PRP test is good to go!

Gpuowl would require extension to larger fft lengths, to attempt P-1 or PRP. Mihai has not posted a commit on GitHub for gpuowl since March 24. Amount of ram on GPUs we can afford is an issue; gpuowl at ~1Gbit Mersennes takes ~140 passes in P-1 stage 2 with 21 buffers, so ~3.32Gbit would be ~450 passes with ~6 buffers, on a 16GiB Radeon VII GPU or similar (Tesla P100).

2021-08-03 04:13:45 asr2/radeonvii2 990004423 P2(33M,990M) D=210, nBuf=21
2021-08-03 08:38:46 GpuOwl VERSION v7.2-53-ge27846f
2021-08-03 08:38:46 config: -user kriesel -cpu asr2/radeonvii2 -d 2 -maxAlloc 14G -proof 10 -use NO_ASM -autoverify 10 -block 1000 -autoverify 11
2021-08-03 08:38:46 device 2, unique id ''
2021-08-03 08:38:46 asr2/radeonvii2 990004423 FFT: 56M 4K:14:512 (16.86 bpw)

For -maxAlloc 15G nBuf=24 at 0.99G, stage 2 runtime ~10 days in v7.2-53 to unusually high bounds; probably ~4 days in v6.11-380. So for ~1 Gdigit, if Gpuowl were extended, ~ 1.5 - 4 months for P-1 stage 2 on a Radeon VII.

The alternative is Mlucas v20.0 with its new P-1 capability, which I'm looking into. Run times for P-1 would be even longer, perhaps 4-8 months on a Xeon Phi 7250.

Mlucas v20.0 or earlier does not have PRP proof generation yet. That's what Ernst plans to add next IIRC. PRP test run times would be very long (decades on my hardware).

[kriesel]

Originally Posted by clowns789

I factored 3321928171 to 91 bits with no factor.

That's the right place to stop for your GTX1080. RTX20xx, GTX16xx, RTX30xx, probably 92 bits. (Mfakto max 92 bits; Mfaktc 95 max)

[clowns789]

Originally Posted by kriesel

The alternative is Mlucas v20.0 with its new P-1 capability, which I'm looking into. Run times for P-1 would be even longer, perhaps 4-8 months on a Xeon Phi 7250.

Interesting, I'm glad Ernst added that feature. If necessary, I have access to a dual hex Ivy Bridge server with 128 GB ECC RAM for testing. Not quite as many cores as a Xeon Phi 7250, but higher clock speeds and should at least verify that it's doable.

Originally Posted by kriesel

That's the right place to stop for your GTX1080. RTX20xx, GTX16xx, RTX30xx, probably 92 bits. (Mfakto max 92 bits; Mfaktc 95 max)

Interesting as well! I have a 3060 Ti, but going to 92 bits would take about 10 weeks. We could test P-1 to ensure it runs on Mlucas or other software and then go to 92 bits if there is hardware available to us that can complete P-1 in a reasonable time frame thereafter.

[kriesel]

Originally Posted by clowns789

If necessary, I have access to a dual hex Ivy Bridge server with 128 GB ECC RAM for testing. Not quite as many cores as a Xeon Phi 7250, but higher clock speeds and should at least verify that it's doable.

I have a 3060 Ti, but going to 92 bits would take about 10 weeks. We could test P-1 to ensure it runs on Mlucas or other software and then go to 92 bits if there is hardware available to us that can complete P-1 in a reasonable time frame thereafter.

The TF will be useful and can be done now as preparation. I've already eliminated 2 exponents that had had substantial prior trial factoring applied. TF up to 92 bits can be pursued on multiple exponents in parallel, with multiple GPUs, as I am doing. It's possible P-1 would find a factor in stage 1, or stage 2, eliminating an exponent. I feel the goal of P-1 factoring in OBD would be more about producing PRP-ready candidates than eliminating them. (No-factor after both adequate TF and adequate P-1.) There's no rush; PRP on OBD takes years or decades on currently available consumer hardware including used Ivy Bridge servers such as the dual-12xHT I have.

I've been familiarizing with Mlucas V20.x and finding and reporting bugs and running P-1 on "smaller" exponents.
Like for other applications, Mlucas run-time scaling must be determined, and hardware coefficients determined, as preparation.

Try doing the same. Including on the server you mentioned if you can have very extended and heavy use of it. (majority of cycles and ram occupancy for a year)

There's not much point in attempting an OBD P-1 run until those preparations are accomplished and fixes for most issues are provided for end users, especially the ram constraint I found. I expect P-1 run times will be ~ several months to a year or more per attempt for good bounds depending on hardware and other loading.

Also, there's no computing credit given or server reporting mechanism for P-1 results (factor or no-factor) for exponents > 1G.