What do we do next?

[chalsall]

Originally Posted by Mark Rose

They are fast at LL/DC, but they're much faster at TF.

But... Again...

This is the Great Internet Mersenne PRIME Search.

As always, every participant is contributing work on a voluntary basis. Thus, they get to do anything they want with their own kit / money / time. I'm simply a facilitator.

It might seem a bit strange that the admin of GPU72 is suggesting people start doing more LL/DC'ing than TF'ing with their GPUs. But that is, actually, "What Makes Sense" if you accept the fundamental goal of GIMPS.

[thyw]

(Most numbers are made up to represent my point. I don't know the exact figures.)
The first priority is LL, and accelerating it by TF/P-1 would be benefical.
Also the chance of a missed prime by bad LL is pretty low, and you cannot make DC and LL go at the same speed, near each other. So this danger will always exist, and reliable computer checking (triple checks, those lists of cpus by bad results) helping it lower.

Exponents are getting bigger and as you can see there are plenty of (as i see it as an outsider - i'm not registered to GPU72) "excess" horsepower to help LL.
Also what's the point of factoring 100M numbers (now) when the LL front will take years to reach that?

And we know that "TF work to do" will be the first to be done/run out - after that either go higher or deeper - (LL-TF) gap expand or shrink. Would you spend that horsepower now for accelerating in 1 year or in 7 years? Also gpus too are getting stronger (don't know the ratio to CPUs), and they are already way ahead of the LL. (Unless you want a second front.)

(About the DC/LL gap over the years:
How did the gap expand/shrink? I mean work ratio for these ranges, not the difference of the exponent. 10M to 30M has different work amount change than 30M to 50M if I understand it correctly.
I think most people(?): slow computer - DC, faster - LL. I don't know the speed difference between these, but i think the DC couldn't catch up unless most people abadon the current LL/TF - which would slow down progress directly or indirectly. And still there are simply too many users have "slow" PCs.)

[Lemonlurker]

Since I'm new to this, I decided to take a factor that gpu72 gave me to do to 76 up to 78 for "fun". That was probably a waste of time, so I'm going to stick to whatever the lowest exponent work it'll give me (73) for now since it takes only between 14-15 minutes.

Also someone I know from a completely different unrelated community has turned out to be top 25 on the gpu72 overall progress charts. It's a small world.

[chalsall]

Originally Posted by Lemonlurker

Since I'm new to this, I decided to take a factor that gpu72 gave me to do to 76 up to 78 for "fun". That was probably a waste of time, so I'm going to stick to whatever the lowest exponent work it'll give me (73) for now since it takes only between 14-15 minutes.

It wasn't a waste of time, exactly. It was what we refer to as "sub-optimal".

The reason for this is there has been very careful analysis (mostly by James; peer reviewed by many) as to exactly where the economic curves cross with regards to maximising the GIMPS effort of finding Mersenne Primes using GPUs. As in, where does it become more efficient to do a LL test on a candidate rather than additional TF'ing *on the same hardware*.

Since there is still so much GPU TF'ing firepower available, and we're so far ahead of the LL'ing effort, there is an argument of taking some of the higher ranges to 77 "bits" instead of "only" 76. Going to 78 was not advisable; but hey, it's your kit / time / energy / money -- and you /could/ have eliminated a candidate which likely will next be handed to a CPU.

It's a simple resource management problem space.

Originally Posted by Lemonlurker

Also someone I know from a completely different unrelated community has turned out to be top 25 on the gpu72 overall progress charts. It's a small world.

Geeks of a feather....

[Lemonlurker]

Currently I have submitted 1215 mhz days to gpu72 and my expected factors found is almost 1. My gpu runs at around 1100 mhz days and the last time I checked CUDAlucas to see how long a DC test would take, it was around one day so I am guessing it would be more or equally productive to do LL or DC testing on my gpu. (Ignoring the fact that the DC work I am given is usually in the 40 millions and the TF work is nearly at 100 million)

The assignments page for gpu72 is set to only TF and P-1 work. Is there a reason we couldn't get gpu LL or DC work from the page like I can attempt to with manual assingment on mersenne.org? Or do you wish to keep the work solely purposed towards factoring? I think I see them in the two disabled work sections at the bottom of the page.

[chalsall]

Originally Posted by Lemonlurker

The assignments page for gpu72 is set to only TF and P-1 work. Is there a reason we couldn't get gpu LL or DC work from the page like I can attempt to with manual assingment on mersenne.org?

There was a time people could get LL and DC work from GPU72 via the web forms. This used a bit of a trick, and a spider, to grab low expired candidates and then hand them out to trusted LL and DC'ers.

But then there was a community effort (early 2014) to optimize the assignment procedures from PrimeNet to avoid unknown entities from being assigned low candidates, and holding up progress and milestones. Please see this page for some details on what kind of credentials are needed for each class of assignment.

My advise is to get LL and DC assignments directly from Primenet to work on with your GPU(s). And I would encourage you to explore doing DC'ing with your GPU(s).

Regardless of what you decide to do, thank you for your cycles!

[kladner]

Originally Posted by chalsall

My advise is to get LL and DC assignments directly from Primenet to work on with your GPU(s). And I would encourage you to explore doing DC'ing with your GPU(s).

Regardless of what you decide to do, thank you for your cycles!

@Lemonlurker: If you want to have DC/LL results show up in GPU72, you should acquire them through Prime95 or mprime with the following in prime.txt, in the [PrimeNet] section:
ProxyHost=gimps.gpu72.com:80
This puts you on the GPU72 radar, and gets you an assignment key from GIMPS. You can then stop Prime95 (important) and copy/cut the most recent assignment from worktodo.txt, save worktodo, restart P95, and paste the assignment into CUDALucas' worktodo.txt.
You can submit the CuLu results via
https://www.mersenne.org/manual_result/