Why sign up with GPU to 72?
 

I recently downloaded mfakto and followed the instructions in the readme, requesting a manual exponent via mersenne.org and going to work at it. I then find that you can be assigned work via GPU to 72, but I don't quite understand what the point of doing so is. Why go through the GPU72 website and not mersenne.org? Is it simply the fact that the range reserved for GPU72 is somehow more appropriate? I understand that both ultimately come from and return to PrimeNet, but I don't understand the apparent middle-man here.

Apologies if this is the wrong area to post in, or if I am missing something. Though I have been running Prime95 for years, this is the first time that I have ventured outside of the automatic operation of the program.

Thank you.

No, they're all good questions.

The basic issue is that PrimeNet is a very forgiving system: assignments have huge expiration dates, and you can extend them quite long.

Unfortunately, this means that all the assignments that are more important can sometimes be assigned to someone who takes their time (which is totally fine) but then some people get anxious/twitchy.

One of the side effects is that Trial Factoring exponents to a sufficiently high level was often not done -- people were being assigned LL tests that were 2-3 bits too low. But because LL tests tend to take forever, no one could reserve such an exponent for trial factoring. Whenever such an exponent is "released" (that is, expired due to lack of progress by the assignee), it usually gets immediately reassigned for LL because there's not enough quick TF workers around.

The idea behind GPU272 is to grab these so called "expiries" -- exponents that are well behind the "leading edge wavefront", but that hadn't been sufficiently trial factored. Using a few tricks, it's easy to reserve large quantities of low expiries, which is done by the GPU272 website. It would then assign TF jobs to "trusted" workers, i.e. those who could do the TF very quickly and thus quickly release the exponent back to PrimeNet for LL testing.

That was the initial goal, anyways; in the last few months, GPU272 has more-or-less successfully trial-factored all the exponents behind the wave, and thus the work doesn't need to be done as urgently/quickly.

So at this point, it doesn't really make a difference if you get work through GPU272; the main advantage is that it provides much prettier graphs for its members than PrimeNet does (partially due to a much smaller user base). The exponents involved may also be a bit smaller than you can get for TF from PrimeNet (due to those tricks mentioned) but this is less true recently per the previous paragraph.

The "middleman" came together with the newer and stronger GPUs. As those are many times faster then a normal CPU when doing TF, it makes sense to take some exponents few bits deeper, BEFORE try costly LL tests on them. This will eliminate exponents before trying costly LL tests on them. You can do this directly with PrimeNet (GIMPS) server if you like, but you will (never) get the "right"** assignments, as other people grab them as they become available. You need to continuously stay focused on the gimps database, if you want the "best"* assignments. What GPU-2-72 does, is continuously watching the ranges, due to involved automatic spiders, and give you "best"* assignments.

*) Now, how you define this "best" is up to a long discussion in the forum. Either maximizing chances to find a prime (if you do LL, by selecting the smallest available exponent, etc), or compacting this or that front.

**) [URL="http://www.mersenneforum.org/showthread.php?p=276748#post276748"]For an analogy[/URL], you can still use your harddisk (or windows) directly with the OS drivers, and do defragmentation of the hdd (or registry clearance) every 3-6 months, because the system is slower and slower, or you can install better hdd drivers (respective a clever registry first aid program) which will occupy the disk in a cleverer manner (respectively store/clean the registry on the way), keeping your system in shape.

edit: @Dub: you beat me to it as I was looking for the link :razz:
edit2: as Dubslow pointed out, GPU72 reserves low exponents as [B]LL [/B]assignments from primenet, but gives you them as [B]TF[/B] assignments. When you do your job (fast) and report your result (to primenet!), gpu72 detects it automatically, and release the LL assignment. So, now the exponent is available for LL, but is better TF-ed. This was the initial goal for GPU72. Many exponents were eliminated (by founding factors) and a lot of (LL) work saved by TF-ing. Not all people understand that gpu72 has his own niche, it is not a competition for primenet/gimps.

I think both of you about perfectly answered my question. Thank you! I very much appreciate the detail.

Now, I suppose I should sign up. :D

You are welcome. Signing up is up to you.
No need to say that gpu-72 (as it is now, assuming it will not evolve to something else in the future) will become completely futile if and when P95 will become GPU-aware, a thing which many of us "hope" to get, but we will not get, as George is very busy, and Oliver did a wonderful job with mfaktX, so many voices say to keep them as separate tools/paths.

[QUOTE=LaurV;320704]You are welcome. Signing up is up to you.
No need to say that gpu-72 (as it is now, assuming it will not evolve to something else in the future) will become completely futile if and when P95 will become GPU-aware, a thing which many of us "hope" to get, but we will not get, as George is very busy, and Oliver did a wonderful job with mfaktX, so many voices say to keep them as separate tools/paths.[/QUOTE]
I'd kinda like the GPU to be incorporated into P95 as it would give better automation to the process. In addition, you'd have more people playing with their GPUs if it was just a few click to set up.

[QUOTE=LaurV;320704]You are welcome. Signing up is up to you.
No need to say that gpu-72 (as it is now, assuming it will not evolve to something else in the future) will become completely futile if and when P95 will become GPU-aware, a thing which many of us "hope" to get, but we will not get, as George is very busy, and Oliver did a wonderful job with mfaktX, so many voices say to keep them as separate tools/paths.[/QUOTE]

It's not that some people want them separate, it's rather that they have completely different development and release cycles. So every time Oliver updated mfaktc, there would be a new version of Prime95.

Maybe Prime95 can have a hook for mfaktX. I have read somewhere that Prime95 has a hook to do stage-two ECM with an external program…

It would also be nice to have mfaktX available in the form of NT service. I am not sure if it is possible, I have, again, read somewhere that consumer versions of Nvidia GPUs have something in them specifically that prevents them to be used from a service.

Can somebody confirm if the above is true?

Thank you.