Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Generalized Fermat Prime Search :
Simple explanation for Genefers and task size?
| Author |
Message |
|---|
|
|
|
I'm not a mathematician. I've got a degree in Physics, so I can do some maths, but not to the level of this project. Keeping that in mind, can someone provide dumbed down answers?
I know the Genefer 20 tasks are much bigger than the Genefer 15 tasks, there's a 2^20 or 2^15 somewhere in the range you're looking in.
My questions are:
1) Why are the tasks monumentally different sizes? Is each task looking for a single prime number, so the 20s take much longer than the 15s? If they're looking for several, can they not be split into smaller tasks?
2) Why aren't there massively less primes to look for in the 15 so it should be finished by now? And on the main page, we see bigger queues for the smaller genefers.
3) Is it more important to find huge primes, or to fill in the gaps further down so the set is complete?
4) Is this all just for fun or does cool physics and stuff come out of it? I know smaller primes are useful, but what do the absolutely massive ones get used for? | |
|
 Michael Goetz Volunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
                               
|
I'm not a mathematician. I've got a degree in Physics, so I can do some maths, but not to the level of this project. Keeping that in mind, can someone provide dumbed down answers?
I know the Genefer 20 tasks are much bigger than the Genefer 15 tasks, there's a 2^20 or 2^15 somewhere in the range you're looking in.
My two questions are:
1) Why are the tasks monumentally different sizes? Is each task looking for a single prime number, so the 20s take much longer than the 15s? If they're looking for several, can they not be split into smaller tasks?
2) Why aren't there massively less primes to look for in the 15 so it should be finished by now? And on the main page, we see bigger queues for the smaller genefers.
Great questions, Peter. Thanks for asking.
1) Why are the tasks monumentally different sizes?
The processing time for a number, assuming all else is equal, is roughly proportional to the square of the number of digits in the number. As it turns out, when you go from GFN-15 to GFN-16, the number of digits doubles. That means the processing time quadruples. So GFN-16 takes 4x as long as GFN-15, GFN-17 takes 16x as long, GFN-18 takes 64x as long, GFN-19 takes 256x as long, GFN-20 takes 1024x as long, GFN-21 takes 4096 times as long, and GFN-22 takes 16,384 as long as a GFN-15.
Is each task looking for a single prime number, so the 20s take much longer than the 15s?
Yes, at PrimeGrid, each task does a single number.
Why aren't there massively less primes to look for in the 15 so it should be finished by now?
There's believed to be an infinite number of primes at each GFN level. However, they are much easier to find at the lower GFNs. The projects can go on forever, so long as the software is able to handle the numbers in questions. The software does have limits, however, the developer of Genefer, the program we use to test GFNs, is actively maintaining the software and has been updating Genefer so that we can continue using it as the numbers grow larger.
And on the main page, we see bigger queues for the smaller genefers.
Correct. Those are not the total number of remaining tests. There's an infinite number of possible tests. It's not even the total number of available tests. It's merely what we have loaded into the server right now. We keep the number of loaded tasks low so that the database is a manageable size. There's more tests loaded for the smaller GFNs because they run much quicker, so we need a bigger supply of them to service customer requests for more work.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
It's b^15 and b^20.
1) Each increase in exponent requires some multiple of calculation (4x? Can't quite remember). The number of digits involved in each are also monumentally different. 15 is 500k, 20 is 6.6M which increases resource requirements and therefore time. Each WU is for one prime candidate, no task splitting.
2) While the number of primes per exponent is infinite, the density of primes is greater at lower exponents. The range of search is absolutely enormous (many millions of tasks even after sieving) thanks to the great programming efforts of the PG devs and is being constantly expanded so we can keep searching.
Lower GFNs complete much faster and therefore more tasks need to be available for users than larger ones that take much longer.
3) Really depends on the individual's preference. All the primes will eventually be found for a given available searching range. If any GFNn were to end, then more resources would go to looking at other values, etc.
4) These aren't useful for cryptography given their special and known nature, and nothing special physics wise has been found, yet. I do it for the fun of finding primes and hoping to help solve the conjectures. And who wouldn't love to have their name attached, even if briefly, to the largest prime of a certain type (or ever) known?
____________
Eating more cheese on Thursdays. | |
|
|
|
|
Thanks Michael for the detailed answers. That makes a lot more sense to me now. I think you answered before I edited the post and added two more questions. Does the project itself have an opinion on which are the most important ones to search for? And is there anything useful to come from these numbers - presumably any correlation of things in maths is asking for a eureka moment and the furthering of everything in general. Somebody once said to me "maths is the basis of physics is the basis of everything".
Grebuloner: I picked the smallest genefers for 3 reasons - 1) to "tidy up" the lower numbers (I go onto Private GFN for that too), 2) because my multiple GPUs on multiple computers fly through them which looks cool on my Boinctasks list spread across 3 monitors, and 3) because I use antique GPUs which often screw up sometimes needing dismantled and repaired [1]. It's very irritating to have wasted a few days computation on a corrupted genefer extreme.
To both of you, are the numbers in a range infinite or not? You seem to disagree.
[1] Broken/old/unreliable GPUs accepted greatfully! I like repairing stuff! I'm in the UK if anyone has stuff they don't want.... | |
|
|
|
It's b^15 and b^20.
Do you mean b^32768 respectively b^1048576? /JeppeSN | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
Does the project itself have an opinion on which are the most important ones to search for?
We would like to find a GFN-21. None are yet known. So that's a priority.
Likewise, we would like to find a GFN-22. There's none of these that have been discovered yet either. It's also a priority.
Finally, DYFL -- essentially GFN-22 but with much higher b values -- is a priority because, if found, it would be the world's largest known prime.
And is there anything useful to come from these numbers - presumably any correlation of things in maths is asking for a eureka moment and the furthering of everything in general.
While some other projects have a mathematical objective, such as proving the Sierpinski conjecture, this one does not. However, there are two goals common to all the projects: educating participants, and advancing software science. The drive to find primes more efficiently has led to several significant software advances. Granted, these advances are primarily useful when searching for primes, but who knows when something will be found to be of use in some other disciple. Who knew that going to the moon would give us Velcro?
It's very irritating to have wasted a few days computation on a corrupted genefer extreme.
That's a large part of the reason we grant so much credit to running long tasks like that.
To both of you, are the numbers in a range infinite or not? You seem to disagree.
Infinite. I assure you there's no disagreement. :) GFNs are of the form B2N+1, so for every GFN, there's another one at (B+2)2N+1. However, how many GFN primes exist is not known, but is believed to be infinite. But that is not proven.
____________
My lucky number is 75898524288+1 | |
|
|
|
It's b^15 and b^20.
Do you mean b^32768 respectively b^1048576? /JeppeSN
Agreed, that's why I said something to the power of 2 to the power of the genefer number. Where my something was b. It's not b^15. It's b^2^15.
| |
|
|
|
While some other projects have a mathematical objective, such as proving the Sierpinski conjecture, this one does not. However, there are two goals common to all the projects: educating participants, and advancing software science. The drive to find primes more efficiently has led to several significant software advances. Granted, these advances are primarily useful when searching for primes, but who knows when something will be found to be of use in some other disciple. Who knew that going to the moon would give us Velcro? I guess.... although that's a bit of a longshot when doing prime searches. For some reason this project attracts a lot more folk than the other maths problems though. I find it way harder to rise up the ranks with this one. When I stick my GPUs on Moo etc, I pass many many people at once.
Infinite. I assure you there's no disagreement. :) GFNs are of the form B2N+1, so for every GFN, there's another one at (B+2)2N+1. However, how many GFN primes exist is not known, but is believed to be infinite. But that is not proven. Do the number of them found slow down as you get to larger numbers? I don't mean take longer to calculate, I mean get further and further apart tending towards there will never be another one?
| |
|
|
|
|
How come in the preferences it says genefer 16 tasks don't take as long as genefer 15s? | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
Do the number of them found slow down as you get to larger numbers? I don't mean take longer to calculate, I mean get further and further apart tending towards there will never be another one?
Yes they do.
However, within a singe "n", GFN numbers grow very slowly, so the rate at which primes become rarer is also very slow.
But when you go from, say, GFN-15 to GFN-16, for example, the primes become about twice as rare, in addition to taking four times as long to test each candidate.
____________
My lucky number is 75898524288+1 | |
|
|
|
How come in the preferences it says genefer 16 tasks don't take as long as genefer 15s?
Generally speaking, a GFN-16 task crunching will take longer than a GFN-15 crunching on the same hardware.
Note that in preferences it says the "Recent average GPU time" is lower for GFN-16 than for GFN-15. This is an average of all GPUs searching for primes recently.
February is the official PG prime hunting month (Tour de Primes - TdP) and lots of people want to find a prime that qualifies for the P badge, and the easier prime to search for is GFN-16.
As so, lots of new and fast GPUs are searching for GFN-16, while older slower GPUs are searching for GFN-15 (which does not qualify for a TdP badge).
This makes the recent average for GPU GFN-16 crunching go down and the opposite for 15.
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
|
|
|
Ah, it would be more helpfull if this showed "time to complete on a (insert a GPU model)" then they could be compared directly to each other and also to your own card. | |
|
|
|
|
I can only guess that the recent average was useful before the fast proof tasks were implemented.
Back then, you'd have to compete with a wingman to return your result fist and get the glory of being the discoverer of a prime, instead of double checker (DC).
Knowing the average crunching time time could help decide if or not to get into that kind of crunching.
Nowadays, except for SGS, AP27 and PPS-Sieve (hope I'm right here), one is always 1st when crunching.
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
|
|
|
The always first system must have vastly sped up the search for primes. | |
|
|
|
|
Indeed !
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
|
|
It's b^15 and b^20.
Do you mean b^32768 respectively b^1048576? /JeppeSN
Oops! Yes. I blame early Sunday morning brain fog.
____________
Eating more cheese on Thursdays. | |
|
|
|
Oops! Yes. I blame early Sunday morning brain fog.
___________
Eating more cheese on Thursdays. Try more Cheese on Sundays.
| |
|
|
|
Thanks Michael for the detailed answers. That makes a lot more sense to me now. I think you answered before I edited the post and added two more questions. Does the project itself have an opinion on which are the most important ones to search for? And is there anything useful to come from these numbers - presumably any correlation of things in maths is asking for a eureka moment and the furthering of everything in general. Somebody once said to me "maths is the basis of physics is the basis of everything".
Grebuloner: I picked the smallest genefers for 3 reasons - 1) to "tidy up" the lower numbers (I go onto Private GFN for that too), 2) because my multiple GPUs on multiple computers fly through them which looks cool on my Boinctasks list spread across 3 monitors, and 3) because I use antique GPUs which often screw up sometimes needing dismantled and repaired [1]. It's very irritating to have wasted a few days computation on a corrupted genefer extreme.
To both of you, are the numbers in a range infinite or not? You seem to disagree.
[1] Broken/old/unreliable GPUs accepted greatfully! I like repairing stuff! I'm in the UK if anyone has stuff they don't want....
My range reference was specifically to Primegrid and the software we run. As noted in other threads as software has been updated, there are technical upper limits to the numbers we can currently test. Ending a range/search might be because of a software limitation (like WW).
The great thing about PG is we have the opportunity to pick our projects for whatever reasons we want. I want to eliminate some k's in the conjectures, but I also love running up the badges so I change my focus around that (and my power bill...). Glad you found yours! :)
____________
Eating more cheese on Thursdays. | |
|
|
|
|
Yes, with the wide scope of Primegrid, it's nice to select a specific part to work hard on. Like the fiddly bits on the Norwegian coast (if anyone gets that possibly misremembered reference).
Maybe I ought to change the colour of all those silver badges. Drat, they're all CPU tasks. It's GPUs I have free at the moment.
My power also heats the house of tropical parrots, so I don't call it a cost, in fact I guess it's a tax deduction (although the UK government seldom asks for many details). Might aswell make the electrons do something useful before they give off heat. And yes I've tried heat pumps, they cost more than they save. | |
|
|
|
The drive to find primes more efficiently has led to several significant software advances. Granted, these advances are primarily useful when searching for primes, but who knows when something will be found to be of use in some other disciple. Who knew that going to the moon would give us Velcro? Or maybe there's a basic part of mathematics we're missing that relates primes, and once you discover it everything will change and you'll get a Nobel Prize.
| |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
Ah, it would be more helpfull if this showed "time to complete on a (insert a GPU model)" then they could be compared directly to each other and also to your own card.
Indeed, it would. But maintaining such statistics would be a nightmare.
Besides, every computer is different in myriad ways, from motherboard chipset to RAM timings. The timings on someone else's computer, even with the same CPU, might be significantly different than the timings on your computer. Even with better statistics you're still left with the unpleasant truth: If you really want to know how fast it will run on *your* computer, you are going to have to test it on *your* computer. Even which task is faster can vary from computer to computer.
We do have a page (https://www.primegrid.com/gpu_list.php) that collects statistics on each GPU model and how fast it runs each of our GPU tasks. Exactly what you want, right? Maybe not so much. There's so much variance in performance between supposedly identical devices that the data is not nearly as useful as you would expect. The sad reality is the best way to get good data is to run your own tests.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
No, what I'm looking for is for one computer (say a test one at Primegrid) to have run one of every task, so you can say a genefer 15 takes 1 minute, a genefer 16 takes 4 minutes, etc.
It may not be exactly 4 times longer on my computer, but it's a good enough estimate if you're looking to pick a subproject that takes x amount of time. And if you named the GPU used for the test, I could decide my card is probably half the speed of that and adjust the numebrs accordingly. | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
No, what I'm looking for is for one computer (say a test one at Primegrid) to have run one of every task, so you can say a genefer 15 takes 1 minute, a genefer 16 takes 4 minutes, etc.
It may not be exactly 4 times longer on my computer, but it's a good enough estimate if you're looking to pick a subproject that takes x amount of time. And if you named the GPU used for the test, I could decide my card is probably half the speed of that and adjust the numebrs accordingly.
Three problems, and the reason I said it's a nightmare. And the reason we don't do it.
The first is that the tasks change, sometimes rapidly. The benchmarks need to be rerun, as needed. It's not a one-time thing. It's an ongoing maintenance issue.
The second problem is that the servers are to busy to spend time running benchmarks; shared cloud computers are unsuitable for running benchmarks, and our personal computers both tend to change, and we have other things we want to do than to run one of everything. That takes a lot of time. Running benchmarks isn't trivial; to insure good data you have to make sure the conditions are consistent. That means the compute must be in an identical condition for each test, preferably otherwise idle. Between CPUs and GPUs, that's a lot of tasks, a lot of time, and a lot of wasted CPU cycles.
The third problem is that different tasks within a sub-project can have significantly different run times.
If it was easy, trust me, we'd have done it.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
I'm sure one of us would volunteer to run one of each app, just to get a general idea. No it won't be 100% accurate, because as you've said every PC is different. But it would give a vague idea of this app takes 15 times longer than this one, and a vague idea of someone with a CPU 3 times faster than the test machine will take a third of the time. A static page could then be made with these figures, so anyone coming along and wishing to run something which will take an hour on their particular GPU can make a pretty good guess.
As for the redundant post, if you edit your post to two spaces and nothing else, it vanishes. Seems to be a thing built into all Boinc forums. Not sure why they don't just have a delete button.... | |
|
|
|
I'm sure one of us would volunteer to run one of each app, just to get a general idea. No it won't be 100% accurate, because as you've said every PC is different. But it would give a vague idea of this app takes 15 times longer than this one, and a vague idea of someone with a CPU 3 times faster than the test machine will take a third of the time. A static page could then be made with these figures, so anyone coming along and wishing to run something which will take an hour on their particular GPU can make a pretty good guess.
Why can't it be you? You seem the most eager for it.
How would you determine relative run times from a single information point? If the benchmark is an AMD Ryzen 2700 at 4 GHz running an LLR2 subproject WU in 1 hour with one thread, how long should my Intel Core i9 10980XE at 4 GHz take to run the same WU? Or my Intel X5675 at 4 GHz?
I would also like to point out (because it's something I do) that it's pretty easy to just go to the top participants or computers pages and just click through the visible hosts/tasks to find systems of the same family or CPU and see how fast they are working through them.
____________
Eating more cheese on Thursdays. | |
|
|
|
Why can't it be you? You seem the most eager for it. I'm quite happy to do so. I never excluded myself. But I'd prefer to know any details the admins wish to give to do it well. And to be sure the information will be used.
How would you determine relative run times from a single information point? If the benchmark is an AMD Ryzen 2700 at 4 GHz running an LLR2 subproject WU in 1 hour with one thread, how long should my Intel Core i9 10980XE at 4 GHz take to run the same WU? Or my Intel X5675 at 4 GHz? cpubenchmark.net gives relative speeds between CPUs. techpowerup gives floating point speeds of GPUs. It may not be precise for primegrid, but it's a damn good estimate.
I would also like to point out (because it's something I do) that it's pretty easy to just go to the top participants or computers pages and just click through the visible hosts/tasks to find systems of the same family or CPU and see how fast they are working through them. Unfortunately that doesn't show how many they are running at a time. On GPUs I run up to 4 tasks simultaneously, which to the server looks like my card is about 3 times slower. for the test, I would run a single task on the GPU, and make sure there was plenty of CPU available to keep it fed.
| |
|
|
|
|
Peter that's actually a great idea and one most hardware reviewers do. Use a standardized test bench to compare all of their games, production work loads, streaming capabilities on so that they can change only one set of parameters when testing. While it would be nice if there was a test bench at an admins house that was a R7 5700x with 16gb of insert brand memory at cl16 at 3xxxmhz and a 3070 gpu on a whatever x570 board. I think it would incur many problems with people wondering why wasn't "amd xxxx gpu picked, why wasn't intel i# picked or why run such fast/slow ram.)" It also would be a tremendous expense as it would be expected to iterate with each generation of hardware.
I think it'd be super neat, but I don't think it'd be practical from an expense and time investment. | |
|
|
|
|
Sorry just read your last comment. Please feel free to start a new thread on your own hardware saying gfn 15 on x gpu after a 24h average ran this long, gfn 16 ran this long and so on. it'd be really cool if you did single thread vs multi thread also for cpu so people can maximize throughput. Like my i7-4770 is best on 1 core per pps but my i5-6600 is better on 2 core because it has 2 less mb of cache | |
|
|
|
Sorry just read your last comment. Please feel free to start a new thread on your own hardware saying gfn 15 on x gpu after a 24h average ran this long, gfn 16 ran this long and so on. it'd be really cool if you did single thread vs multi thread also for cpu so people can maximize throughput. Like my i7-4770 is best on 1 core per pps but my i5-6600 is better on 2 core because it has 2 less mb of cache I'll do it if the admins put up a page for such comparisons. My thoughts are I would run my best machine on just 1 GPU task on 1 GPU, and 1 CPU task on the CPU (with the advised number of threads). It would do 1 of every subproject, one at a time. As you suggested, I could also then try again with different numbers of cores.
For anyone with a different CPU or GPU, they can work out the rough difference in speed, or perhaps add their own test. Or I could make up a list of common ones or place a link to cpubenchmark.net and techpowerup
My best machine is a Ryzen 9 3900XT with a Radeon R9 Fury GPU. I have other machines nowhere near as fast in the CPU department, and all other GPUs are Radeon R9 290X. | |
|
|
|
|
Hey Michael, your sig is too big for https://www.calculator.net/big-number-calculator.html and that's only genefer 19. I told them to upgrade their system :-) | |
|
|
|
|
Who keeps the full list of every single prime ever found? | |
|
|
|
Who keeps the full list of every single prime ever found?
I don't know a place where this can be found.
The links below appear on PG home page:
PrimeGrid Primes by Project, if you select "All primes", currently can show 89 pages of prime findings, with 88,707 primes found.
PrimeGrid AP27 lists APs higher than AP22, with the respective ("small") primes in the AP.
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
Who keeps the full list of every single prime ever found?
No such thing.
There's too many to store, and they're too easy to generate. Nobody bothers recording the little stuff.
If you want the first 2 billion primes, just google it and you'll be able to download them. But you won't find anywhere that has everything.
Most of us have probably looked for it at one time or another, only to end up disappointed.
____________
My lucky number is 75898524288+1 | |
|
|
|
Who keeps the full list of every single prime ever found?
No such thing.
There's too many to store, and they're too easy to generate. Nobody bothers recording the little stuff.
If you want the first 2 billion primes, just google it and you'll be able to download them. But you won't find anywhere that has everything.
Most of us have probably looked for it at one time or another, only to end up disappointed.
But there must be some record of what areas have been searched, or you guys wouldn't know where to look with the tasks you send out. Do you just have records that say "we've found everything between x and y"?
If the primes found are to make some breakthrough in maths, surely there needs to be records of them? For example one of your projects finds arithmetic progressions. | |
|
|
|
|
I take it there's no interest in me doing a speed test then? | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
                
|
I take it there's no interest in me doing a speed test then?
I suppose not. What works best for your system isn't what works best for mine.
What interests me most is how my systems perform, not really how your systems perform.
Case in point. Look at the fastest GPUs list.
Do you really think that a 3070 is 10% faster than a 3080 Ti at GFN-16?
It appears that way in the list because the 3080 Ti is underutilized with 1 task,
so typically 2 tasks are running on the 3080 Ti. Each one takes a bit longer
than the single task on the 3070, but two of them are getting done in that time. | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,606,292
RAC: 174,937
|
Who keeps the full list of every single prime ever found?
No such thing.
There's too many to store, and they're too easy to generate. Nobody bothers recording the little stuff.
If you want the first 2 billion primes, just google it and you'll be able to download them. But you won't find anywhere that has everything.
Most of us have probably looked for it at one time or another, only to end up disappointed.
But there must be some record of what areas have been searched, or you guys wouldn't know where to look with the tasks you send out. Do you just have records that say "we've found everything between x and y"?
If the primes found are to make some breakthrough in maths, surely there needs to be records of them? For example one of your projects finds arithmetic progressions.
We search for primes of specific forms only. We do this because those specific forms are easy to search. Otherwise, searching the huge numbers we search would not be possible.
The key part there is "specific forms". You asked about all primes. We have records for the types of primes we search. That's a very small subset of all primes.
____________
My lucky number is 75898524288+1 | |
|
EA6LESend message
Joined: 4 Feb 21
Posts: 2
ID: 1345478
Credit: 1,932,045,666
RAC: 14,218,471
         
|
I take it there's no interest in me doing a speed test then?
I suppose not. What works best for your system isn't what works best for mine.
What interests me most is how my systems perform, not really how your systems perform.
Case in point. Look at the fastest GPUs list.
Do you really think that a 3070 is 10% faster than a 3080 Ti at GFN-16?
It appears that way in the list because the 3080 Ti is underutilized with 1 task,
so typically 2 tasks are running on the 3080 Ti. Each one takes a bit longer
than the single task on the 3070, but two of them are getting done in that time.
Hi!
This is my post ever here :)
I agree with the OP, we should have a list with best time performed by specific hardware and configuration tricks to achieve it. a lot of us we go with the default settings and don't use the hardware to its full potential.
one thing I can report is that the Core i9-12800k is having its best times with the hyperthreading and e-cores disabled.
Also, a strange thing to report is that PPS (sieve) is a bit faster on a RTX 3080 if the bionic manager windows is selected. if another window is selected I have up to 40 seconds more per task. so, we need some little tricks to post in new post about how to get the best from your hardware. | |
|
Scott BrownVolunteer moderator
Project administrator
Volunteer tester
Project scientist
Send message
Joined: 17 Oct 05
Posts: 2380
ID: 1178
Credit: 17,934,948,750
RAC: 11,009,949
                                               
|
|
Okay, having tested things myself on a wider variety of hardware than most over far more years than most on PG (I have been here more than 17 years), I can say definitively that your hardware speeds, etc. are pretty much irrelevant to just about everybody else. On the face, that makes little sense, so allow me to offer a few explanation examples to demonstrate why that is the case:
1) Different applications have different software resulting in different performance comparison times/ratios. A good example of this is GPU applications. Ampere cards generally perform better than older NVidia GPU families, but that is not a consistent gap. On AP27, which is an OpenCL application, Ampere is MUCH faster than even Turing cards, and older cards are abysmal at it by comparison (e.g., an RTX 3060 substantially outperforms an RTX 2070). On the other hand, on GFN (which uses different aspects of OpenCL), Turing and Ampere cards are not so different in performance (e.g., the RTX 2070 is slightly faster than the RTX 3060). And if we turn to the older PPS sieve application (which is CUDA rather than OpenCL), even the older cards such as Pascal cards (e.g., a GTX 1080Ti) can be relatively competitive with the newer generations. And that is just NVidia...AMD cards have their own quirks in performance comparisons and even just some outright application differences (e.g., PPS sieve is a slower OpenCL application for AMD vs. the CUDA app for NVidia).
2) Different hardware installations can produce vastly different performance numbers. Some easy examples for GPUs include installation of new cards into older boards (e.g., even an old GTX 1080 will perform much slower than it should when installed into a PCIe x16 version 2.0 slot as well as when installed into x4 or x1 slots/risers). Dual-GPU installs also often produce very different performance numbers compared to even the exact same card placed in a single install configuration. But this is also true for CPUs. For example, an i7-8700 with single-channel RAM (1 stick installed) compared to the same i7-8700 with dual-channel RAM can run times more than double those of the latter configuration. In other words, there are literally thousands of permutations for hardware configurations (RAM, Power Supply, PCIe generation, etc.) that make creating a useful universal performance configuration almost impossible.
3) Context matters more than we often think. A frequently cleaned system with cooler ambient temperatures and better system cooling will run substantially faster than a rarely clean system in a hot room with a poorly ventilated case (and all that ignores the option of liquid cooling).
4) Sometimes, BOINC just lies. Well, maybe lies is a bit strong, but it certainly misleads things. For example, in a dual-GPU installation, BOINC sees both GPUs as the same thing...whatever the GPU is in slot 0. So If I have a GTX 750Ti in slot 0 and an RTX 4080 in slot 1, BOINC reports some blazingly fast times for the GTX 750Ti cards in the system. Reverse the install so that the RTX 4080 is in slot 0 and we then have some of the slowest RTX 4080s on the planet reported. One can certainly bypass this by digging down into actual output from applications, but that is a lot of tedious work to create a performance list.
I could probably list some other issues with trying to create a performance lists (e.g., different drivers often equals different performance, LINUX vs. Windows, etc.), but hopefully I have made my point. Trying to create a useful comprehensive performance comparison for PG applications to share with all PG users is a fool's errand (I know, I have tried to work on this a couple of times). One can definitely examine performance in a useful manner for one's own machines, but making a list that works for most everyone is just near impossible.
| |
|
|
|
|
Scott Brown wrote: Okay, having tested ....
Very good, Scott !
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
|
|
I take it there's no interest in me doing a speed test then?
I suppose not. What works best for your system isn't what works best for mine.
What interests me most is how my systems perform, not really how your systems perform.
My test would be to compare each app in Primegrid, nothing to do with my machine. It would be so people could tell one app would take 5 times longer than another. And they can also compare the specs of my machine to theirs and divide the times by three.
Case in point. Look at the fastest GPUs list.
Do you really think that a 3070 is 10% faster than a 3080 Ti at GFN-16?
It appears that way in the list because the 3080 Ti is underutilized with 1 task,
so typically 2 tasks are running on the 3080 Ti. Each one takes a bit longer
than the single task on the 3070, but two of them are getting done in that time.
I wouldn't look in the list to compare GPUs, as I don't know what those people are doing with the GPUs. I would look at the specs for the two cards on techpowerup. If the single precision is double the speed, it'll be about twice as fast in primegrid.
| |
|
|
|
We search for primes of specific forms only. We do this because those specific forms are easy to search. Otherwise, searching the huge numbers we search would not be possible.
The key part there is "specific forms". You asked about all primes. We have records for the types of primes we search. That's a very small subset of all primes.
Ok, so you obviously keep records of what you've found. Others are presumably looking for other specific forms and have records. Is there not anywhere central you all report to to compare notes?
| |
|
|
|
Hi!
This is my post ever here :)
I agree with the OP, we should have a list with best time performed by specific hardware and configuration tricks to achieve it. a lot of us we go with the default settings and don't use the hardware to its full potential.
one thing I can report is that the Core i9-12800k is having its best times with the hyperthreading and e-cores disabled.
Also, a strange thing to report is that PPS (sieve) is a bit faster on a RTX 3080 if the bionic manager windows is selected. if another window is selected I have up to 40 seconds more per task. so, we need some little tricks to post in new post about how to get the best from your hardware.
Yes, I forgot to mention just now, I would also give all the apps a go with different numbers of threads, so people could tell which was most efficient.
| |
|
|
|
Okay, having tested things myself on a wider variety of hardware than most over far more years than most on PG (I have been here more than 17 years), I can say definitively that your hardware speeds, etc. are pretty much irrelevant to just about everybody else. On the face, that makes little sense, so allow me to offer a few explanation examples to demonstrate why that is the case:
1) Different applications have different software resulting in different performance comparison times/ratios. A good example of this is GPU applications. Ampere cards generally perform better than older NVidia GPU families, but that is not a consistent gap. On AP27, which is an OpenCL application, Ampere is MUCH faster than even Turing cards, and older cards are abysmal at it by comparison (e.g., an RTX 3060 substantially outperforms an RTX 2070). On the other hand, on GFN (which uses different aspects of OpenCL), Turing and Ampere cards are not so different in performance (e.g., the RTX 2070 is slightly faster than the RTX 3060). And if we turn to the older PPS sieve application (which is CUDA rather than OpenCL), even the older cards such as Pascal cards (e.g., a GTX 1080Ti) can be relatively competitive with the newer generations. And that is just NVidia...AMD cards have their own quirks in performance comparisons and even just some outright application differences (e.g., PPS sieve is a slower OpenCL application for AMD vs. the CUDA app for NVidia).
Fair enough, but even a vague idea would be useful for a beginner to look at. So they know which app to pick to get done in the timescale they want etc. Or whether 10 cores or 1 core per task is best.
3) Context matters more than we often think. A frequently cleaned system with cooler ambient temperatures and better system cooling will run substantially faster than a rarely clean system in a hot room with a poorly ventilated case (and all that ignores the option of liquid cooling).
Things don't slow down with heat unless they reach the temperature limit. People usually notice that.
4) Sometimes, BOINC just lies. Well, maybe lies is a bit strong, but it certainly misleads things. For example, in a dual-GPU installation, BOINC sees both GPUs as the same thing...whatever the GPU is in slot 0. So If I have a GTX 750Ti in slot 0 and an RTX 4080 in slot 1, BOINC reports some blazingly fast times for the GTX 750Ti cards in the system. Reverse the install so that the RTX 4080 is in slot 0 and we then have some of the slowest RTX 4080s on the planet reported. One can certainly bypass this by digging down into actual output from applications, but that is a lot of tedious work to create a performance list.
Which is why I said never to use the results list to work out time. We need results from someone performing tests.
I could probably list some other issues with trying to create a performance lists (e.g., different drivers often equals different performance, LINUX vs. Windows, etc.), but hopefully I have made my point. Trying to create a useful comprehensive performance comparison for PG applications to share with all PG users is a fool's errand (I know, I have tried to work on this a couple of times). One can definitely examine performance in a useful manner for one's own machines, but making a list that works for most everyone is just near impossible.
It would give a vague idea for people to start from. Something like:
For GPU x:
gen15 = 3 minutes
gen16 = 12 minutes
gen17 = 48 minutes
For CPU y doing proth prime search:
1 core = 3 hours
2 cores = 2 hours
3 cores = 1.8 hours
| |
|
EA6LESend message
Joined: 4 Feb 21
Posts: 2
ID: 1345478
Credit: 1,932,045,666
RAC: 14,218,471
|
|
maybe a new thread in each subproject where we can post the time it takes for Y processor or X GPU. eventually tricks or optimizations to achieve a shorter time. | |
|
|
|
maybe a new thread in each subproject where we can post the time it takes for Y processor or X GPU. eventually tricks or optimizations to achieve a shorter time.
Perhaps, but someone would need to collate the results together. | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
Fair enough, but even a vague idea would be useful for a beginner to look at. So they know which app to pick to get done in the timescale they want etc.
It would give a vague idea for people to start from. Something like:
For GPU x:
gen15 = 3 minutes
gen16 = 12 minutes
gen17 = 48 minutes
If you haven't noticed already, there are typical run times for each app in the preferences page.
It's a general average of efficient and inefficient setups, and of fast and slow systems.
You should look at these as ballpark runtimes. The relative magnitudes of runtimes between apps will reflect on your runtimes.
So start with small apps, see how your runs compare with the "benchmark", and scale your expectations to the larger apps accordingly.
If you want to optimize your setup, that's all up to you. If you want to know HOW to do it, there are plenty of posts in these forums.
I've just spent 7 days tweaking GFN 16 tasks for CPU and GPU on ONE computer.
One change per day, and recording the throughput (firsts per day).
At the beginning of February I started with 1 task per core (6) and 2 simultaneous GPU tasks, producing 2079 firsts/day.
Then I went to 1 task per hyperthread (12) and 2 GPU tasks: 2138 firsts/day. + 59
Then I added 1 GPU task (3): 2183 firsts/day. + 46
Then reduced 1 CPU task (11): 2185 firsts/day. + 2
Then reduced another CPU task (10): 2197 firsts/day. + 12
Then reduced another CPU task (9): 2170 firsts/day. - 27
Oops I left the Discord page running all day, so I need to test that again for another day.
You get the idea.
In the meantime, I'm not making big changes. This computer found 1 GFN-16 this month, so far.
By the way, this computer has a second graphics card, much weaker.
Using it just to drive the monitor. BOINC doesn't play well with disparate GPUs. | |
|
VatoVolunteer tester
Send message
Joined: 2 Feb 08
Posts: 841
ID: 18447
Credit: 645,278,403
RAC: 549,056
                 
|
|
feel free to start a thread and post whatever you think is interesting or useful.
that's distinct from asking someone else to do busywork that serves no purpose (imho)
____________
 | |
|
|
|
If you haven't noticed already, there are typical run times for each app in the preferences page.
It's a general average of efficient and inefficient setups, and of fast and slow systems.
You should look at these as ballpark runtimes. The relative magnitudes of runtimes between apps will reflect on your runtimes.
So start with small apps, see how your runs compare with the "benchmark", and scale your expectations to the larger apps accordingly.
As has been discussed already, those are meaningless, since they don't take account of some people running one and some people running 4 at a time per GPU. Hence I saw genefer 16 claiming to un in a fraction of the time of genefer 15. I bet they don't take into account how many CPU cores on CPU tasks either.
If you want to optimize your setup, that's all up to you. If you want to know HOW to do it, there are plenty of posts in these forums.
Why should a new user have to trawl through forums? This sort of thing should be on a page of it's own.
I've just spent 7 days tweaking GFN 16 tasks for CPU and GPU on ONE computer.
One change per day, and recording the throughput (firsts per day).
You're willing to mess around that much for a single machine, but won't do a simple collation of other people's results for a web page?
In the meantime, I'm not making big changes. This computer found 1 GFN-16 this month, so far.
By the way, this computer has a second graphics card, much weaker.
Using it just to drive the monitor. BOINC doesn't play well with disparate GPUs.
I find it fine on some machines but not others. My main machine has always had two seperate cards. I want the fastest one in there for games, and I need two as I have 5 monitors, and none of my cards take more than 4. It was Tahiti + Baffin, now it's Tahiti + Nano. Just a few bits of messing around to direct certain apps like Milkyway onto the Tahiti only, since it does proper double precision unlike modern crap.
Why don't you drive the monitor with the main card?
| |
|
|
|
feel free to start a thread and post whatever you think is interesting or useful.
that's distinct from asking someone else to do busywork that serves no purpose (imho)
I'm not asking others to do things they don't have time for. I'm saying I'm willing to do whatever is required. I'll run tests here on my machine(s), others can provide some other data if they wish, and I'll even make the webpage. But unless I know someone will place it on this server somewhere.... | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
Why don't you drive the monitor with the main card? What for? I don't play games with it, and I don't have a 4k or 8k monitor. | |
|
|
|
Why don't you drive the monitor with the main card? What for? I don't play games with it, and I don't have a 4k or 8k monitor.
So you don't need the second card in there.
| |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
Why don't you drive the monitor with the main card? What for? I don't play games with it, and I don't have a 4k or 8k monitor.
So you don't need the second card in there.
Correct, in a way. In fact it was the newer card which was not necessary.
I just wanted a faster GPU for computing. The older card has nothing wrong with it and it still serves a purpose.
On the other hand, to make room for the newer card,
(whether the old card stayed or not wasn't a factor),
I had to remove 4, 12+ years old 1 TB hard disks.
Only regret there is that they still work. | |
|
|
|
Why don't you drive the monitor with the main card? What for? I don't play games with it, and I don't have a 4k or 8k monitor.
So you don't need the second card in there.
Correct, in a way. In fact it was the newer card which was not necessary.
I just wanted a faster GPU for computing. The older card has nothing wrong with it and it still serves a purpose.
It in fact serves no purpose at all, since your new card has a display output.
On the other hand, to make room for the newer card,
(whether the old card stayed or not wasn't a factor),
I had to remove 4, 12+ years old 1 TB hard disks.
Only regret there is they still work.
You could always buy a proper sized case.
| |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
It in fact serves no purpose at all, since your new card has a display output. My monitor doesn't have displayport nor HDMI ports.
I ditched the HDMI-DVI adapter long ago.
HDMI connectors have the poorest resistance to falling out.
On the other hand, to make room for the newer card
You could always buy a proper sized case.
Sure, but I'd rather buy a 32:9 monitor. Every cent counts.
| |
|
|
|
It in fact serves no purpose at all, since your new card has a display output. My monitor doesn't have displayport nor HDMI ports.
I ditched the HDMI-DVI adapter long ago.
HDMI connectors have the poorest resistance to falling out.
Mine don't fall out, they're horizontal, how could they do that?
You could always put the adapter on the other end of the cable.
I think I own about 15 adapters. Anything to anything. VGA/DVI/HDMI/display port in either direction.
On the other hand, to make room for the newer card
You could always buy a proper sized case.
Sure, but I'd rather buy a 32:9 monitor. Every cent counts.
That's a very wide monitor. And two orders of magnitude more expensive than the case.
| |
|
Dave
Send message
Joined: 13 Feb 12
Posts: 3171
ID: 130544
Credit: 2,233,156,498
RAC: 568,654
         
|
Mine don't fall out, they're horizontal
There's an image. | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
|
This is really off-topic now.
Sure, but I'd rather buy a 32:9 monitor. Every cent counts.
That's a very wide monitor. And two orders of magnitude more expensive than the case.
A very wide screen is worth the money in some circumstances.
1) Spending less time to make more money. A very wide screen is excellent for very wide spreadsheets.
2) The computer case isn't on my desk taking up real estate. The monitor is.
3) Some very wide monitors have a picture-by-picture feature to display output from 2 computers on one screen.
A note about HDMI connectors. They are designed to plug in and forget about.
They do not hold up well over the long term with insertion and removal 4 times each week that accompanies hybrid working conditions.
Cheaper connector pins made of copper alloy with tin plating will lose its effectiveness after 10-15 mating cycles.
Whereas, a connector made with quality contacts like beryllium copper or phosphor bronze with thick gold plating can last for 500-1000 cycles. https://www.amphenol-cs.com/connect/learn-more-about-connector-mating-cycles.html
So you can buy expensive cables that hold up to such wear for a few years,
but you have no control over the selection of materials in the computer's or monitor's connector.
When device's connector is worn out, you'll be spending a lot more than the price of a cable.
In hybrid working conditions, a computer's connector suffers the brunt of mating cycles.
Better hope they used the $0.45 HDMI connector rated for 10,000 mating cycles. | |
|
|
|
This is really off-topic now. It's my topic and I don't care. It's just natural conversation drift, it's all part of becoming human. Congratulations.
A very wide screen is worth the money in some circumstances.
1) Spending less time to make more money. A very wide screen is excellent for very wide spreadsheets.
2) The computer case isn't on my desk taking up real estate. The monitor is.
3) Some very wide monitors have a picture-by-picture feature to display output from 2 computers on one screen. Cheaper to get two normal ones. You do realise you can connect multiple monitors? I have 5.
A note about HDMI connectors. They are designed to plug in and forget about.
They do not hold up well over the long term with insertion and removal 4 times each week that accompanies hybrid working conditions. Why would you plug and unplug often?
Cheaper connector pins made of copper alloy with tin plating will lose its effectiveness after 10-15 mating cycles. Incorrect. I use cheap connectors and have had some in and out over 100 times. Perhaps you meant 10-15 thousand? All connectors can handle thousands of connections.
When device's connector is worn out, you'll be spending a lot more than the price of a cable. Which is why I laugh at Iphone users. The bit which wears out is the springy bit, which is on the Iphone! On Android, it's on the cable. A bit cheaper than the phone, which in their case will report you to Apple for the criminal offence of opening it.
In hybrid working conditions, Never heard of that expression, what does it mean?
| |
|
|
|
Mine don't fall out, they're horizontal
There's an image.
https://www.healthyandnaturalworld.com/best-exercises-to-firm-and-lift-your-breasts/
Hang on, I'm male. Wrong website. | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1140
ID: 55391
Credit: 1,022,871,643
RAC: 1,717,262
|
A very wide screen is worth the money in some circumstances. Cheaper to get two normal ones. You do realise you can connect multiple monitors? I have 5.
I already have 2 monitors. I want to replace them both.
A 49-inch UWQHD monitor would be good enough. I'm not looking for an IMAX experience in my face.
Cheaper connector pins made of copper alloy with tin plating will lose its effectiveness after 10-15 mating cycles. Incorrect. I use cheap connectors and have had some in and out over 100 times. Perhaps you meant 10-15 thousand?
All connectors can handle thousands of connections.
10 to 15. I didn't make this up. It's a direct quote from the web site of Amphenol, the connector manufacturer.
When device's connector is worn out, you'll be spending a lot more than the price of a cable. Which is why I laugh at Iphone users. The bit which wears out is the springy bit, which is on the Iphone! On Android, it's on the cable.
A bit cheaper than the phone, which in their case will report you to Apple for the criminal offence of opening it.
I'm talking about real computers with teraFLOP GPUs, not the mass surveillance devices that masquerade as phones.
But I don't mind about my employer's computer, it's replaced every 3 to 5 years as a matter of policy.
That's a lot of plugging and unplugging video cables unless you have a laptop docking station.
Generally there is a docking station with 2 monitors at the office, but you have to supply your own
external monitor at home. I opted for a TFF PC instead of a laptop, it's less bulky to carry around.
In hybrid working conditions, Never heard of that expression, what does it mean?
Hybrid work is an arrangement in which an employer allows employees to work
remotely part of the week and to work in the office part of the week.
It works well when everyone is away from the office on the same days of the week.
That's a new concept for most people, forced upon employers by the pandemic.
They had difficulty recruiting knowledge workers to work at the office every day,
because candidates demanded remote work for at least part of each week.
Hybrid work became a company-wide standard work arrangement, and
I haven't needed to go to the office for a full week for nearly 3 years.
You haven't heard about The Great Resignation? (Wikipedia)
Part of it was people who liked working at home so much, when they were required
to return to the office, they quit to work for companies offering full-time remote work.
I've never seen knowledge workers wielding so much power over employers.
In response, the tech industry is laying off people by the thousands recently,
partly to control costs during a recession (that's the excuse), partly because
productivity is expected to rise rapidly due to AI assist,
and partly to show workers who is boss. | |
|
|
|
I already have 2 monitors. I want to replace them both.
A 49-inch UWQHD monitor would be good enough. I'm not looking for an IMAX experience in my face.
The cost of a cheapo full tower case isn't much compared to that.
10 to 15. I didn't make this up. It's a direct quote from the web site of Amphenol, the connector manufacturer.
Amphenol don't make crap, it must be a typo.
I'm talking about real computers with teraFLOP GPUs, not the mass surveillance devices that masquerade as phones.
Indeed, the police have far too much power with phones. Although I did read of a library in a university making good use of phone tech. They counted roughly how many students were in a library room etc by picking up bluetooth/wifi signals, so every other student could use an app to see which areas were busy before they bothered going there for a quiet study.
But I don't mind about my employer's computer, it's replaced every 3 to 5 years as a matter of policy.
Things should be replaced when they need to be, not because there's spare budget which must be spent this year.
That's a new concept for most people, forced upon employers by the pandemic.
Pandemic indeed. The reactions by the government caused more damage than the virus. Now they're at it again with "bird flu".
I haven't needed to go to the office for a full week for nearly 3 years.
There's no need to waste time and money commuting for most work. Of course some stuff is physical and hands on, you can't be a bricklayer from home.
You haven't heard about The Great Resignation? (Wikipedia)
Part of it was people who liked working at home so much, when they were required
to return to the office, they quit to work for companies offering full-time remote work.
Or decided not to work at all. You're looking at one. Although I am trying to breed parrots.
productivity is expected to rise rapidly due to AI assist,
No matter how much technology we get, we still seem to have 40 hour weeks. Something is up. | |
|
ReggieVolunteer moderator
Project administrator
Volunteer tester
Project scientist
Send message
Joined: 10 May 14
Posts: 230
ID: 311759
Credit: 207,728,458
RAC: 81,981
        
|
|
I'm locking this. The last dozen or so messages have been completely unrelated to the original questions and have no place in the "Generalized Fermat Prime Search" forum. Please keep threads on topic. | |
|
Message boards :
Generalized Fermat Prime Search :
Simple explanation for Genefers and task size? |
