Info PrimeGrid Challenges 2025

[TennesseeTony]

The next round is FACTORIAL sieve, GPU only. 180 seconds for a 75% power limited 4090 (versus 90 seconds on AP27, with FAR more points offered on AP27). 10m10s on a 2080Ti @80% power limit (8m on AP27). Sieve projects IIRC have always offered more points, for the lack of fame (finding a prime). Sadly that does not seem to be the case this time around, but at least we are not talking hours on slower cards.

 

[StefanR5R]

Actually the next one was previously meant to be at a new project (or remake of an old project which was called "Factorial Prime Search". So, not the corresponding sieve project.

Recently, the 2025 Challenge Series table was edited to show "Compositoral" as the 7th challenge project, still at the same time slot of 18-25 September 08:00 UTC (7 days).

In June, PrimeGrid started a new project called Factorial / Compositorial Sieve.

The project is necessary step in preparation for new Factorial Prime Search and Compositorial Prime Search on PrimeGrid.

Now the big question is: Will the September challenge be held at "Factorial/Compositorial Sieve"? I don't think so; why would they list the challenge project as "Compositorial" then? Or will they start a Compositorial Prime Search project in time for the challenge? I haven't seen anything in the message board about it. Perhaps somebody said something at discord (which I don't read).

At the PrimeGrid preferences web page, none of the existing projects is marked as World Bamboo Day Challenge challenge project yet.

 

[Ken g6]

Recently, the 2025 Challenge Series table was edited to show "Compositoral" as the 7th challenge project, still at the same time slot of 18-25 September 08:00 UTC (7 days).

They edited it down to 5 days now.

 

[StefanR5R]

Mottos of the 2025 PrimeGrid Challenge Season:
Flexibility — Who copes best with the diverse project requirements and reacts fastest?
Tactics — Which team manages fast project switching best?

;-)

 

[Skillz]

I think TAAT is the answer to both. Maybe.

 

[Markfw]

So what does the upcoming challenge "Compositorial" use ? CPU ? GPU ? Both ? Maybe by then I will be ready to help .

Looks like GPU only


I have 5 4090's and 6 4080 supers. This should help.

 

[Skillz]

It's not been officially picked yet. The Factorial/Compositorial Sieve project is the prerequisite for the challenge. It will be a CPU project if enough of the Factorial/Compositorial Sieve is done as it's basically generating the CPU work for the challenge. Just get to get more guys to help run that GPU project so we would have enough tasks to at least make it through a challenge.

 

[Markfw]

It's not been officially picked yet. The Factorial/Compositorial Sieve project is the prerequisite for the challenge. It will be a CPU project if enough of the Factorial/Compositorial Sieve is done as it's basically generating the CPU work for the challenge. Just get to get more guys to help run that GPU project so we would have enough tasks to at least make it through a challenge.

confused. Are you saying that the SR5 sieve has to run now ?

 

[mmonnin03]

There's a sieve app before an LLR app to quickly weed out possibilities. The competition was meant for the LLR version but PG isn't done sieving from what I gather. I'm sure someone that cares about primes will correct all that.

What is a prime sieve method, and how did they help Zhang, Maynard and Tao?

At children's school we learned about the Sieve of Eratosthenes for sieving our primes from an interval of natural numbers. I was surprised to hear that "sieve methods" were used to make

math.stackexchange.com

 

[Markfw]

There's a sieve app before an LLR app to quickly weed out possibilities. The competition was meant for the LLR version but PG isn't done sieving from what I gather. I'm sure someone that cares about primes will correct all that.

What is a prime sieve method, and how did they help Zhang, Maynard and Tao?

At children's school we learned about the Sieve of Eratosthenes for sieving our primes from an interval of natural numbers. I was surprised to hear that "sieve methods" were used to make

math.stackexchange.com

OK, still confused. What app in PG do I need to run to help get the Sept race ready to go ?

 

[Skillz]

Factorial/Compositorial sieve on GPUs.

 

[Markfw]

Ok, so run the GPU job I thought was for the competition, but it just gets the CPU client ready ?

 

[StefanR5R]

@Markfw, the status of the September challenge is still [almost] the same as I noted in January:

Some remarkable bits of the — technically tentative — 2025 challenge schedule:

• The projects of two challenges which are scheduled for later this year are not active yet.
  • Factorial Prime Search, scheduled for September, is a new project. Or rather, a reboot of an old project.
    The PRST program which is used in several of the active subprojects does have tests which are aimed at factorial prime numbers. I don't know what the Factorial sieving status is.
  • GFN-22, scheduled for December, is suspended as long as no GFN-21 prime is found. Maybe they'll unsuspend it for the challenge regardless.

One little detail has changed: This new and not yet active project has recently and silently been switched from Factorial Prime Search to Compositorial Prime Search, which is a similar subject matter.

I for one don't know if PrimeGrid needs any sieving help in order to start this project. They haven't posted anything in this regard. Maybe they said something on their Discord chat, or maybe not, impossible for me to say because I am not a participant there.

Since they did not make anything known to the general BOINC using masses, I am certain that they do not need help from the general BOINC using masses in setting up this next challenge. :-) So if you prefer to fold some more proteins in the meantime, this should be a perfectly good plan.

The situation is a bit similar to 2024, when they launched the new Primorial Prime Search project shortly before the August 2024 challenge. Beforehand, Primorial Sieving was performed at the separate Private GFN Server project, if I recall correctly.

 

[StefanR5R]

The Compositorial Prime Search project will apparently go live on September 10:

The "Compositorial (PRST)", short: "prstCPS" application was put onto the server two days ago.

The main page shows it with 0 available tasks right now, and with ~400,000 digits sized primes. This is smaller than PPSE and GFN-16, that is, this one is now the PrimeGrid subproject with the smallest prime finds. (And like currently PPSE and GFN-16, Compositorial Prime Search finds are not T5k eligible.)

If you go to PrimeGrid preferences and enter the editing link, this is what is shown right now:

Compositorial Prime Search (CPS)
Project will be started 2025-09-10
(n!/n#)±1 (n# is the product of all primes less than or equal to n)

Supported platforms:

• Windows: 32 bit, 64 bit
• Linux: 64 bit
• Multi-threading is supported and IS recommended. Click here to set the maximum number of threads.
• Uses fast proof tasks so no double check tasks are needed. Everyone is "first"!

Deadline: 4 days (up to 30 days)
Recent average CPU time: 0:22:00
FFT size: 2x72K (uses up to 2304K cache per task in worst case)

They haven't created a thread for this project at their message board yet.

As an owner of a couple of AMD Zen 2 CPUs, I am happy to learn that this project fits well into the smaller CCXs of those (smaller than Zen 3 and later). That said, it remains to be seen how quickly the "leading edge" of the project will advance when it is live.

 

[Markfw]

The Compositorial Prime Search project will apparently go live on September 10:

The "Compositorial (PRST)", short: "prstCPS" application was put onto the server two days ago.

The main page shows it with 0 available tasks right now, and with ~400,000 digits sized primes. This is smaller than PPSE and GFN-16, that is, this one is now the PrimeGrid subproject with the smallest prime finds. (And like currently PPSE and GFN-16, Compositorial Prime Search finds are not T5k eligible.)

If you go to PrimeGrid preferences and enter the editing link, this is what is shown right now:

They haven't created a thread for this project at their message board yet.

As an owner of a couple of AMD Zen 2 CPUs, I am happy to learn that this project fits well into the smaller CCXs of those (smaller than Zen 3 and later). That said, it remains to be seen how quickly the "leading edge" of the project will advance when it is live.

None of the blue things that say "click here" appear to work. This is a CPU project ? and what does everyone thing that best number of threads per job ?

 

[mmonnin03]

It is CPU only.

 

[StefanR5R]

what does everyone thing that best number of threads per job ?

Short answer: Go with singlethreaded jobs, and run only as many jobs on a host as there are real cores (not SMT threads).

Long answer:

If you wanted to optimize for *speed*, then ≈8 per job, and only as many jobs at once per host so that SMT threads remain unused, and set CPU affinities such that each job is confined to a single last-level cache domain.

But if you want to optimize for *throughput* — and this is what most want for a credits-based challenge like these ones —, then the first question to answer is how many jobs at once per host to run. Or in case of AMD Zen: How many jobs at once per CCX.

The following is for desktop and server AMD Zen (dense server excluded, 3D V-cache variants excluded, variants with disabled cores excluded) :

Zen 1...2 have 4c/8t CCXs with 16 MB L3$/CCX. 16,384 kB/CCX / 2,304 kB/job ≈ 7 jobs/CCX.
Round this down to 4 jobs/CCX and test whether 1-threaded jobs or 2-threaded jobs give higher throughput.

Zen 3...5 have 8c/16t CCXs with 32 MB L3$/CCX. 32,768 kB/CCX / 2,304 kB/job ≈ 14 jobs/CCX.
Round this down to 8 jobs/CCX and test whether 1-threaded jobs or 2-threaded jobs give higher throughput.

The likely outcome of appropriately precise throughput tests is that either 1-threaded jobs work slightly better, or that 2-threaded jobs work a tiny bit better but at the cost considerably more energy per task. A big benefit of 1-threaded jobs is that you don't have to assign CPU affinity, at least not on Linux. 2-threaded jobs need CPU affinity set such that both threads of a job run on the same CCX, otherwise throughput suffers.

For a throughput test, the PRST exe and a suitable number of the (n!/n#)±1 form are required. One way to get the latter is to wait until September 10 and then take the candidate number of a real workunit.

PS, about speed vs. throughput: The PrimeGrid page which I cited gave an estimate of 22 minutes CPU time on average CPUs, which is short. Also, the "fast proof" verification scheme is enabled in this project = "everyone is first" (every prime finder will be credited as first finder). These are more reasons why one should not bother with optimization for speed, but should rather optimize for throughput.

PPS, the PrimeGrid page which I cited said "Multi-threading is supported and IS recommended". I believe this recommendation is mostly for CPUs with less cache per core (≤ 2 MB last level cache per core, notably), e.g. various Intel desktop and mobile CPUs and several AMD mobile CPUs. Such CPUs can only run fewer jobs at once before they run into a serious memory bandwidth bottleneck. With fewer jobs at once, they should engage respectively more cores per task, in order to leave no core unused.

 

[mmonnin03]

WUProp hours posted for the new PG app

The application Compositorial (PRST) from project PrimeGrid is active

 

[StefanR5R]

Meanwhile, the "edit PrimeGrid preferences" page switched to

Recent average CPU time: No Data
FFT size: 0K (uses 0K cache per task)

This should correct itself when enough recent results are in, I suppose.

However, the workunit sizes are much larger now than what I quoted in #264. The PrimeGrid main page now shows prime sizes of ~870,00 digits, and that these primes would enter T5K at rank 4200.