PrimeGrid Challenges 2022

[waffleironhead]

so apparently my config is set got 8c or 8 threads per task. Is this still good ? for the CUL and WOO tasks ?

Yeah, Thats what I run for my 3700x and 5700g. Should be just fine for your rigs.

 

[Markfw]

So it appears that a 7950x does as much in this as a 7V12 Rome or a 7763 Milan. And thats only using 16 thread instead of 32, where the Milan and the 7V12 have SMP off !

Like 4 hours vs 44 hours, but 8 times as many tasks. Could be 10 to 1.

 

[cellarnoise]

Is the 7950x at stock or what are the restrictive settings if any from stock? RAM at 6000?

I think I need one of these by the end of the year or wait to see if they come out with a 7950x 3D cache version?

 

[Markfw]

Is the 7950x at stock or what are the restrictive settings if any from stock? RAM at 6000?

I think I need one of these by the end of the year or wait to see if they come out with a 7950x 3D cache version?

Ram EXPO 6000 CL30, so sort of stock. CPU settings, PBO/CO at -25 gives me 142 watt or a bout 10-20% slower than stock.

 

[StefanR5R]

So it appears that a 7950x does as much in this as a 7V12 Rome or a 7763 Milan.

CPU settings, PBO/CO at -25 gives me 142 watt

Makes sense. Power efficiency in vector math seems to be up a lot in Zen 4, compared to Zen 2…3.

Any special config for this upcoming event ? (Nov 1st). Other than selecting CUL and WOO.

I'll check it out later today, I think.

 

[StefanR5R]

Start time: Request work after November 1, 05:00 UTC = 01:00 EDT = October 31, 22:00 PDT.

End time: Report results before November 11, 05:00 UTC = 00:00 EST = November 10, 21:00 PST.

___________________________________________

From "Edit PrimeGrid preferences" and "Subproject status":

llrCUL (n*2^n+1)

recent average CPU time = 98 hours
20% long job credit bonus
leading edge n = 22,981,098 (6,918,008 digits long)​

llrWOO (n*2^n-1)

recent average CPU time = 170 hours
20% long job credit bonus
leading edge n = 22,024,837 (6,630,144 digits long)​

This ratio of recent average CPU time looks strange, given that llrCUL is already at somewhat larger test candidates and uses slightly larger FFT lengths by now.

I tested the leading edge of both on a Haswell CPU.

llrCUL: "zero-padded FMA3 FFT length 2400K" -> 18.75 MB FFT data size
llrWOO: "zero-padded FMA3 FFT length 2304K" -> 18.0 MB FFT data size​

I looked up some random workunits at the PrimeGrid web site and saw similar FFT lengths on computers which use an AVX-512 FFT implementation instead of the FMA3 implementation.

Zen 2 CPUs have 16 MB last-level cache per CCX (and up to 4 cores per CCX). I'll have to test mine if they run better with 1:1 binding of tasks to CCXs, or 1:2 binding.

Zen 3 and Zen 4 CPUs with their 32 MB last-level cache per CCX (and up to 8 cores per CCX) might work best with 1:1 binding of tasks to CCXs, but that's just a guess.

 

[Markfw]

I just started my first 5950x task, eta 11 hours. But my 7950x's are just killing it at 7 hours ! Thw 12700F is doing about 7 hours also ! But only one task instead of 2 (its e-cores are disabled)

All the Rome and Milan EPYC hosts are over 2 days for 8-16 tasks(SMT OFF).

 

[Markfw]

With help from other DC'ers, I have gotten the EPYCs to about 24 hours for 8 tasks on a 64 core box. I assume 16 tasks on a 2 x 64 core box. I have everything prepped with my last unit finishing right before the start. (4:10 to go and 4:52 until the start)

 

[Ken g6]

Ready or not, we're off!

And if you weren't ready, you have ten days to get it sorted out before the end.

 

[Markfw]

I started a few tasks.

 

[emoga]

It would be cool if a Cullen prime was found in this 10 day long challenge.

The last Cullen was found in 2009 (link)

While the last Woodall was in 2018 (link)

Seems like a good time to find one

 

[crashtech]

It would be cool if a Cullen prime was found in this 10 day long challenge.

The last Cullen was found in 2009 (link)

While the last Woodall was in 2018 (link)

Seems like a good time to find one

Makes me think finding a Cullen prime is less likely than winning the lottery, but I am running both regardless.

 

[Orange Kid]

@crashtech you find the prime, I'll win the lottery. 😄😁🙃

 

[crashtech]

@crashtech you find the prime, I'll win the lottery. 😄😁🙃

Likewise, I am sure. Guess I better buy a ticket!

 

[Ken g6]

Day 1 stats:

Rank___Credits____Username
1______1814055____markfw
2______1760662____xii5ku
4______1591507____crashtech
11_____938745_____w a h
18_____607031_____biodoc
28_____320523_____Orange Kid
36_____253072_____cellarnoise2
37_____222716_____parsnip soup in a clay bowl
61_____110249_____Ken_g6
86_____58902______waffleironhead
96_____58578______Lane42
158____1063_______mmonnin
178____213________[TA]Skillz

Rank__Credits____Team
1_____7737322____TeAm AnandTech
2_____4634239____Czech National Team
3_____3391952____SETI.Germany
4_____2399059____[H]ard|OCP

Hey, we have a different #1 this time! Has @Markfw led in PG before?

 

[Markfw]

Day 1 stats:

Rank___Credits____Username
1______1814055____markfw
2______1760662____xii5ku
4______1591507____crashtech
11_____938745_____w a h
18_____607031_____biodoc
28_____320523_____Orange Kid
36_____253072_____cellarnoise2
37_____222716_____parsnip soup in a clay bowl
61_____110249_____Ken_g6
86_____58902______waffleironhead
96_____58578______Lane42
158____1063_______mmonnin
178____213________[TA]Skillz

Rank__Credits____Team
1_____7737322____TeAm AnandTech
2_____4634239____Czech National Team
3_____3391952____SETI.Germany
4_____2399059____[H]ard|OCP

Hey, we have a different #1 this time! Has @Markfw led in PG before?

I have never had so many fast cores before. Over 1200 and 32 of them are the fastest you can buy,., insanely fast ! ( the 7950x's)

 

[StefanR5R]

Stats in this challenge will jump around a bit, given that one or another team is also bringing single-threaded cloud VMs again.

 

[StefanR5R]

llrCUL (n*2^n+1)

recent average CPU time = 98 hours​

llrWOO (n*2^n-1)

recent average CPU time = 170 hours​

This ratio of recent average CPU time looks strange, given that llrCUL is already at somewhat larger test candidates and uses slightly larger FFT lengths by now.

Now that a lot of people work these two projects, the web site shows much more sensible figures:

llrCUL = 95 hours recent average CPU time
llrWOO = 91 hours recent average CPU time​

 

[crashtech]

So we'll be waiting a few more days to see what happens with the single-core renters. Should be interesting!

 

[Markfw]

34 hours for an 8C task is my slowest box. a 7742 with only 7 channel memory.

 

[cellarnoise]

Zen1 is horrible in this! Running 4c tasks and likely around 34+ hours also even with tasks pinned to a core cluster...

 

[crashtech]

Zen1 runs AVX2 code at half speed, so it's probably not worth running. Also pre-Haswell Intels like Ivy and Sandy are not worth it anymore either.

 

[cellarnoise]

Your speaking the truth!

I spent some time sweet talking this Zen1 puter, even put on some soothing tunes for a good bit of time. No increase in output!

That kind of made me upset, so then after a yelling at the thing for awhile, guess what? Still no increase in its output...

Computers can be so irrational sometimes.

This evening I'm thinking about threaten it with being mothballed and maybe the trash bin. Let's see if that encourages it some?

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
1______4976054____markfw
3______3605071____crashtech
4______3554199____xii5ku
13_____1910688____w a h
18_____1328901____biodoc
21_____1235358____parsnip soup in a clay bowl
29_____820446_____[TA]Skillz
31_____759190_____cellarnoise2
40_____641879_____Orange Kid
43_____478203_____mmonnin
87_____222014_____Ken_g6
106____174130_____waffleironhead
119____146760_____Lane42
148____88050______Mardis
191____54479______mnelsonx

Rank__Credits____Team
1_____19995431___TeAm AnandTech
2_____12781899___Czech National Team
3_____8838095____SETI.Germany
4_____7026057____Antarctic Crunchers

Markfw doesn't look like a fluke. At least not yet. I wonder how long a single thread takes on one of these tasks?

 

[Markfw]

Day 2 stats:

Rank___Credits____Username
1______4976054____markfw
3______3605071____crashtech
4______3554199____xii5ku
13_____1910688____w a h
18_____1328901____biodoc
21_____1235358____parsnip soup in a clay bowl
29_____820446_____[TA]Skillz
31_____759190_____cellarnoise2
40_____641879_____Orange Kid
43_____478203_____mmonnin
87_____222014_____Ken_g6
106____174130_____waffleironhead
119____146760_____Lane42
148____88050______Mardis
191____54479______mnelsonx

Rank__Credits____Team
1_____19995431___TeAm AnandTech
2_____12781899___Czech National Team
3_____8838095____SETI.Germany
4_____7026057____Antarctic Crunchers

Markfw doesn't look like a fluke. At least not yet. I wonder how long a single thread takes on one of these tasks?

8 threads on a 7950x (the fastest out there) takes over 6 hours. So the best would be 48 hours ? but more like 96 ? Well in those 4 days I will have racked up a LOT of jobs. I am doing over 120 a day. So they would need 480 cores just to keep up with me. Well, I guess we will see in 2 more days. Last I checked tng, I was still pulling away from him, 438k in a few hours compared to 294k. I have 832 threads running, mostly optimized for cache and affinity. At only 50% load (but with SMT on). An as we speak, I have about 20 more units that will be dumping in the next hour.

Not saying I will take number one overall, but I think I have a shot at it.

Edit: in case I counted wrong, add up the number of 8c tasks, times the number of tasks for that line. So the last line is 6 CUL tasks*8 cores each or 48 core for that one line. The line one up is 10 WOO tasks*8 = 80, so those 2 lines are 128 for a 256 thread box. It should be close to 832. As far as times, I think you will see the 7950x's are about 7 hours, the 5950x's are about 9-11 hours, and the EPYC's are 23-34 hours. The 12700F has the e-cores disabled, so its just 8 P-cores and about 7 hours like the 7950x's, but its stock (200+ watts) and the 7950x's are 142 watt. I ran out of power in the house, so I have one 5950x off. Unless I don't want to use my TV.