PrimeGrid Challenges 2023

[Markfw]

Thanks Ken !

 

[Ken g6]

Day 8.5 stats:

Rank___Credits____Username
2______62216610___markfw
5______27398834___xii5ku
6______24130896___crashtech
9______20428017___w a h
16_____10340428___cellarnoise2
36_____4228117____Fardringle
41_____3608699____biodoc
46_____3230562____Orange Kid
50_____3028052____waffleironhead
58_____2423993____mmonnin
107____1195631____Ken_g6
152____639975___10esseeTony
157____562861_____johnnevermind
206____225555_____[TA]Skillz
243____164328_____kiska
340____5033_______Pokey
348____2584_______Letin Noxe

Rank__Credits____Team
1_____163830184___TeAm AnandTech
2_____87956603___Czech National Team
3_____71279893___Ural Federal University
4_____69737533___Antarctic Crunchers

 

[TennesseeTony]

Thanks Ken!

 

[Markfw]

Thanks Ken!

 

[StefanR5R]

............................................. day 6.5 .............. day 7.5 .............. day 8.5
TeAm AnandTech ......... 117,790,115 ....... 138,939,886 ....... 163,830,184
the next two teams ....... 115,371,421 ....... 140,398,315 ....... 159,236,496
delta ................................ +2,418,694 ........... -1,458,429 ......... +4,593,688

 

[Ken g6]

Day 9.5 stats:

Rank___Credits____Username
2______69215380___markfw
5______29580133___xii5ku
6______27278056___crashtech
9______22420367___w a h
15_____14594633___cellarnoise2
38_____4659862____Fardringle
44_____3938771____biodoc
47_____3878852____Orange Kid
51_____3355658____waffleironhead
57_____2869105____mmonnin
109____1305917____Ken_g6
153____738721___10esseeTony
164____562861_____johnnevermind
218____225555_____[TA]Skillz
250____164328_____kiska
256____120002_____Letin Noxe
351____5033_______Pokey

Rank__Credits____Team
0_____187006780___(the next two teams)
1_____184913241___TeAm AnandTech
2_____104014265___Czech National Team
3_____82992515___Ural Federal University
4_____79595249___Antarctic Crunchers

Probably time for some to set No New Tasks. Or maybe start one of the several projects that will shut down soon.

 

[Markfw]

Day 9.5 stats:

Rank___Credits____Username
2______69215380___markfw
5______29580133___xii5ku
6______27278056___crashtech
9______22420367___w a h
15_____14594633___cellarnoise2
38_____4659862____Fardringle
44_____3938771____biodoc
47_____3878852____Orange Kid
51_____3355658____waffleironhead
57_____2869105____mmonnin
109____1305917____Ken_g6
153____738721___10esseeTony
164____562861_____johnnevermind
218____225555_____[TA]Skillz
250____164328_____kiska
256____120002_____Letin Noxe
351____5033_______Pokey

Rank__Credits____Team
0_____187006780___(the next two teams)
1_____184913241___TeAm AnandTech
2_____104014265___Czech National Team
3_____82992515___Ural Federal University
4_____79595249___Antarctic Crunchers

Probably time for some to set No New Tasks. Or maybe start one of the several projects that will shut down soon.

I started setting no new tasks hours ago, and left a 2 hour buffer, as once you kill a task, and add no new work, the others may finish faster. I have about 52 units still that may finish, but I think unless somebody dumps some huge bunkers, we have 1st overall, and I have 2nd as an individual.

As far as setting other work, right now my electric bill is way out of hand (estimated at $1000) so as each computer runs out of tasks, they will be shut down. I will be in minimal mode for a month or so.

 

[mmonnin03]

I needed to complete 1 more task to get my 25k WUProp hours for this app. I asked for more work, BOINC gave me a bunch of work (well beyond by queue request) of like 20 tasks and they were all the full task with at least _3 to _7. Lots of people aborting work at the end of competition trying to get the shorter proof tasks.

 

[StefanR5R]

I asked for more work, BOINC gave me a bunch of work (well beyond by queue request) of like 20 tasks [...]

In the past, if I wanted to run [m] simultaneous tasks with [n] threads each, I used to set my app_config.xml's like this:

<cmdline>-t [n]</cmdline> = as usual
<avg_ncpus>1</avg_ncpus> = fake​

Then I configured the boinc clients to use only [m] logical CPUs. This way, requests for new work did not end up with responses which contained far too many new tasks.

Still, configuring a work buffer depth of [x.y] days did not work quite as I wanted this way. So I tried the old

<cmdline>-t [n]</cmdline>
<avg_ncpus>[n]</avg_ncpus>​

again when the competition started (and allowed the client to use [m×n] logical CPUs), to see if things have improved in the meantime. They haven't. Work requests were ending up with far too many tasks in the response again. So I added profiles with limits of tasks in progress in the web preferences (incorrectly called "Max # of simultaneous PrimeGrid tasks" there), with limits which roughly corresponded with my desired work buffer depths.

Either way sort of works, but I don't like it that I only have imprecise control over the work buffer depth.

Probably time for some to set No New Tasks. Or maybe start one of the several projects that will shut down soon.

With the latter, in case that you still have some SoB to finish, you risk that the client intelligently suspends the boring old SoB work and runs the shiny new non-competition work instead.

 

[Ken g6]

More or less final stats:

Rank___Credits____Username
2______74226152___markfw
5______32608922___xii5ku
6______28936664___crashtech
9______23515076___w a h
15_____15218763___cellarnoise2
38_____4867463____Fardringle
46_____4048391____biodoc
54_____3465946____waffleironhead
60_____2869105____mmonnin
108____1416210____Ken_g6
142____849683___10esseeTony
161____672729_____johnnevermind
227____278716_____Letin Noxe
229____225555_____[TA]Skillz
255____164328_____kiska
355____5033_______Pokey

Rank__Credits____Team
1_____197579176___TeAm AnandTech
1.5___195660966___(the next two teams)
2_____109639116___Czech National Team
3_____86021850___Ural Federal University
4_____85406479___Antarctic Crunchers

Yes! We did better than the next two teams combined!

 

[StefanR5R]

@Ken g6, thanks for the quips! :-)

And for the whole lot of stats, of course.

Great job, TeAm!

 

[Skillz]

You guys are lucky I managed to return one main task so we could get that win.

 

[Markfw]

You know I still can't believe that no one even replied to my post in CPUs about how my 4 Genoas destroyed all Intel, including the 3595x, and almost 2000 Intel cores (older, by 7 years, yes). Anyway, I know, and you all know the power of Genoa.

 

[Skillz]

You know I still can't believe that no one even replied to my post in CPUs about how my 4 Genoas destroyed all Intel, including the 3595x, and almost 2000 Intel cores (older, by 7 years, yes). Anyway, I know, and you all know the power of Genoa.

Why?

You had more than just 4 Genoa's running in the competition. Those 4 Genoa's had a combined core count of 288 cores.
You had, I think 6x 7950x CPUs to finish. That's another 96 cores.
Two 7V12s. That's 128 cores.
Maybe some other stuff I can't remember or you didn't mention.
Total: 512 cores

Pavel has 70 Xeon servers, with dual CPUs. Most of his CPUs are V2 Xeons that came out in 2013. That's nearly 10 years old. Not 7.
His Xeons range from 10 core 2660 V2 to 12 core 2695 V2 CPUs. 70 servers with dual CPUs is 140 CPUs.
If they were all 12 core parts that's a total of 1,680 Cores.

Are Genoa EPYCs extremely good? Yes.
Can 4 Geona CPUs beat 140 Xeons? No.

 

[Markfw]

Why?

You had more than just 4 Genoa's running in the competition. Those 4 Genoa's had a combined core count of 288 cores.
You had, I think 6x 7950x CPUs to finish. That's another 96 cores.
Two 7V12s. That's 128 cores.
Maybe some other stuff I can't remember or you didn't mention.
Total: 512 cores

Pavel has 70 Xeon servers, with dual CPUs. Most of his CPUs are V2 Xeons that came out in 2013. That's nearly 10 years old. Not 7.
His Xeons range from 10 core 2660 V2 to 12 core 2695 V2 CPUs. 70 servers with dual CPUs is 140 CPUs.
If they were all 12 core parts that's a total of 1,680 Cores.

Are Genoa EPYCs extremely good? Yes.
Can 4 Geona CPUs beat 140 Xeons? No.

Yes, but the 6th 7950x was not live until about 1/2 way through the competition. Same with the dual 7V12'as and the 7763's. I would have to check the dates, but they were several days after the start.

I guess my point is the same as you are saying, they are extremely good. I wish I knew what I was up against on day 1, as I would have used the 7V12's and the 7763's earlier. I will may have lost, but I put up a good fight.

 

[StefanR5R]

Well, it's both about what hardware you bring and how you are using it.

LLR runs best with FMA3 support or better. And SoB-LLR operates on 30 MBytes of FFT data at the current search range. If the CPU can't cache all of these — or can, but the user didn't take the necessary configuration steps —, the CPU will merely trod along with more or less of a handicap.

On the Intel side:

Haswell-EP (and its derivative, Broadwell-EP) with >12 cores manage, but they are old and therefore energy-intensive to operate. Before Haswell(-EP), there is no FMA3 support.

Skylake-SP and its derivatives presumably need about >22 cores in order to have sufficient cache (because there is a lot less L3$/core than in Haswell-EP). But last time I checked, LLR2 was unable to utilize respectively many software threads on Skylake-SP and its ilk. It hit a ceiling at about 8 threads, hence, the numerous AVX-512 units idled a lot. I haven't checked for quite a while now whether or not these scaling issues have been addressed in the meantime.

All other Intel CPUs just don't have enough cache. SoB-LLR is bottlenecked by memory accesses on them. (Plus potential additional bottlenecks, depending on CPU generation and the user's setup.)

On the AMD side:

Zen 2 with its TSMC 7nm process and native 256-bit FMA3 support was nice and all, but has got far too small CCXs for SoB-LLR (4 cores, 16 MB L3$). Zen 3 (same manufacturing process and same vector instruction set, but double CCX size) is what you would prefer over Zen 2 for SoB-LLR.

Zen 4 brings the same CCX size as Zen 3, such that SoB-LLR fits into the CCXs of both of them. But Zen 4 is also manufactured on a newer process (thus, is even more energy efficient), and has got some core architecture tweaks (such as AVX-512 support) which successfully improve per core and per clock throughput versus Zen 3.

Comparing big AMD CPUs with big Intel CPUs, Zen (2,3,4)'s division into CCXs is one of the reasons why they are so fast and efficient. But for the same reason, the user needs to go through some more hoops to set up an application like SoB-LLR on Zen (2,3,4), because Windows and even Linux do not schedule the tasks appropriately. With Linux and AMD server CPUs, you get almost optimal results with respective BIOS hints to the OS, but intervention at the application software level (i.e. logical CPU affinity) still brings best results. So the user has a price to pay for Zen (2,3,4)'s added complexity compared to Intel gear.

 

[Orange Kid]

Thanks for all the days of stats.

Oh, sure, I pass biodoc and I'm no longer in the stats. How much did he pay you. 🤔 😉

 

[Ken g6]

Thanks for all the days of stats.

Oh, sure, I pass biodoc and I'm no longer in the stats. How much did he pay you. 🤔 😉

Huh. Must be gremlins someplace. I copy and paste from my script that gets the stats from PG.

Rank___Credits____Username
2______74226152___markfw
5______32608922___xii5ku
6______28936664___crashtech
9______23515076___w a h
15_____15218763___cellarnoise2
38_____4867463____Fardringle
43_____4210432____Orange Kid
46_____4048391____biodoc
54_____3465946____waffleironhead
60_____2869105____mmonnin
108____1416210____Ken_g6
142____849683___10esseeTony
161____672729_____johnnevermind
227____278716_____Letin Noxe
229____225555_____[TA]Skillz
255____164328_____kiska
355____5033_______Pokey

Rank__Credits____Team
1_____197579176___TeAm AnandTech
2_____109639116___Czech National Team
3_____86021850___Ural Federal University
4_____85406479___Antarctic Crunchers

Well, I did say they were "more or less final". 😉

Come to think of it I should have known there was one more contributor. This was a "Seventeen or bust" challenge. 😛

 

[Ken g6]

In case anyone wasn't aware, PrimeGrid will be down for some hours tomorrow while they switch servers. Mike has promised the switch will go smoother than WCG's. 😉

 

[crashtech]

That bar is so low it's underground.

 

[Markfw]

primegrid homepage is now updated for the Xmas challenge, as well as the options description. a little over 7 days to go.

 

[StefanR5R]

Here are the data allocation sizes — i.e. processor cache demand per task — at the current leading edge [...]
31 October - 5 November
ESP-LLR: 18.0 MB

I checked again for the eight remaining ks, for min n and max n per k. Allocation sizes are 18.0…20.0 MB now (on Haswell).

Edit,
if somebody needs a test specimen: 238411*2^24201228+1 takes 20.0 MB in "all-complex FMA3 FFT", or 19.7 MB in "all-complex AVX-512 FFT" respectively, and gives 31,512.33 credit.

 

[StefanR5R]

So the next challenge is all about

• disproving that there are any undiscovered Sierpiński numbers left between the currently known two smallest ones,
• whether Exhaustive Search on 'Puters or ExtraSensory Perception will be the more effective tool to solve above riddle,
• the mistaking of base ten for base eight,
• not letting oneself be fooled by ever changing local clock offsets.

On the latter issue: We are supposed to fetch work after…
…Wednesday, November 1, 00:00 AEDT,
…Tuesday, October 31, 14:00 CET/ 13:00 GMT,
…Tuesday, October 31, 09:00 EDT/ 08:00 CDT/ 07:00 MDT/ 06:00 PDT,
and to report results before…
…Monday, November 6, 00:00 AEDT,
…Sunday, November 5, 14:00 CET/ 13:00 GMT,
…Sunday, November 5, 08:00 EST/ 07:00 CST/ 06:00 MST/ 05:00 PST.

 

[mmonnin03]

Haha DST used to change this weekend in the US but it was pushed back a week so now it's next weekend after Halloween.

 

[Markfw]

Haha DST used to change this weekend in the US but it was pushed back a week so now it's next weekend after Halloween.

Yes, 46 hours and change to go !