Info PrimeGrid Challenges 2025

[Skillz]

OK, these have ETAs that could change, as I just updated lasso, but here is the times after changing
View attachment 130085

These tasks grow exponentially fast. 6 days from now those numbers wont mean anything.

 

[StefanR5R]

These tasks grow exponentially fast. 6 days from now those numbers wont mean anything.

Zen 1...2: 16,384 kB/CCX / 6,144 kB/job = 2⅔ concurrent jobs/CCX ≈ 2 concurrent jobs/CCX
Zen 3...5: 32,768 kB/CCX / 6,144 kB/job = 5⅓ concurrent jobs/CCX ≈ 4 concurrent jobs/CCX

1 day later: 7,680 kB/job; and 6 days to go until start of the challenge. The above job counts per CCX will likely be obsolete by then. I am a bit worried how useful Zen 2 will remain during the challenge.

 

[Skillz]

1 day later: 7,680 kB/job; and 6 days to go until start of the challenge. The above job counts per CCX will likely be obsolete by then. I am a bit worried how useful Zen 2 will remain during the challenge.

I think when using AVX512 for these tasks they use more L3. Based on the rate these things are going I don't know, but it's not looking good.

 

[Markfw]

BTW, these use AVX-512 a lot. The Zen 4 8854 is taking ~1.5 hours, and the 9950x is taking ~35 minutes. Thats a lot more than just 3.5-4.6 ghz.



I will try my Turin later and see what times it gets.

 

[Ken g6]

And we're off! "(uses up to 12288K cache per task in worst case)" currently.

 

[Markfw]

All tasks that I have running are way slower than the benchmarks I ran before this started. Something sure changed.

 

[StefanR5R]

The "edit primegrid preferences" page still says

FFT sizes: 2x72K to 2x384K (uses up to 12288K cache per task in worst case)

However, I am already seeing up to

Using Montgomery reduction AVX-512 FFT length 2x432K​

and

Using Montgomery reduction FMA3 FFT length 2x480K​

in boinc/slots/*/stderr.txt.

If 2x384K coefficients means 12.0 MiBytes size of the coefficients, then that's 2 arrays of 384*1024 coefficients with 16 Bytes?? (128 bits??) per coefficient. In contrast, the FFTs which the other LLR/PRST subprojects at PrimeGrid use, are single arrays of 64 bits wide coefficients… Not sure if this Montgomery reduction is really 128 bits math, or if it is 64 bits math but can't store each coefficient in place but is rather swapping back and forth between two 64 bits coefficients (times 2 arrays of those).

Anyway; if this also holds true for the FMA3 implementation, the that one is already 1t 15.0 MBytes cache footprint, uncomfortably close to what Zen 2 for example can carry in cache.

 

[Skillz]

@StefanR5R they think the FFT/L3 requirements for this sub project is different than the others.

The formula is 2*N*8 instead of 2*N*16.

Therefore a FFT size of 2x384K = 6MB

Which is perfectly fine for Rome.

 

[Ken g6]

Day 0.5 stats:

Rank___Credits____Username
1______4434844____markfw
3______3374805____Icecold
4______2764602____ChelseaOilman
5______2453328____crashtech
8______1315428____[TA]Skillz
22_____631283_____Orange Kid
31_____369641_____cellarnoise2
64_____140473_____w a h
70_____128749_____biodoc
76_____115711_____IEC
78_____111368___10esseeTony
93_____85358______Ken_g6
109____73322______waffleironhead

Rank__Credits____Team
1_____15998919___TeAm AnandTech
2_____5796618____Antarctic Crunchers
3_____5638217____Czech National Team
4_____3363494____SETI.Germany