Info PrimeGrid Challenges 2024, sieve-free edition

[waffleironhead]

8 core 7700x is 20 hours
RX7600 is 9.5 hours
Rx6500 23.3 hours
Rx6400 27.6 hours

 

[Skillz]

GPU time comparison:

4090 about 2 hours
2080 Ti about 6h10m
2080 about 7h30m
1080 Ti about 10h45m
1070 about 16 to 17 hours, depending on the factory clock

22 core Broadwell Xeon with 55MB cache is 26 to 42 hours. Ouch.

Also...

P100 about 10 hours 40 minutes
3070Ti about 5 hours 30 minutes
3060 about 9 hours 25 minutes
3090Ti about 5 hours 15 minutes.

My 4090s are doing them in about 1 hour 40 minutes on one host and 1 hour 32 minutes on another host.

Stats taking from the PG elapsed time estimate column.

 

[Orange Kid]

4070ti super - 2.5 hours
3070ti - 5.25 hours
3070 - 6 hours
Using the average at PG

 

[Skillz]

I need to replace the PSU on my 6x 3070 Rig as I don't have any of them in the challenge right now. At least, I hope it's the PSU that's crapping out.

It could be one (or all) of the PCIe extensions on the cards and I really hope it's not one of them. It's so much working trying to figure that problem out.

 

[Pokey]

I'll be crippled for a while, (output wise). A network switch died and My rigs downstairs are cut off from civilization.

 

[Skillz]

I'll be crippled for a while, (output wise). A network switch died and My rigs downstairs are cut off from civilization.

Oh crap that's not good. At least switches are relatively cheap(ish) to replace and easy at that.

I think I have a switch on my network going bad. Need to swap it out to see, but connecting to some of my hosts remotely seems to cause problems with random disconnects. I think it's the switch.

 

[Ken g6]

Day 4 stats:

Rank___Credits____Username
5______98024694___crashtech
7______82299034___[TA]Skillz
8______76228390___markfw
15_____44585927___Pokey
27_____26421523___ChelseaOilman
36_____18180889___cellarnoise2
39_____16838478___biodoc
40_____16811680___mmonnin
41_____15202829___w a h
134____5151106____Ken_g6
159____3768819____waffleironhead
189____2741787____SlangNRox
198____2441493____johnnevermind
227____1886783____Fardringle

Rank__Credits____Team
1_____410588753___TeAm AnandTech
2_____312092372___Antarctic Crunchers
3_____287327998___Planet 3DNow!
4_____218430826___SETI.Germany

 

[Ken g6]

Also:

RTX 4070: 3.75 hours
GTX 1060 3GB: 29 hours

 

[Markfw]

Also:

RTX 4070: 3.75 hours
GTX 1060 3GB: 29 hours

3090TI 3 hours 20-30 minutes. Its a pig for the wattage it sucks

 

[jdiddy]

If anyone is interested:

Apple M4 Max (32 GPU cores): 7.5 hours
Apple M4 (10 GPU cores): 27 hours
RTX 2070 Super: 8.25 hours

 

[crashtech]

That M4 Max time is damned impressive for a low-power chip. Possibly the best ppd/joule ever on this project.

 

[Ken g6]

Day 5 stats:

Rank___Credits____Username
5______125129456___crashtech
7______104240356___[TA]Skillz
8______95559501___markfw
16_____53502962___Pokey
27_____34331960___ChelseaOilman
30_____30461056___cellarnoise2
40_____21419290___mmonnin
41_____21302125___biodoc
54_____16786689___w a h
138____6525663____Ken_g6
163____4624136____waffleironhead
181____3821445____johnnevermind
192____3430443____SlangNRox
248____1886783____Fardringle

Rank__Credits____Team
1_____523027188___TeAm AnandTech
2_____400255229___Planet 3DNow!
3_____380461239___Antarctic Crunchers
4_____283505078___SETI.Germany

Halftime!

 

[Pokey]

OK, I'm back in the game, not that you all need me, but problem(s) solved. There was more than one, of course. Isn't there always? Switch, cable and a surge protector. We are all working and back in touch with the world. 👌

 

[Markfw]

I am now 9th, I think I can hold off tenth

 

[StefanR5R]

I am now 12th, but it doesn't look like I could hold off the one who is 17th 16th 15th 14th... ;-)

 

[Ken g6]

Day 6 stats:

Rank___Credits____Username
5______145592947___crashtech
7______129546160___[TA]Skillz
9______115385440___markfw
14_____68151258___Pokey
26_____41894827___ChelseaOilman
31_____35826926___cellarnoise2
39_____26795612___biodoc
42_____26036700___mmonnin
66_____18013454___w a h
135____8070419____Ken_g6
170____5368786____johnnevermind
181____4795420____waffleironhead
202____3944627____SlangNRox
271____1886783____Fardringle

Rank__Credits____Team
1_____631314681___TeAm AnandTech
2_____463416611___Planet 3DNow!
3_____442119359___Antarctic Crunchers
4_____346806767___SETI.Germany

@TennesseeTony's like reminded me to post, so thanks!

 

[waffleironhead]

My production has fallen off. I blame epic games for giving away free copies of Return to Moria. My whole family got copies and we have been playing it all weekend and every night so far.

 

[Orange Kid]

My production has fallen off as well, but it seems the tasks are suddenly taking a lot longer to complete. Either they are all primes or someone switched up something.

 

[StefanR5R]

(Liking your post for the heads-up, not for the increased task times.)
I noticed reduced PPH at several individuals and teams, but not for myself yet. But I am maintaining a 1 day deep work buffer because I am suspicious of my Internet connection. If the workunits have become more difficult (for specific hardware) without proportional increase of credits, I could be encountering this almost 1 day later than many others...

 

[StefanR5R]

PS, as most here know, the PrimeGrid folks adjust credits for each workunit based on computational difficulty, which they are able to gauge beforehand with good accuracy. But in case of Genefer, CPUs and GPUs use different transforms, and the hardware model for credits prediction can only be aligned with one or the other. I don't know though if a divergence of these is indeed happening at the moment.

 

[Ken g6]

Day 7 stats:

Rank___Credits____Username
5______162539500___crashtech
7______145614241___[TA]Skillz
9______132392070___markfw
13_____79518802___Pokey
26_____47250654___ChelseaOilman
30_____41050414___cellarnoise2
38_____30607109___biodoc
40_____29962698___mmonnin
74_____18356517___w a h
137____8931923____Ken_g6
167____6228217____johnnevermind
177____5654244____waffleironhead
206____4467124____SlangNRox
285____1886783____Fardringle

Rank__Credits____Team
1_____714465617___TeAm AnandTech
2_____494215630___Planet 3DNow!
3_____483265071___Antarctic Crunchers
4_____391559216___SETI.Germany

Slightly early to include news from PrimeGrid:

Hello. I've noticed the last couple tasks for Genefer 21 4.04 (OCLcudaGFN) went from a completion time of about 3 hours and 45 minutes to over 7 hours on the same hardware. Looking under tasks at completed items, the run time has nearly doubled but the credit remains the same. Unsure what to make of it. Any ideas?

It was explained in our Discord server that a change of transform (Number-Theoretic Transform, NTT) happened for GFN-21 on GPU at b = 2,019,124. It has the consequence you are seeing. /JeppeSN

It only affects GPU at this time, not CPU. I think credit is based only on CPU performance.

 

[Pokey]

I have set all my rigs at 80% PL and the resulting point reduction, while small, might insure I will be stuck in 13th place for the duration. So, Stefan, you may be safe.

 

[StefanR5R]

I could be encountering this almost 1 day later than many others...

Alas, I saw just now my hosts reporting the first two results of this workunit type.

For the sake of the team to which I defected to I'll look into whether or not it's worth to switch more GPU board power and to add the better ones among my CPUs. (I only have GPUs engaged so far and they produce more than enough heat already.) But I don't have hopes that either of this is going to make a meaningful difference.

So, Stefan, you may be safe.

Not anymore?

 

[StefanR5R]

I'll look into whether or not it's worth to switch more GPU board power

RTX 4090 @375W runs at ≈2.67 GHz, @450W at 2.775 GHz. But this depends on the specimen, some may clock a little slower.

and to add the better ones among my CPUs.

I tested the following hosts with the workunit n = 21 (naturally), b = 2,020,926 (that's one of those which are past the b limit which @Ken g6 cited from JeppeSN), credit = 171,517.55.

• RTX 4090 @ 2.75 GHz + Kabylake at 3.4 GHz:
  1 task at a time, 2:27 h:m, 1,680 kPPD, 515 W, 3.27 kPPD/W
• EPYC 9554P @ 400W:
  8 tasks at once, 8 threads per task, 22:19 h:m, 1,480 kPPD, 510 W, 2.89 kPPD/W
  4 tasks at once, 16 threads per task, 13:31 h:m, 1,220 kPPD, 495 W, 2.46 kPPD/W
• dual EPYC 7452 @ 180 W I think:
  8 tasks at once, 8 threads per task, 82:15 h:m, 400 kPPD, 445 W, 0.90 kPPD/W
  4 tasks at once, 16 threads per task, 47:43 h:m, 340 kPPD, 410 W, 0.84 kPPD/W
• dual Xeon E5-2696 v4 @ 2.6 GHz:
  2 tasks at once, 22 threads per task, 22:22 h:m, 370 kPPD, 515 W, 0.71 kPPD/W

CPU affinities were set. "h:m" is task duration extrapolated from a few minutes long test. Wattages are total host power.

Edit,
Zen 4's performance as well as perf/W look surprisingly good compared to Ada GPUs, not to mention any of the earlier GPU generations. But this is after GPUs got hit by this nasty b limit.

 

[StefanR5R]

stderr.txt of the GPU results show that genefer22g used 48 MBytes of coefficients for the older workunits with b < 2,097,152, and 72 MBytes of coefficients for the current workunits. I gather from pschoefer that the previous GPU transform used a 2*INT32 datatype and and the current GPU transform is using a 3*INT32 datatype.

PrimeGrid's message board has got tables of transforms and their b limits for Genefer versions 2.3.0, 3.1.2, and 3.2.9, but not for genefer22 yet (Genefer versions 22 and 23), which differ more or less from the earlier implementations. Fortunately for CPUs, genefer22 is able to run GFN-21 with vectorized arithmetic until b = 45,000,000, unlike the earlier genefer versions which had to fall back to slow classic x87 arithmetic at rather low b limits already. On GPUs, the changes are perhaps not as significant, except that the topmost b limit was bumped to 2,000,000,000, and that no FP64 transform has been reimplemented yet for respectively capable GPUs.