PrimeGrid Challenges 2021

[Markfw]

OK, so I got to thinking, how would a 1950x with 4 channels of memory work, with the same 16 cores as a 5950x... Its says 19 minutes for one unit, and 19 hours for the other... I think I will wait until the morning and see if things straighten out.

So, I could not stand it. Now looks like 19 to 20 hours will be it. WOW, Zen 3 has it all over Zen 1. By a LOT. more than twice the speed !!!

 

[Ken g6]

Its says 19 minutes for one unit, and 19 hours for the other

Some WUs are double-checks. They run much faster than the initial WUs.

 

[lane42]

225​	boss	SETI.Germany	53 993.84​	136​
226​	MikeCAT	Team niconico	53 272.50​	3​
227​	necavi	BOINC@Poland	52 822.44​	2​

boss must get a lot of em

 

[StefanR5R]

Well, that's because user boss just got started with this challenge last night. He only reported results from "proof tasks" so far; his first "main task" results will come in at some later time.

boss, SETI.Germany: 136 results, 397 credits per result on average
Ryan Propper, Antarctic Crunchers: 480 results, 388 credits/result on average
Icecold, TeAm AnandTech: 496 results, 5,702 credits/result on average

 

[StefanR5R]

OKI, a 2990wx.... ETA 14.5 hours ! on 2 16 cu units !

OK, so I got to thinking, how would a 1950x with 4 channels of memory work, [...]
Now looks like 19 to 20 hours will be it.

Interesting,
....................................... 1950X : 2990WX
cores and caches ................ 1 : 2
TDP ....................................... 1 : 1.4
memory bandwidth ............. 1 : 1.1 *
your reported throughput ... 1 : 1.3

So the GCW-LLR performance on these two seems primarily controlled by the CPU power draw, secondarily by RAM bandwidth.

*) assuming DDR4-2667 : DDR4-2933 per AMD's specs

 

[Markfw]

Interesting,
....................................... 1950X : 2990WX
cores and caches ................ 1 : 2
TDP ....................................... 1 : 1.4
memory bandwidth ............. 1 : 1.1 *
your reported throughput ... 1 : 1.3

So the GCW-LLR performance on these two seems primarily controlled by the CPU power draw, secondarily by RAM bandwidth.

*) assuming DDR4-2667 : DDR4-2933 per AMD's specs

The threadrippers are running 3200 cl14 . The 5950x is running 3800 cl16. Even with 32 cores/64 threads its only running 1/2 of them and still takes 13.5 to 14 hours vs the 9 for the 5950x. While the 1950x is holding 20.25. All these are averages in the last 12-24 hours. All 3 are running off the same config, just 16c config units, and 100% allocated to BOINC.

The 5950x is just a BEAST !

 

[crashtech]

5950x is pretty incredible, It's keeping up with my 2x 2690v3 Xeons. 32 threads versus 56 threads, would not surprise me if the 5950X pulls ahead in the end.

 

[Markfw]

5950x is pretty incredible, It's keeping up with my 2x 2690v3 Xeons. 32 threads versus 56 threads, would not surprise me if the 5950X pulls ahead in the end.

Mine is beating a 2990wx, 64 threads. And I checked, its 100% (or close) cpu usage ! and thats at 3 ghz ! I think those 2690v3's do 2.5 ghz ? I have a 2683v3 doing that. I have a 2683v3 doing nothing, can't sell it.

Edit: I just had a thought. For about $150 net (after selling the old motherboard) I can get a dual 2011-3 motherboard and put the one I have and the spare in. Somebody want to tell me I need 28 more threads and talk me into it ?

 

[Ken g6]

Day 4 stats:

Rank___Credits____Username
4______5142228____crashtech
5______4842178____parsnip soup in a clay
7______4121548____xii5ku
14_____3247919____Icecold
19_____2226422____biodoc
22_____1895495____emoga
48_____860676_____Orange Kid
65_____670441_____HutchinsonJC
72_____594543_____Lane42
84_____464079_____Skillz
88_____432403_____Skivelitis2
100____351194_____markfw
134____261525_____Ken_g6
159____212155_____VirtualLarry
214____99118______SlangNRox
243____53050______waffleironhead

Rank__Credits____Team
1_____25474981___TeAm AnandTech
2_____22955166___Czech National Team
3_____19779422___Antarctic Crunchers
4_____19361954___SETI.Germany

See, @Markfw? That wasn't so hard.

 

[VirtualLarry]

Would a 4C/4T 3.5Ghz Skylake (i5-6400/6500) or an R5 1600 do any good for this challenge? I have hesitated throwing them in, I've got 3x Zen2 CPUs (2x 3600, 1x 3800X) running it though.

 

[crashtech]

Zen1 is going to take twice as long as Zen2 on this project.

 

[VirtualLarry]

Zen1 is going to take twice as long as Zen2 on this project.

Plus, last time that PC was in a PG race, it errored like 30-50% of the WUs. So there's something a bit amiss about it.

 

[Ken g6]

Would a 4C/4T 3.5Ghz Skylake (i5-6400/6500) or an R5 1600 do any good for this challenge? I have hesitated throwing them in, I've got 3x Zen2 CPUs (2x 3600, 1x 3800X) running it though.

I'm running Skylake. Each WU should be less than 24 hours, if you use all cores at once.

 

[StefanR5R]

A 4-core Haswell of mine took 31 hours for a task which I ran one week before the race. I'm not sure at which clock it ran, about 3.4…3.6 GHz, on dual-channel DDR3-1600. (RAM bandwidth matters for these small CPUs.) There was desktop stuff going on too.

Skylake should of course do better at the same core count, but I guess not really much more than in proportion to the increased RAM bandwidth. (Skylake-S, that is.)

Plus, last time that PC was in a PG race, it errored like 30-50% of the WUs. So there's something a bit amiss about it.

…and conservative RAM timing matters as well. ;-)
(As you are certainly already aware of, there can be a few causes for errors at PrimeGrid, but RAM which normally seems stable but isn't actually stable is a quite common cause, from what I read.)

 

[VirtualLarry]

Ruh-roh, main rig with R5 3600 is rebooting on me. I have CPB and PBO DISABLED, so I only cruse along at 3.60Ghz, and I'm water-cooled, with a max temp under 70C. I don't know why I'm rebooting so often. I'm mining on twin GTX 1660 Super cards (undervolted) too, using NH, but I have been for several months, and mostly no ill effects, until I started doing PrimeGrid too. Although I did have one or two freeze-ups, not reboots, while mining. (I was mining on my CPU as well.)

I'll keep an eye on it, maybe the PSU is starting to go, it's a Gold-rated Antec 750W though, which should be built like a tank.

 

[crashtech]

Probably the RAM, as Stefan mentioned. I finally had to swap a kit out that was causing problems on my 3700X even though it passed Memtest over and over. Try setting it a little slower and see if it clears up.

 

[VirtualLarry]

Yeah, the RAM is suspect too. I have 4x8GB Trident RGB 3600 RAM. Also have FCLK set to 1800 to match RAM (a slight memory-controller OC).

 

[Markfw]

Yeah, the RAM is suspect too. I have 4x8GB Trident RGB 3600 RAM. Also have FCLK set to 1800 to match RAM (a slight memory-controller OC).

Put the ram at 3200 and the fclk to 1600, same voltage.

 

[Markfw]

Update: So I enabled the video card for F@H on the 1950x and set cpu to 98% for BOINC. It then did one 16cu unit, and its doing it in 14 hours, like the 2990wx is, but only one at a time.

Edit: So my take on this, is that Zen 1 and most likely Zen 2 can only use one real core per compute, so a 3950x would use 16 CU's and one unit, a 1800x and a 3800x would use 8 CU units, one at a time. But the Zen 3, 5950x can use 2 16 CU units at a time, due to optimizations, and the 5800x could probably do one 16cu unit.

Most likely all older Intel CPUs (anything below Alder lake) need to use the one real CPU rule as well.

 

[Ken g6]

Day 5 stats:

Rank___Credits____Username
3______7647602____parsnip soup in a clay
5______6537448____crashtech
9______5239971____xii5ku
11_____4766848____Icecold
19_____2849631____biodoc
22_____2256313____emoga
50_____1117776____Orange Kid
60_____939316_____HutchinsonJC
65_____890795_____Skillz
74_____762880_____Lane42
80_____698603_____markfw
97_____513349_____Skivelitis2
141____318431_____VirtualLarry
144____315963_____Ken_g6
220____126873_____SlangNRox
236____102640_____waffleironhead

Rank__Credits____Team
1_____35084448___TeAm AnandTech
2_____29395016___Czech National Team
3_____27802868___Antarctic Crunchers
4_____26824164___SETI.Germany

Look! Up in the stats! It's a bird! It's a plane! No, it's Souperman! He's flown into the lead!

 

[VirtualLarry]

141____318431_____VirtualLarry
144____315963_____Ken_g6

Whoohoo! Moving up.

(Let's see how long that lasts... )

 

[StefanR5R]

So I enabled the video card for F@H on the 1950x and set cpu to 98% for BOINC. It then did one 16cu unit, and its doing it in 14 hours, like the 2990wx is, but only one at a time.

You could reduce the thread count per PrimeGrid task, such that two tasks fit again.

Or you could actually leave BOINC at 100%. The PrimeGrid tasks with their relatively high thread count won't use each hardware thread for the full time anyway, because they repeatedly need to wait for RAM accesses or for synchronization between program threads.

However, I suspect the best option for running PrimeGrid in parallel with Folding@Home (one or more GPU slots, no CPU slot of course) is to set BOINC to use 50% of the logical CPUs, assuming SMT is enabled in the BIOS, as per default. (BTW, even if you wouldn't want to run a GPU application along with PrimeGrid, PrimeGrid's LLR-based applications would only profit very little from SMT. There may even be cases in which it does not profit from SMT or HT, or in which HT is even somewhat detrimental.)

Though whatever you do, expect both applications to decrease performance of the other to some degree. PrimeGrid's LLR-based applications are heavy on the vector units (and thereby may decrease processor clocks, depending on the CPU) and heavily use the processor caches, the infinity fabric, and the memory controllers. And I presume F@H's GPU application depends a lot on the latter three too.

The upshot is, a F@H + WCG combo usually works out better for F@H performance than a F@H + PG combo. (Well, there are now less than 2 days left to go in the PG challenge…)

 

[StefanR5R]

@Icecold is back in the top 10! :-)
…and about to overtake xii5ku. >:-(

 

[StefanR5R]

There are 3½ more days to go, will we finish in 1st too? The competitors are quite close.

Looks like we built ourselves a bit of a buffer between us and the competitors yesterday and today.

And, I presume, we aren't quite finished with the buffer building just yet, going at a pace of 14 MPPD now. Though SETI.Germany are busy themselves too (13 MPPD).

 

[VirtualLarry]

Put the ram at 3200 and the fclk to 1600, same voltage.

I did this, last night when it semi-hung. It just rebooted again now. Not sure whats up. Now suspect, mobo, PSU (*), or UPS.

(*) I think that this is most likely, though, if I stp mining, and only run 10 threads on PrimeGrid, and it happens again, I'm going to suspect the CPU itself.