PrimeGrid Challenges 2022

[cellarnoise]

I try to obscure joke some with witchcraft.

I think TTony gets it some as they are close to the Bayou!

I hope anyway and I bullsheet alot...

 

[crashtech]

Ah Jeez, I missed an exciting one!

HI, Tony! See you soon for the Pent!

 

[Skillz]

Ah Jeez, I missed an exciting one!

It was for sure a good one. Team AnandTech basically took 1st and 2nd in the race. lol

I did too although I don't know if my rigs will hold together for another CPU race, they are very unstable CPU mining, sometimes.

I believe the next one coming up can be run on GPUs. But we got a while before it starts. (Mid-end of May I believe)

 

[StefanR5R]

I believe the next one coming up can be run on GPUs.

Yes. (It's GFN-19 on May 25-30.)

There is a CPU-only application for it too, but it's singlethreaded while cache footprint is at least 8 MB per running task.

 

[crashtech]

In case one wanted to run a bit of CPU on Zen based PCs, there'd be a question of whether the scheduler would properly allocate tasks as not to overflow the caches. My guess is that task pinning like what you have done with other PG projects would be necessary.

 

[StefanR5R]

With EPYC firmware, one could use the L3$-as-NUMA-node thing. Or And probably even better, switch down to 2 cores per CCX in the BIOS. ( = in case of Zen 2 with 16 MB L3$ per CCX. This might turn out uncomfortably tight for 2 GFN-19 instances per CCX, but 0.5 MB L2$ per core adds to the non-inclusive L3$, in theory.)

 

[Ken g6]

Bump! A GFN-19 challenge is only 6 days away!

GFN-19 seems like the middle-child of GFN's. It's not so large you might find a record prime, but not so small that you're likely to find a prime at all. That said it's listed as taking about 48 hours on a single CPU core (you can't edit: and should multi-thread), so CPUs might be useful alongside GPUs.

 

[StefanR5R]

About cpuGFN19: Data allocation is 8 MB (earlier post with links). Which means this application can easily run into a memory access bottleneck.

About OCLcudaGFN19: I looked a little through some top regulars' hosts but didn't find any which ran GFN19 recently. So I started one random task on a Pascal GPU of my own. The following was logged:

Supported transform implementations: ocl ocl2 ocl3 ocl4 ocl5​

[...]​

OCL transform is past its b limit.​

OCL3 transform is past its b limit.​

OCL4 transform is past its b limit.​

OCL5 transform is past its b limit.​

Using OCL2 transform​

The OCL transform is the only one which uses FP64. But as this one is not usable in GFN19, server GPUs and some AMD GPUs with a high FP64 throughput design will not have a special advantage here. The OCL2 transform performs INT32 operations.

(From pschoefer's memory: OCL = FP64, OCL2 = 3*INT32, OCL3 = INT64, OCL4 = 2*signed INT32, OCL5 = unsigned INT32.)

Edit: On a GTX 1080Ti, Linux, no app_config.xml, OCLcudaGFN19 ran constantly at 100 % SM utilization and 80 % memory controller(?) utilization, and at or near its board power limit, at least while I watched.

 

[Pokey]

Am I correct to assume the GFN-19 challenge is an event for individuals?

 

[Orange Kid]

It is both an individual and team event. All you need to do though is to set your PrimeGrid preferences to GFN-19 and of course be on TeAm Anandtech.
This one is best run on GPU.

PrimeGrid Challenge Series Overall Results

 

[Skillz]

Am I correct to assume the GFN-19 challenge is an event for individuals?

You can watch your individual points on this one, but it's also a team challenge.

Unlike the Pentathlon where I got you to start running Primegrid early do not do such with the Challenge series.

Tasks only count if they're sent after the start of the challenge and only count if they're returned before the end of the challenge.

This includes pendings also. All pending credit goes towards the challenge total if the above criteria is met. But, should the task fail during or after the challenge the rewarded points will be deducted.

 

[Skillz]

4 Hours until the start of this. 2PM EDT.

Make sure you set your cache to 0 or 0.01 so you only download one task at a time. Returning your task first is a big deal. Means if you find a Prime number you'll be the one who is credited with the discovery.

The project is GFN19 and runs on both GPUs and CPUs with GPUs being the better option. CPUs have a new Multi threading app so if you do run it on your CPU it's best to only run one or two tasks at a time, depending on the CPU.

Also do not forget;
-Do not start too early. Tasks downloaded before the start do not count.
-Do not keep a large cache of downloaded tasks if you want credit for finding the Prime.
-GPUs are better than CPUs.
-CPUs have multi threading support now.
-Make sure you are in the correct team.

 

[Ken g6]

And we're off! (Almost an hour ago.)

 

[mmonnin03]

PG is also the Sprint this weekend. I was hoping it would be a CPU only project.

 

[Skillz]

27 hours in:

Pos.	User	Team	Points
1	crashtech	TeAm AnandTech	7,132,106.06
2	Icecold	TeAm AnandTech	6,995,452.89
3	[TA]Skillz	TeAm AnandTech	6,315,865.89
9	Pokey	TeAm AnandTech	3,872,408.52
11	emoga	TeAm AnandTech	2,857,451.39
21	mmonnin	TeAm AnandTech	1,803,224.79
31	biodoc	TeAm AnandTech	1,085,209.65
37	Orange Kid	TeAm AnandTech	936,704.82
40	cellarnoise2	TeAm AnandTech	788,525.61
68	Endgame124	TeAm AnandTech	460,445.36
76	Skivelitis2	TeAm AnandTech	421,130.93
90	Fardringle	TeAm AnandTech	343,491.75
132	waffleironhead	TeAm AnandTech	179,469.09
152	Lane42	TeAm AnandTech	148,329.05
177	Ken_g6	TeAm AnandTech	101,468.49

 

[crashtech]

Thanks for the stats, Skillz!

 

[Ken g6]

I guess I'm out of practice of posting stats.

I can at least post the team standings:

Rank__Credits____Team
1_____40749167___TeAm AnandTech
2_____18721801___Antarctic Crunchers
3_____18642719___Czech National Team
4_____15808807___SETI.Germany

 

[StefanR5R]

I started late...

GFN-19 seems like the middle-child of GFN's. It's not so large you might find a record prime, but not so small that you're likely to find a prime at all. That said it's listed as taking about 48 hours on a single CPU core (you can't multi-thread), so CPUs might be useful alongside GPUs.

OCLcudaGFN19: On my dual-/triple-GPU computers with Pascal GPUs (TSMC 16 nm, board power limit reduced from 250 W to 200 W) + Intel 14 nm host CPUs (turbo disabled), I am measuring 1,14 MJ energy used per result.

cpuGFN19: On my dual-Rome computers (TSMC 7 nm, Epyc 7452 reconfigured in the BIOS to 180 W socket power limit and cut down to 2 cores per CCX, SMT not disabled but using only 50 % of the logical CPUs in boinc), I am measuring 0,64 MJ energy used per result. (First results are still in the making, they take almost 14 hours.)

Remarks:

• task energy (in J) = power draw (in W) × 3600 s/h / avg. task completions per hour
• avg. task completions per hour = number of concurrently running tasks / avg. task duration (in hours)
• Since cpuGFN19 is said to work with 8 MBytes "hot" data, and Zen 2 CCXs have got 16 MB non-inclusive level 3 cache, I decided to run 2 tasks per CCX. To this end, instead of pinning individual tasks to individual logical CPUs for a perfect spread of tasks across CCXs, I reconfigured the CPUs to 2 cores per CCX by means of a BIOS option, as mentioned.
• Even though I cut down the CPUs this much (disabled half their execution units), yet increased the power limit to cTDPmax, they pull all the power they can get. Average processor clock is 3.08 GHz, which is certainly way past their efficiency sweet spot.
• I don't have data for my 14 nm Intel server CPUs (Broadwell-EP). Based on other applications, they would be using at least 2×, perhaps 2.5× the task energy compared to the Zen 2 server CPUs.

 

[StefanR5R]

About that MJ (Megajoule) thing:

Yesterday, which was a holiday here, I happened to go on a cycling tour, 120 km long (75 miles), 820 m up/ 880 m down (2700' / 2900'). My GPS calculated that this took 4800 kcal = 20 MJ. But I don't know how it calculated this; I ate a lot less than that right before/during/after the tour and don't have any body fat to speak of which I could burn. So my guess is that the tour took about 10 MJ.

So I expended a guesstimated equivalent of ≈16 GFN-19 results (based on my 7nm CPUs' consumption) for that 75 miles tour through somewhat hilly landscape. And that's not counting the energy which the wingmen spend to validate the results...

 

[lane42]

1​	Icecold	TeAm AnandTech	12 760 525.50​	1 634​
2​	crashtech	TeAm AnandTech	12 668 347.84​	1 623​
3​	[TA]Skillz	TeAm AnandTech	12 643 374.15​	1 619​
4​	tng	Antarctic Crunchers	12 573 760.78​	1 61​

That's some race at the top

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
1______14918049___[TA]Skillz
3______14261460___Icecold
4______13965220___crashtech
9______7685321____Pokey
15_____4810553____emoga
21_____3818338____mmonnin
30_____2335886____xii5ku
38_____1913379____Orange Kid
43_____1788127____cellarnoise2
62_____1085209____biodoc
69_____959925_____Endgame124
78_____881791_____Skivelitis2
94_____718299_____Fardringle
121____484323_____kiska
152____327986_____Lane42
156____319935_____waffleironhead
188____195211_____Ken_g6

Rank__Credits____Team
1_____70250255___TeAm AnandTech
2_____34676613___Antarctic Crunchers
3_____33329162___Czech National Team
4_____28175673___SETI.Germany

Still a nice race at the top! Apparently I need to upgrade something.

 

[StefanR5R]

it's listed as taking about 48 hours on a single CPU core (you can't multi-thread),

Until right now I was relying on my older info...

About cpuGFN19: Data allocation is 8 MB (earlier post with links).

...according to which multithreading is not available in cpuGFN19. Well that's still sort of true, but...

-CPUs have multi threading support now.

...at the end of April or early May, a cpuGFN19_mt application version has been made available. The current x86-64 version on the server for x86-64 is from May 2, the one for x86-32 from May 18.

Related thread on the PG message board: Generalized Fermat Prime Search : Multi-threaded GFN-19 3.4.0.2.

(Much of the discussion in this thread is about pinning tasks to logical CPUs, which helps on architectures with segmented last level processor caches, as well as about running the CPU application and GPU application on the same host. There are also suggestions in this thread how to implement the task pinning on Linux; the implementation shown there does this by periodically checking for presence of processes with particular names and overriding their processor affinity masks. — A Windows implementation which works periodically too was posted in the SETI.Germany forum. — I have dual processor computers, therefore I prefer to pin tasks to CPUs not periodically but through their entire lifetime, such that they won't be moved to a different NUMA domain while they run.)

--------

Edit: I posted a request to the project owners for better visible communication of application version updates.

 

[Ken g6]

Until right now I was relying on my older info...

Sorry. We discussed that on Discord, but I guess I didn't post it here. I just edited my post. "Blowing the 't' off the can" as my Mom used to say.

 

[StefanR5R]

As far as I am concerned, there was no "damage" done except that I perpetuated outdated information here. :-)

Due to Pentathlon and RL, I doubt that I would have squeezed in the (little) time required to test the new multithreading option. (After all, I didn't even manage to start on time.) Luckily, the TeAm's rank in this race appears to be somewhat secure.

The multithreading option would probably carve some more throughput out of the Rome CPUs. As for power efficiency, CPU clocks would go down a lot, hence they would operate at a more efficient V-f point; on the other hand, there would be some penalty for thread synchronization. — But more important for kit like mine, the option makes it more attractive to enter CPUs like my Broadwell-EPs. Maybe I'll still add them, but this would require me to run a power cable through the hallway again, like I had during the Pentathlon.

Edit: On second thought, since it is raining here occasionally, my apartment cooling setup with open windows will be too much of a bother to change for the addition of further computers.

 

[Icecold]

But more important for kit like mine, the option makes it more attractive to enter CPUs like my Broadwell-EPs. Maybe I'll still add them, but this would require me to run a power cable through the hallway again, like I had during the Pentathlon.

Edit: On second thought, since it is raining here occasionally, my apartment cooling setup with open windows will be too much of a bother to change for the addition of further computers.

Just for comparison's sake/information, my dual E5-2683 v3 machine, hyperthreading on running 4 tasks at a time (14 threads each) is completing tasks in around 4.5-4.75 hours. That seemed pretty decent to me, but there may be a more optimal configuration it's just what I went with to ensure the tasks would fit in cache.