PrimeGrid Challenges 2021

[StefanR5R]

I'm second now. I might be first soon! :-D

...but not for too long, I presume. :-/

 

[Markfw]

How many 4 ghz cores would it take, to take first ? And what special projects should we select ?

 

[crashtech]

How many 4 ghz cores would it take, to take first ? And what special projects should we select ?

The Primegrid project is a GPU project mainly, GFN-21 will run on CPU but is only effective on high core count CPUs with multithreading set up.

 

[Markfw]

The Primegrid project is a GPU project mainly, GFN-21 will run on CPU but is only effective on high core count CPUs with multithreading set up.

OK, forget it. I was thinking of the project a while back where the 5950x's were kicking butt.

 

[StefanR5R]

A 5950X may easily be faster than a "small" or "old" GPU.
I posted some PPD for GFN-21 two years ago.

There is a catch though: 1. CPUs can only run GFN-21, not the other two subprojects which are available in this challenge (GFN-22 and GFN-DYFL). 2. GFN-22 or/and GFN-DYFL give higher PPD than GFN-21 on the same hardware (if that hardware can run one of those in the first place, and can complete them in a reasonable amount of... days). --- In other words, CPUs are restricted to the one subproject with least PPD, while most GPUs can run any of the three subprojects.

 

[StefanR5R]

About the timing of previous GFN challenges:

2020 December challenge at GFN-18/-19/-20 (10 days):

results were to be reported until Dec 31, validations took until March 1​

2019 December challenge at GFN-21/-22/-DYFL (10 days):

results were to be reported until Dec 22, validations took until February 22​

 

[Ken g6]

I collected Day 1 stats last night, but I forgot to post them:

Rank___Credits____Username
1______7786374____emoga
5______3828390____xii5ku
12_____2342654____crashtech
25_____1118785____biodoc
57_____319750_____Orange Kid

Rank__Credits____Team
1_____15395954___TeAm AnandTech
2_____13949426___Antarctic Crunchers
3_____12576691___Czech National Team
4_____6525862____SETI.Germany

I need to stop staying up after midnight.

 

[StefanR5R]

Very impressive, @emoga!

 

[Markfw]

What would a 3070ti do on this ? I can't have this new baby sitting idle for a week ! Maybe I will put it on the 7551 EPYC box for a week.

 

[StefanR5R]

What would a 3070ti do on this ?

GFN-21 ........... ~5h/task, ~770 k PPD (sample size of 3)
GFN-DYFL .... ~21h/task, ~1.1 M PPD (sample size of 1)

(found in top_hosts.php; note to those who look up results in this way: in case of multi-GPU hosts, always verify the GPU type in stderr.txt of each result, the GPU type in the host record may differ from the one on which the task actually ran)

 

[Markfw]

So, the 3070TI is installed on the 7551 EPYC. First using 460 drivers it said computational error. So I upgraded to 495. Now it says GPU missing, and its not on the menu where it says "use GPU always". And the cpu is running the 21 tasks taking 2 cores of 64 !

:Edit: So I decided to do updates. In the middle of the updates, the GPU came back on the menu, but its not using it yet.
Edit2: During the update BOINC tasks and projects disappeared. Just have to wait now....

 

[StefanR5R]

You need to configure your PrimeGrid preferences to allow only Nvidia GPU work, no CPU work (at the top of the PrimeGrid preferences page).
Further down on that page,

• select only GFN-21= "Generalized Fermat Prime Search n=21 (GFN-21 or Genefer 2097152)" / "NVIDIA GPU" - "OpenCL" if you prefer moderate run time at the order of 5 hours.
• select only GFN-DYFL = "Do You Feel Lucky? (Genefer World Record Attempt)" / "NVIDIA GPU" - "OpenCL" if you prefer maximum PPD and can stomach the long task duration at the order of a day.

At the very bottom of the page, remember to click [Update Preferences].
Afterwards, update the project in boincmgr or boinctasks.

I suggest you abort all already loaded CPU work as soon as you changed the prefs to GPU-only. The 7551 is ineffective at this due to its smallish AVX units, its internal multi-chip/multi-CCX structure, and comparatively small cache per CCX. "cpuGFN21" wants high FMA3 throughput, 8+ cores per task, and ideally >24 MB unsegmented CPU cache per each running task. (BTW, multiple cores per task cannot be configured at the PrimeGrid web preferences. An app_config.xml must be written for that, as detailed in the OP of the challenge thread. But as I said, Zen 1 is not going to work well at this.)

"OCLcudaGFN" on the other hand, i.e. the Nvidia GPU port of it, is of course single-threaded.

 

[Markfw]

You need to configure your PrimeGrid preferences to allow only Nvidia GPU work, no CPU work (at the top of the PrimeGrid preferences page).
Further down on that page,

• select only GFN-21= "Generalized Fermat Prime Search n=21 (GFN-21 or Genefer 2097152)" / "NVIDIA GPU" - "OpenCL" if you prefer moderate run time at the order of 5 hours.
• select only GFN-DYFL = "Do You Feel Lucky? (Genefer World Record Attempt)" / "NVIDIA GPU" - "OpenCL" if you prefer maximum PPD and can stomach the long task duration at the order of a day.

At the very bottom of the page, remember to click [Update Preferences].
Afterwards, update the project in boincmgr or boinctasks.

I suggest you abort all already loaded CPU work as soon as you changed the prefs to GPU-only. The 7551 is ineffective at this due to its smallish AVX units, its internal multi-chip/multi-CCX structure, and comparatively small cache per CCX. "cpuGFN21" wants high FMA3 throughput, 8+ cores per task, and ideally >24 MB unsegmented CPU cache per each running task. (BTW, multiple cores per task cannot be configured at the PrimeGrid web preferences. An app_config.xml must be written for that, as detailed in the OP of the challenge thread. But as I said, Zen 1 is not going to work well at this.)

What about GFN-22 ?

Edit: I did what you said. Using 495 drivers, I still get computational error. I abort and update, and it loads another DYFL and same error. doing nothing now.

I have updated, and tried select only 21,22 and DYFL one at a time. All three get computational error immediately.

 

[crashtech]

@Markfw , have you tried:

sudo apt install ocl-icd-libopencl1

sudo apt install ocl-icd-opencl-dev

sudo reboot

 

[StefanR5R]

What about GFN-22 ?

I didn't find completed GFN-22 results from an RTX 3070 Ti from which I could calculate PPD. However, I do know from RTX 3080 that GFN-22 takes 90% of the runtime compared to GFN-DYFL, but gives only 85 % of the PPD. Hence, it's not really worth it on this class of GPUs. Although its PPD is at least better than GFN-21.

Using 495 drivers, I still get computational error. I abort and update, and it loads another DYFL and same error. doing nothing now.

At the beginning of the boinc event log, does it show the 3070 Ti both as a CUDA and as an OpenCL device? (It should.)
When you go to one of the failed results at your user page on the web site, what is showing up as the stderr.txt output?

 

[Markfw]

@Markfw , have you tried:

sudo apt install ocl-icd-libopencl1

sudo apt install ocl-icd-opencl-dev

sudo reboot

That fixed it ! Except the ETA is 50 days......
on a 22 task

 

[crashtech]

That fixed it ! Except the ETA is 50 days......
on a 22 task

It should update its estimate, unless there is something seriously wrong it will be long, but not that long.

 

[Markfw]

It should update its estimate, unless there is something seriously wrong it will be long, but not that long.

at 1.5% done it was 17.4 minutes, so I estimate about 20 hours. How is that for a 22 ?

 

[crashtech]

at 1.5% done it was 17.4 minutes, so I estimate about 20 hours. How is that for a 22 ?

The result I have for a 3070 is 22.3 hours and for a 2080ti, 20.31 hours, so what you are seeing is in the ballpark.

 

[crashtech]

I think DYFL will be better for that card, though.

 

[Markfw]

I think DYFL will be better for that card, though.

I just edited prefs. thats the only thing I selected. After this finishes(18.5 hours total, 16.5 from now), it will be doing that.

 

[Markfw]

I added my 3080TI. Not sure of the ETA yet. I will reply when I have something better than 0.5% done.

 

[crashtech]

@Markfw , It's hard for me to want to take GPUs of F@H, traditionally I give Folding everything I have all December long, BUT... we've been very competitive this year in the PrimeGrid challenge series, so I've put some stuff on it, and it's great that you are helping, too!

 

[Markfw]

Looks like it might be 15 hours !

Oh, and the reason I decided to help, even without my most powerful 2 cards, I am doing over 40 million ppd, slightly more than I estimated at the start of the race @! The new drivers really helped.

 

[crashtech]

I've only taken one 3070 off of F@H, but the moment PG is over it's going back on.