Suggestions for best prime projects

[ryzenmulti]

Hi all. Would like to pick the collective brains of the seasoned experts here, and take you along this journey to see where it goes. In the next few days (hopefully!!) I am bringing on line 2 new boards, each has 2x 48C AMD EPYC and 2x Titan V GPU's. So in total 192 cores (not SMT) and 4 GPU's. Linux operating system each with 64GB ram. I can fix the GPU fan speeds with linux with much less bloatware overhead. I am prepared to install windows if there are challenges that work on windows only.

What is the collective wisdom on how to best put these machines to use (project/application settings) to target whatever we are trying to achieve for the Tour de Primes? I'd like people to think of these as team machines and pick your experience to see how to best apply it for the challenges. Max CPU can be quite a bit larger if this is of benefit - I use max CPU of 6 now for current machines.

Obviously cc_config to use all GPU's ... what project preferences for CPU/GPU would maximise the return for the team?

(offline, these will be working on my own personal projects)

[chooka03]

Hmmm for TdP....tough question mate.

I guess running PPSE on the cpu would find the most primes. I'm not sure what the Titan's are best used for with Primegrid. You'll have to do a few comparisons.

I run 1 x PPSE / core as they are small tasks. Again with the Titan's, you would have to run say 1 x GFN16 lets say, get a few results then apply the app_config to 0.5 w/u and see if there's a increase in results. (ie 580/sec to run 1 task but 470 seconds running 2. Two would be better throughput.

Or you could put the Titan's onto GFN-18 and try to jag a big one. Again... not sure how fast/slow that will be. Smaller primes might be better suited to the Titan?

If I had that hardware right now..... probably PPSE due to the shear number of cores. Maybe GFN-16 on the gpu's because gfn-16 won't be available for this challenge next year. There's going to be a HUGE drop in primes found. I kinda look forward to that.

 

[ryzenmulti]

Thanks Chooka. I've been playing around with different WU's for a few days each to benchmark CPU and GPU. CPU's doing PPS, PPSE (good credits) Sierpinski/Riesel Base 5 Problem (poor credits) - have never used less than 6 cores per task. Some PPSE's take up to an hour for good credits - some are very short.

{edit} I see PPSE does not recommend multithreading, but the PPS does. jmswn89 has found T5K primes on PPS - so have biased PPS over PPSE ????

The titans do a GFN16 in under 90 seconds on average (not multithreading); they do a mega 'do you feel lucky' in just under 24 hrs. So getting credits is nothing. CPU's I have tried in groups of 6, 7, 8 and currently 9. I do this to allocate free CPU's for the GPU and improves CPU completion

e.g. on a 24 core machine, 7 cpu per task gives 3 cpu tasks and 2 full spare cpus for the gpu; I have noticed real GPU performance improvements doing this (I never have all CPU's doing CPU work); For the 96 core epyc I use 9 (but will go back to 7 (13x7 = 91) for improved big task speed. I never want to bottleneck the GPU and rely on settings. Going to d an assault on GFN16's tomorrow to see what 24hrs can produce - waiting for a few megas to complete

[chooka03]

I never use hyperthreading (SMT) when crunching any Primegrid. I allow the spare threads to run the GPU.
I think I remember a project once (can't remember which) and each time I reserved a core, the gpu tasks took less time. But I got the the point of not wanting to give up any more cores for the small gpu boost.

You know what.... I just looked and I never noticed that hyperthread WAS recommended for PPS! Well there you go.
Yes, I always reserve at least 2 threads if I'm running a gpu task. I set my 3950X to 95%.

Interesting...... I should try that (running SMT)....... but I might just stick with whats working. They take long enough on some of my pc's!

You might find you can load the Titan up with multiple wu's. I have found the sweet spot at 2 wu. Three tasks and it actually became slower. Same with my slugs- Radeon VII's :) 230sec running 2 consecutive wu's. The 4070Ti's do 2 in 85sec. 

[ryzenmulti]

Thanks chooka, sounds like I am on the right path. I disable SMT and H/T on everything so I wouldn't recommend enabling it. Have now set the titans to 2 wu's as above. Rest the project to GFN 16/17/18 and PPS LLR (not extended). Will leave it be for 24 hours and see what comes out of it.