I am thinking ahead to a Sprint or the Pentathlon. So I may be posting some ramblings about various projects. Nothing Earth shattering, just results, thoughts, and anything I may have learned.
Amicable GPU is a tough cookie. The older 10*20 app, people were running up to 16 tasks at a time to get the GPU utilized 100%. No more. 10*21 app is a resource hog.
First of all, one task will now send your card to 100% usage.
Secondly, each task requires 8GB of SYSTEM RAM. Make sure your page file is enabled!
Third, each task will use a lot of CPU upfront, calculating the 8GB of Primes that go into RAM, then CPU usage drops dramatically. Leave at least one thread free per task.
Fourth, because of the third note above, processor performance matters for this GPU project.
Under your preferences, you can select....uhm....kernel size I think it is called. Tasks are worth 6,836.19 points.
I used kernel size 23 for my much newer Nvidia cards, and kernel size 21 for my poor old R9-280X's in this test.
RTX 2080Ti - 300 seconds - 70 seconds of CPU (R9-3800X) sample size of 3 tasks. Usage unknown - Linux
GTX 1070 - 1707 seconds - 121 seconds CPU (i7-5820K) sample size of 1 task. Usage 100% - Windows
R9-280X - 3930 seconds - 730 seconds of CPU (FX8320E) sample size of 2 tasks. Usage 98% - Windows. BOINC wisely did not fire up the 3rd card due to only 16GB of system RAM
I remember being pretty shocked at the new RAM requirement, one of my headless machines had only 8GB and running Amicable would cause it to become unresponsive. Took me a while to figure out what the problem was!
Thanks, Skillz. Nvidia-smi just gives an instantaneous 'snap-shot', correct? I used it for Collatz, and got 94%. Good to know we no longer have to do a bunch of hoop jumping to get max results. Bad to know the next set of numbers will cripple the current gen of GPUs.
