Folding Forum

Replying to @PantherX - I have two systems dedicated for folding: Ryzen 9 3950X + RX560 GPU - one CPU slot, 12 cores (SMT off), one GPU slot 2P EPYC 7601 - two CPU slots, 32 cores (SMT off) The other two systems are used throughout the day: Ryzen 9 3950X + RX5500XT - one CPU slot, 12 cores (SMT off)...

What do you guys need - more 10-12 CPU slots or fewer bigger slots? Or mix of both?

Finally got a core 0xa8 WU on the 64 CPU folding slot. TPF varies significantly over the run. This is most likely the result of memory and CPUs being scattered around due to no affinity, then the scheduler or automatic NUMA balancing trying to make things better. Only 10 threads launched though even...

A really quick fix (for me) would be to enable the GROMACS mdrun -pin -pinoffset -pinstride and -ntomp (and allow overriding the -ntmpi) - these could be submitted via slot extra-core-args. On EPYC pinning is a must (to take advantage of the huge L3) and if there is shared memory, shared data, sync ...

My normal slot config is two CPU slots, each with 64 cores. Currently one of those is paused and I have a mix of sizes 4/8/12/16 CPU running to characterize this a bit more. Below is a grab from core 0xa8 on the 12 CPU slot. If not obvious from the log, SMT is disabled so this is a dual proc machine...

I am seeing significantly worse performance with core 0xa8. A 128 physical core 2nd gen EPYC machine that gives 4M - 5M PPD on core 0xa7 with TPF in the 10-20 second range is instead delivering 100K PPD with TPF in the 4-5 minute range. That machine has two 64c folding slots. Since FAH does not have...