Page 1 of 1

folding.stanford.edu/stats/os - Inaccurate Data?

Posted: Mon Jul 03, 2017 3:48 pm
by ryoungblood
Are we sure the stats shown regarding the DCN compute power are accurate? Despite some pretty substantial increases in team output, the TFLOP metrics here have pretty much remained constant/dropped.

How are NVIDIA GPU cores contributing to this metric?

Re: folding.stanford.edu/stats/os - Inaccurate Data?

Posted: Mon Jul 03, 2017 5:22 pm
by bruce
1) I don't have any information

2) The analysis codes (either OpenMM or GROMACS) used by the FAHCore uses the loop counters to estimate total FLOPS for each WU. These numbers are summed over time.
Time wasted not-computing gets ignored.

Whatever (in-)efficiency numbers (based on a relationship of the number of FLOPS actually accomplished with the number of GPU cores) get included too.

Re: folding.stanford.edu/stats/os - Inaccurate Data?

Posted: Tue Jul 04, 2017 11:59 am
by foldy
I guess it is summer time. So while curecoin team doubled their PPD from 300M PPD to 600M PPD, other folders stopped their GPUs because the home gets too hot.

Assuming a gtx 1080 ti makes 1M PPD then the curecoin team increase would match 300 gtx 1080 ti which make 11 TFlops each so this is 3300 Tflops.
The overall tflops are currently 80000 Tflops. So curecoin team added 3300 Tflops to 83300 but others paused folding for summer time so minus 3300 Tflops and we have 80000 Tflops again.

It is also possible that some people which already did folding lately joined curecoin team because of the high coin prices. So the overall Tflops do not change but more folders change to curecoin team because currently it pays off.

Will be interesting at Chistmas when we have the highest FAH TFlops/PPD peak if a new record will be set above 100 PetaFlops.
https://docs.google.com/spreadsheets/d/ ... 1016224441