Page 1 of 1
What does FAH use? 32bit float vs 32 bit int?
Posted: Wed Sep 09, 2020 1:28 am
by MeeLee
Does FAH use more Integer, or more Floating operations for processing WUs?
Re: What does FAH use? 32bit float vs 32 bit int?
Posted: Wed Sep 09, 2020 3:39 am
by bruce
In theory, FAH computations are primarily FP32 plus limited amounts of FP64 and Int.
Surely somebody has done some actual instruction profiling and can back up my estimates (call them guesses if you want) with measured information, even if it's for GROMACS.
Re: What does FAH use? 32bit float vs 32 bit int?
Posted: Sat Sep 12, 2020 2:59 am
by JohnChodera
Some molecular simulation packages can make great use of INT32 to mimic fixed-precision, but unfortunately, the newer GPUs don't seem to support INT32 for the fancy tensor cores.
~ John Chodera // MSKCC
Re: What does FAH use? 32bit float vs 32 bit int?
Posted: Sat Sep 12, 2020 5:52 am
by bruce
I though I saw some announcement data on int32 for Ampere but I wasn't particularly interested.
There's probably some advantage to 31 bit fixed precision as opposed to 24 bit floats ... probably including speed on a CPU but probably not on a GPU.