real*4 or real*8?

Lattice Boltzmann Method LBM: Theory
#1

Hi,
is it always necessary to use double precision (real*8) in this method for u, v, w, rho, f, f_tmp? I try to apply LBM on very fine grids therefore the economy of memory is important.

#2

Hi Kaff,

As I talked with some guys which use GPUs there is no difference for them to use single precision instead of double.

Alex

#3

Of course GPU guys have no problems using floats instead of doubles. :wink:
I am not sure whether single precision is sufficient in some cases. I could imagine that in simulations with tau close to 0.5 there could be problems.

Timm

#4

This paper by P. Dellar[/url] reports issues with round-off errors at very small Mach number. As a work-around, instead of increasing the machine size of floating point representation, the paper reports that good results are achieved with [url=http://www.lbmethod.org/howtos:reduce_roundoff]the round-off optimization procedure by P. Skordos.