Re: NEC-LIST: CEM question

> Can some help to figure out the computational resources needed for
> such a problem? I am trying to figure out if a souped-up PC can
> handle such a problem.
>
> If one were using FDTD it is 6 variables per node (Ex,Ey,Ez,Hx,Hy,Hz),
> TLM is 18 variables per node.
>
> Let see, if the dimension of the computational volume is 1mX0.7mX1m(?)
> (may be a reasonable size but at 10 GHz..) where L=3 cm one may end up
> modelling (discretise) the volume into 3 mm size cells( using the
> LAMBDA/10 rule) i.e. 334 X 234 X 334 or so cells/nodes ( 26,104,104 )
> and pulling a figure from thin air, at say 80 MB per 1E6 nodes, it
> fill up approx. 2.1GB.

The old code of Luebbers and Kunz, still available at:

     ftp://ftp.emclab.umr.edu/pub/aces/psufdtd/

requires only about 1 GByte for such a problem (it uses 36 Bytes per
cell). With the HP Fortran compiler on HPUX, the speed of this code is
pretty good. I obtained 0.25 micro-second per cell per timestep on our
HP risc processors (which peak at 1700 Mflops). On a fast PC I would
still expect less than 1 us / cell /step.

>
> At 10GHz the period is 334E-12 s and at 3 mm spacing between nodes the
> propagation time per node is approx. 10E-12s...

Timestep must be sqrt(3) or more times smaller, acc. to Courant, so if
you want, say, 10 periods, you get about 579 timesteps, which leads
(for 26104104 cells and 1 us / cell /step) to 4.2 hours.

Of course, 10 periods would be a but small, if high-Q effects are
present.

Also, the newer (commercial) codes, might be somewhat faster or
slower. But the above should be a fair indication.

Jos

--
  Dr. Jozef R. Bergervoet                      Electromagnetism and EMC
  Philips Research Laboratories,             Eindhoven, The Netherlands
  Building WS01                                     FAX: +31-40-2742224
  E-mail: bergervo_at_natlab.research.philips.com    Phone: +31-40-2742403