"fastfib" in the gf package faster than "fib"

Stavros Macrakis wrote:

> If all you want is two processors, you can certainly run a two
> separate Maxima processes on your local machine.  I was suggesting
> that you might want to think about the general case.  But then, I'm
> still not clear what you're trying to accomplish.
> 
> By the way, if you're doing *floating-point* calculations, probably
> the most cost-effective platform (FLOPS per dollar) right now is
> graphics accelerator cards (!), and there are people working on real
> physics using them.  Not something I see Maxima using any time soon,
> though.
>

We have bought one in our lab. However one must be aware that one needs
to do a lot of programming to send the computations to the nVidia board
or use modified programs such as Matlab which know how to use it. 
Using Matlab means losing a lot of speed. And finally if i remember well,
the board does only single precision floating point, which for my needs is
next to useless. In many cases of interest the computation involves
quantities differing by several orders of magnitude, and having a lot
of precision is the only way to get significative answers at the end. For
example i have to diagonalize 100x100 matrices, and i need to keep around
100 decimals to get correct answers. I do that using symbolic computations
programs, e.g. mathematica. If more speed was necessary, i would use
e.g. Gnu GMP. I see a very fast single-precision board useful mainly for
statistical mechanics computations, Monte-Carlo or similar stuff.


-- 
Michel Talon