On Tuesday 16 September 2008 20:22:00 AlgoMantra wrote:
> > So I've already assumed that I need a million cpu's.
> >
> >
> > Yes, I noticed, but this assumption is incorrect, in fact this kind of
> > thing is easier on a single CPU as there is less need to sync stuff.
> >
> > I'm not going to argue this any further as I could probably just have
> > written the whole thing in the time that was now spend on this.
> >
> > No offence; you can try my method or you could send me a outline of what
> > you need exactly (it sounds like Wolfram's CA's on a circular line?) and
> > I'll see when I get round to writing it (probably tomorrow or so) but I'm
> > not going to play "is / isn't" over this. I mean that in the most
> > friendly possible way.
>
> So do I, my friend, but now I think you're so far off the issue that we
> should just close this discussion. I already have the answer I needed, so
> lets just leave it here. There is no point arguing more.
Sorry to jump into this discussion, after you have given it up...
But how realtime do you need this to be? Do you really need your answers in
one or two instruction cycles of a bunch of CPU's running at their full
speed?
I cannot really believe that (assuming you want the output to be audible or
visible). And if you don't, you can easily slice up time in larger chunks and
do calculations for more cells at a time, based on the algorithm Kassen
proposes.
I think you have to separate between the concept of the algorithm and the
actual implementation on real hardware of it. There is usually a pretty big
gap between that, because of how CPU's and the hardware around it is built.
You can write beautifully simple high level code, but at the nitty gritty
hardware level, it is just passing through voltage levels through some chips,
and interpreting the results of that... even in your multi-milliion cpu
scheme, things would not be completely simultaneous...
Synching that many cpu's adds a lot of extra delay, especially since in the
algorithm data is dependant on neighbouring data. If you delve into parallel
computing and architecture of such systems a little, you would see that
communication between the cpu's is actually quite costly.
So, to get around all these issues, I suggest you just write your algorithm in
conceptual code and publish that. Conceptual art does not suffer from
implementation issues. That is, as long as you keep it conceptual.
sincerely,
Marije
Received on Fri Sep 19 2008 - 12:48:14 BST