[ExI] simulation

[Bryan Bishop]

On Saturday 15 December 2007, Eugen Leitl wrote:
> On Sat, Dec 15, 2007 at 11:17:17AM -0600, Bryan Bishop wrote:
> > And what if we build our own fiber optics to connect the nodes
> > together?
>
> I thought you were talking linking up volunteer's machines via
> residential broadband. If it's a local cluster, e.g. Myrinet will

We were, at first. You showed that lag would make it useless.

> set you back some 1 k$/node. That leaves you with 2 k$/node,
> or 2 MBuck/1 kNode. Don't forget that you'd need 200 k$/year

See below why it might not cost so much.

> for juice, minus air conditioning and admining (allow for
> 1 MBuck/year operational costs otherwise).

I haven't solved the juice or aircon problem yet. Solar cells wouldn't 
work, not unless I also have a factory making those (again, see below), 
but then we'd need the land to put them on and some massive battery 
and ... doesn't sound good. Of course, putting them in orbit would give 
us as much space as we want, to beam the power back down to the ground, 
but then we're adding a whole next level to these plans: how would we 
make the required rocket? Maybe there's a simpler solution.

> > I wonder why I said we need self-rep, all we need is the ability to
> > make the factories and then let them churn out the cpu nodes. Might
> > take a
>
> We already have those -- they're called silicon foundries, and
> take multiple gigabucks to build and run.

Nope. Those are the big, advanced fabs. I've been doing my homework, and 
the entire semiconductor fabrication process can be done in the home, 
with the right equipment. The originals who started the industry did 
just the same thing: they began with wax and very impure wafers and 
probably 500 micrometer technology. Anyway, the hardest part of doing 
it on our own will be the silicon ingot pulling (which requires 
thousands of degrees Fahrenheit) and the CVD chamber, the lithography 
is simple enough and the chemical washes are also ok. Making 
microprocessors on a first try is unrealistic, but I am confident that 
with OpenRISC that it might one day be possible. And if we're doing 
this much, figuring out how to make an ethernet card can be included 
such as your Myrinet reference.

It will all be inefficient, slow, and terrible, but if it's all open 
source and the results are being published and other people start 
experimenting too, then improvements will be made.

> > while for them to make enough nodes, operating at peak efficiency,
> > but at least we'd be on track.
>
> Let's wait for desktop fabs, which will allow you to generate
> dedicated hardware for neural emulation while being affordable
> (0.1-1 M$). Maybe in another 25-30 years.

Yes, that would be useful. But let's not "let's wait". Let's get it 
done.

- Bryan