From: Eugen Leitl (eugen@leitl.org)
Date: Fri Apr 12 2002 - 03:35:49 MDT
On Fri, 12 Apr 2002, Samantha Atkins wrote:
> If you need to do any reasonably fast, or even as fast as neurons
> propagate information (200hz), exchange of much information with
Once again, a modern PC system can't handle GBit/s bitrates. Not just
processing, just receiving or sending it will max out the CPU. (Marginally
better architectures could reduce the load considerably, but the basic
problem still stands: current networking to processing capabilities relate
to trying to drink from a waterhose (or, rather, a high-pressure
pipeline)).
Current ballpark figures of residental WANs is 40 kBytes/s (and about 15
ms latency) is sure not GBit rates, but 40 kBytes/s of a tight code
communication is not something that can be shrugged off.
You can do a lot if you're being smart about the constraints of the
medium. (Inversely, you can piss away orders of magnitude ineffectively if
you're being dumb about it). For instance, Ruediger mentioned the temporal
jitter ruining spiking code precision, which doesn't occur if you put
timing info into the packets (clocks precision is the limiting factor, not
the ping timing distribution), and it doesn't have to be a spiking
encoding on internode comm, if the medium is adverse to it. Spiking codes
don't even map to current computers efficiently, which can do high-speed
arithmetic operations on 32 and 64 bit data types at each clock cycle
essentialy for free (it's a stupid way to use transistors, but since the
architecture is hardwired at the foundry, you have to live with it).
> systems as far away as the other side of the world (or even more than
> about a thousand miles) then you have problems due to speed of light
Your hidden assumption is that global communication is required, always.
This is not true for the biological CNS. Your second hidden assumption is
that the system is a singleton. Instead, it will be most likely a
hierarchical architecture, a swarm intelligence, or a diverse population.
> limits. You might have local units and drones that far away but I
The problem is not the speed of light, at least not as long as we don't
have purely photonic cut-through networks. Current latency is dominated by
the store-and-forward paradigm. Having this said, mammalian CNS does about
120 m/s vs. 299792458 m/s. That's a size scaling factor of roughly 10^6,
if you can do relativistic speed cut-through.
> don't see how you can make a mind a million times faster than human
> minds distributed over that kind of area. But perhaps the flu bug
> attempting to bite me tonight is addling my wits. What am I missing?
Hope you'll get better soon.
As to 10^6 speedup, of course you wouldn't want a distributed system. But
we're talking about bootstrap here. The 10^6 speedup is reserved for the
time when the system designs and builds its own hardware, which is a
distinctly superhuman stage of development.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:26 MST