Hardware Progess: $165/Gflop

From: Amara D. Angelica (amara@kurzweilai.net)
Date: Tue Jan 06 2004 - 04:16:31 MST


Forwarded

-----Original Message-----
From: Robert J. Bradbury [mailto:bradbury@aeiveos.com]
Sent: Sunday, January 04, 2004 5:46 AM
To: Amara D. Angelica
Subject: Re: FW: Hardware Progess: $165/Gflop

Amara Angelica has forwarded some of the recent discussion regarding
GFlops cost to me -- below are a few comments.

I believe Daniel said:

> Moore's law originally referred to cost optimal
> number of components in an integrated circuit
> doubling every year. The trend for integrated
> circuits subsequently has been about 18 months
> per doubling. A 'Moore's Law pace' is therefore
> a change of a factor of two per 18 months, and
> can refer to things other than the number of
> components in an IC.

Fine -- but CPU speed isn't the problem!
Look at:
  Dongarra, J., "Trends in High Performance Computing and
  the Computational Grid", Annual High Performance Computing
  and Communications Conference, April 2-4, 2002, Newport,
  Rhode Island.
  http://www.hpcc-usa.org/pics/02-pres/dongarra.pdf
Slide:
    "Where DOes the Performance Go? or
   Why SHould I Care About the Memory Hierarchy?"
    (pg 20 of 30).

Shows the real problem is memory access time.

> I beg to differ. This chart:
> http://www.top500.org/lists/2003/11/PerformanceDevelopment.php#
>
> shows that supercomputer performance has in fact
> been increasing in just such a fashion.

Supercomputers are a very poor measure of Moore's Law. One
can fool yourself into thinking that performance is increasing by simply
throwing more processors at the problem. And that is largely what they
have been doing. But that has limits. You will find that companies like
SGI, now Cray, IBM and others are having to put at least as much
engineering into the data switches between processor nodes as they put
into the processors themselves.

In fact (though I can't find it now) I've seen one chart for the next
Semiconductor Industry Road Map that shows traditional increases slowing
down. I think this in large part is due to hitting the limits on chip
heat density without more sophisticated (and more expensive) cooling --
you also have to start asking how many pins its going to take to get 55
Amps (100W/1.8V) into a microprocessor?

> > I wonder why you didn't
> > include GBit Ethernet in your comparison.
>
> I don't think the motherboard I priced would
> support that much LAN traffic.

Well then I think you may have a *big* problem.
Assume the brain contains ~60 billion neurons (a reasonably justifiable
number) producing 100 action potentials a second with a fanout (# of
synapses/neuron) per neuron of 1000 and a minimum data value per synapse
of 1 byte (i.e. how much information gets transmitted at each synapse
per action potential). With those parameters the internal brain
bandwidth is ~6000 TB/sec. You can juggle these numbers a bit but you
aren't going to change them by many orders of magnitude.

I'm not saying you can't get this -- it works out to about the
*internal* data bandwidth of 10,000 Fuzion 150 high end graphics cards.
Note the stress on *internal*. The external bandwidth is 2 orders of
magnitude lower.

> Why do you laugh? That figure comes from Hans
> Moravec. While I consider it optimistic, he has
> what I think is a reasonable means of coming up
> with the figure.

Sure but it is primarily dealing with the processing power required to
match a limited subset of the brain. He never goes into the problem of
getting power in, getting waste heat out and interprocessor and
internodal delays particularly when scaling up to human brain
capacities.

> Although I've been working on the concept of a
> replicating factory in the form of a single
> unit that can copy itself, our society as a whole
> has been working towards automating production.

Its already to a large extent been worked out.
For a general overview see:
  Kinematic Self-Replicating Machines by Freitas and Merkle
  http://www.MolecularAssembler.com/KSRM.htm
and for the first serious study see:
  Advanced Automation for Space Missions
  http://www.islandone.org/MMSG/aasm/
Both of these are discussed on the Freitas web site:
  http://www.rfreitas.com/

But we already have self-replicating factories -- they are called cells
and they are extensively used in some types of drug production and
agriculture.

> Imagine if you will that the railroads and shipyards
> fully automate handling of shipping containers
> (those 8x40 foot steel boxes). That factories
> automate the unloading and loading of those boxes
> with raw materials and finished products, etc.

This has largely alreaedy been done. Visit a major shipping port like
Seattle. The cost of the humans to operate the container moving cranes
is very small compared to the capital investment in the port, ships,
containers, flatcars, etc.

> Once all the pieces get automated, your entire
> industrial infrastructure acts as an automated system.

The key point is *all* the pieces -- we are a long way away from that.

Robert Bradbury



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:43 MDT