[rescue] San Antonio, TX area rescue

[Joshua D Boyd]

On Fri, Dec 06, 2002 at 04:25:30PM -0800, Francisco Javier Mesa-Martinez wrote:
> On Fri, 6 Dec 2002, Joshua D Boyd wrote:
> 
> > On Fri, Dec 06, 2002 at 03:25:17AM -0500, dave at cca.org wrote:
> >
> > > One weird thing about the Connections Machines is that lisp was
> > > highly optimized for them - it was supposedly the best language
> > > to program the beast in. Arguably, the CMs were "lisp supercomputers"
> > > (a slight exageration, but not by much).
> >
> > It's not that strange if you take into account that the original purpose
> > was to be an AI machine, and this was the heydey of lisp for AI.
> 
> Nope, this is a common misconception. Thinking machines got its name
> because at that time, the brain was supposed to be a SIMD like machine,
> and the C-1 (and later the c-2/200) were the biggest SIMD machines
> proposed at that time. Therefore they claimed that this is a "thinking"
> machine since it mimics part of the neuronal processing with its single
> bit serial processors. Plus the machine uses storage and processing
> clusters, much like neurons are used for storage/processing in the brain
> (plus nearest neighbor interconnection fabrics which also mimic neuronal
> synapses). This is purelly for a machine that mimics the "processing" of
> the brain, but has nothing to do with the cognition part of it, which is
> wat AI usually deals with.
> 
> The CM machines were geared towards scientific processing, not AI. Since
> the machine was not expectacularly fast under *lisp (surprise there), and
> then they found out that actual FP performance sucked, thus they added a
> bunch of Weitek co-pros to speed up the FP ops.

Scientific processing isn't what Hillis say's their aim was in his 1989
article in Physics Today (archived at
http://www.kurzweilai.net/meme/frame.html?main=/articles/art0504.html?m%3D3).
To quote: "I was trying to design a computer fast enough to solve common
sense reasoning problems. The machine, as we envisioned it, would
contain a million tiny computers, all connected by a communications
network. We called it a 'Connection Machine.'"

Later in that article, Hillis says "Feynman's insistence on looking at
the details helped us discover the potential of the machine for
numerical computing and physical simulation. We had convinced ourselves
at the time that the Connection Machine would not be efficient at
"number-crunching," because the first prototype had no special hardware
for vectors or floating point arithmetic. Both of these were "known" to
be requirements for number-crunching. Feynman decided to test this
assumption on a problem that he was familiar with in detail: quantum
chromodynamics."
 
> Usually parallel lisp in SIMD tries to map recurrence and evaluation loops
> into a systolic pipeline sort of processing. Since you are limited by the
> I/O constraints of the system, you start one evaluation at one IO port and
> let it propagate through the array. You can do things like futures, and
> such to take care of evaluations which are still pending in the recursion
> pipeline.
> 
> In any case the strength of the CM's was C* which had very good support
> for SIMD parallelization of loops and plural variables, which allowed the
> machines to operate as very deep "Vector" machines, so most of the
> previous code in CRAY et al could be easily ported. However you have
> monstruous latencies, huge throughput though...

I'd love to see more about C* and *Lisp someday.

-- 
Joshua D. Boyd