Distributed microcontroller computing

[Ben Bradley]

An 'appropriate' microcontroller that can handle that could cost as
little as 50 cents (not counting its share of printed circuit board
real estate, support components, power supply, etc). But consider a
$200 desktop PC does 32 bit operations (four times as much data per
operation) at about 2 gHz (100 times the speed of an 8-bit
microcontroller), so offhand it looks like you're getting the same
computational horsepower with one desktop PC as with 400
microcontrollers at the same cost.

You may be able to do better with 32-bit microcontrollers, or
especially with FPGA's, but in any case you should always calculate
the operations-per-dollar that each solution gives you.

The Connection Machine! ;-)


How many chips are you planning to use? Are you saying it will take
5 X 3 PC-years? What happens if you do this same thing over an ordinary
LAN?

I've written 8087 code, to do the inner loop for a Mandelbrot set on an
AT. (anybody remember them?)

Remember which, the Mandelbrot set or the AT?

There was a "Circuit Cellar" column in Byte many years (okay, a
couple DECADES ago!) ago where Steve Ciarcia made a parallel-processor
board using 8051's (or was it 8048's back then?), each running a
floating-point math package, and each used to calculate points for the
Mandelbrot set. ISTR there was one main processor to communicate with
the PC and the other processors. Something like eight processors was
as fast as a 286 (I forget if that's with or without a FP
coprocessor), and you could put up to 64 processors on the board. This
was neat toy, but still not very useful even back then.

Don't all PCs these days (I assume by
"5pcs" you mean, "five personal computers") have a math coprocessor?

They're "built in" to the main processor, starting with the 486.
The cost per logic gate has continued to go down, so they've stuck
more and more crap on/in the CPU.

 

[Roger Hamlett]

An 'appropriate' microcontroller that can handle that could cost as
little as 50 cents (not counting its share of printed circuit board
real estate, support components, power supply, etc). But consider a
$200 desktop PC does 32 bit operations (four times as much data per
operation) at about 2 gHz (100 times the speed of an 8-bit
microcontroller), so offhand it looks like you're getting the same
computational horsepower with one desktop PC as with 400
microcontrollers at the same cost.

Also, I'd be very suspicious, that if the code will split down like this,
the SSE abilities in a modern PC could be used. It may be that the PC
code, needs some more work using this sort of ability. I'd be suspicious
that this sort of thought, coud bring perhaps three or more orders of
magnitude improvement. As an example, I have a surface 'shaping' task,
which written in C, and implemented using a normal PC, would require
somthing over 1000hours to run. This was re-written to use SSE3, and
implemented on an A64 X2, and solves in less than one second...

You may be able to do better with 32-bit microcontrollers, or
especially with FPGA's, but in any case you should always calculate
the operations-per-dollar that each solution gives you.

The other option, depending on what sort of 'code' structure is needed, is
to implement the basic arithmetic in a FPGA, and have each one handle
multiple cells. This would keep the gate count needed in the gate array
down, below that neeed for a general purpose micro, and allow more cells
per array. However even using the latter, a 8051 for example, is available
as a general 'core' for use in such an array, with a less than 10K gate
count, and cores are available now routinely, with 200K gates, operating
at 500MHz, allowing a single chip to implement a 'macro cell' of perhaps
16 processors. Some of the basic PIC's are available in such a form, and
it might well be possible to push 32 of the smaller designs into such a
module.

Remember which, the Mandelbrot set or the AT?

There was a "Circuit Cellar" column in Byte many years (okay, a
couple DECADES ago!) ago where Steve Ciarcia made a parallel-processor
board using 8051's (or was it 8048's back then?), each running a
floating-point math package, and each used to calculate points for the
Mandelbrot set. ISTR there was one main processor to communicate with
the PC and the other processors. Something like eight processors was
as fast as a 286 (I forget if that's with or without a FP
coprocessor), and you could put up to 64 processors on the board. This
was neat toy, but still not very useful even back then.


They're "built in" to the main processor, starting with the 486.
The cost per logic gate has continued to go down, so they've stuck
more and more crap on/in the CPU.

Best Wishes