/*
* INMOS TRANSPUTER PW/424
* Parallel Processing - Unparalleled Potential
*
*
* June 1986
*
*/
/* Introduction */
You're watching a demonstration of power, speed and exceptional performance.
You're watching a graphic sequence generated by the unique parallel processing
capability of an inmos transputer. At last you are watching - a whole new
ballgame.
/* Butterflies */
The technology which propels this butterfly is generated by a single
microprocessor the IMS T414 32 bit transputer. It's a complex task to drive
the graphic screen and generate the animated picture. While at the same time
calculating the perspective of a three dimensional object in a two-dimensional
medium. A second butterfly is added and now both the unrelated processes are
running on the same transputer. Yet another yellow butterfly is drawn this
time by a second transputer and displayed on an adjacent screen. The
synchronization of the two yellow butterflies is maintained by passing
messages along a link between the two transputers. This unique ability of
transputers is further illustrated with the orange butterfly. As the butterfly
flies along the link between the two transputers, perfect synchronization is
maintained. This is despite the fact that at one point of the flight the two
halves of the butterfly are being generated on adjacent screens by the two
transputers. It's this ability to communicate freely between one transputer
and another which puts this microprocessor ahead of the field.
/* Linear scaling */
The problem with conventional microprocessors is that they communicate using a
shared bus and because only one transaction can take place at a time the
result is a communication bottleneck. Not so with the transputer. Its links are
not shared. Each one connects just two transputers. This allows each
communication link to operate concurrently and yet independently of all other
links. In fact communication capability increases as more transputers are
added into the system. But as the curve on the graph dramatically shows this
isn't possible with a conventional microprocessor. As more are added they
reduce the performance of the overall system. And the peek would typically
occur with about three to five processors on a single bus. Contrast this with
the red curve which shows that the performance of a multi transputer system
actually increases linearly as more are added.
/* Ray tracing */
Joining transputers together means that the speed in which a problem can be
solved increases proportionally as this display dramatically illustrates. Here
the left hand screen is being driven by one transputer while the one on the
right is driven by an array of eight. The left hand display is actually been
drawn four times faster than if an IBM PC/AT was being used. As similarly as
the right hand display is driven by eight transputers it is being built another
eight times faster. It's a striking illustration of concurrency in action.
Independently computing each pixel color and assembling the lines involves many
millions of calculations. But the task is distributed over a number of
transputers. Spreading the workload in this way leads to a dramatic reduction
in the time taken to build the complete display. The massive computational
power available from a network of transputers is demonstrated by a mathematical
object of infinite complexity: The Mandelbrot set.
/* Mandelbrot set */
To generate this image requires up to one hundred and fifty millions floating
point operations. But this is just the start. The boundary of the Mandelbrot
set is a fractal. This means that if you zoom in on any point of the display
you find even greater complexity. In effect the operation can be performed to
infinity always finding more and more complexity within the image. A Turbo
Pascal version of the program running on an IBM PC/AT will take between six
and fifteen hours to generate each image. Here various views of the Mandelbrot
set are being generated interactively by a network of ten transputers. They
take between two seconds and two minutes to generate the images.
/* Conclusions */
Parallel processing also means flexibility because of the ease of merging
functions. This startling image results from combining the ray tracing of
Newton's cradle and the Mandelbrot set displays. It's yet another example of
the unparalleled potential of parallel processing using the inmos transputer.
At last - it's a whole new ballgame.
/* EOT */