
TRANSPUTER
continued from
page 77
Divide
and
conquer
The idea
is to
divide and con-
quer; let's
illustrate with
an anal-
ogy.
If,
for example, you
find that
your
secretary is not
getting her
work done
because of constant
telephone interruptions,
you
hire
a receptionist
and divide the
workload
between the
two, there-
by improving
productivity.
In
a similar manner.
a comput-
er can
divide and distribute
its
workload
by offloading portions
of it
to other processors.
Whereas
dedicated
hardware
-disk con-
trollers and video
adapters,
for
example-
improve
system per-
formance
by
doing some
of the
CPU's work, those
types of con-
trollers typically
differ from
gen-
eral- purpose
processors
in that
they require
step
-by
-step instruc-
tions from
the CPU about
how
to
operate.
By
contrast,
a processor gets
its orders
directly from
the
soft-
ware application
without
involv-
ing the
CPU at all. That
leaves the
CPU free
to do other
things, and
thereby
increases overall
produc-
tivity. In
addition, a general
-pur-
pose processor
can be adapted
to
a number
of
tasks, whereas
the
typical hardware
controller
can
only
do one
thing, albeit very
well.
Basically, there are two dif-
ferent types of
processors:
coprocessors and parallel
pro-
cessors, although the distinction
between them
is
at
times hazy.
Strictly speaking, a coprocessor
is a processor dedicated to one
task and one task alone -like the
receptionist
in
the example. Par-
allel processors, on the other
hand, can function in different
situations.
In what follows, well discuss
several parallel and coprocessing
products for the IBM family. Keep
in mind, however, that compara-
ble products for
the
Apple
and
other
systems
are available.
Number crunching
The most popular and best
known
coprocessors are the
math coprocessors from Intel:
the 8087, 80287, and the 80387.
Those IC's work in conjunction
with
the 8088/86,
80286,
and
80386 CPU's,
respectively,
and
usually plug directly
into
a
socket
on the PC's
motherboard.
As the name suggests,
a
math
coprocessor does
nothing
but
mathematical calculations
-a
task that a general-
purpose
mi-
croprocessor
is very slow at.
The
chip
doesn't
do windows
and it
can't read a disk -but
it can mul-
tiply two
numbers as much
as
100 times
faster than a CPU
by
itself.
lb give you
an idea of how
a
math
coprocessor can improve
performance, consider the
math -
intensive CAD program,
Auto -
CAD.
Without a math
coprocessor, even simple
Auto -
CAD drawings can
take several
minutes to compute for display
using just the CPU as
a number
cruncher.
However,
when a math
coprocessor is installed,
Auto -
CAD recognizes
its
presence
and
diverts all
math routines to
it
rather
than
forcing them
through the CPU. Then,
while
the coprocessor
is busy munch-
ing
mantissas,
the CPU is analyz-
ing the results and placing them
on the screen.
The outcome of
that
partnership is a tenfold
in-
crease in speed.
You want to
do
windows?
If it's screen performance you
want, there are coprocessors
for
that, too.
Popular IC's include
'I xas Instruments 34010,
Intel's
82786, and others
from Hitachi,
National Semiconductor,
and
others. Like the
math
coprocessors
discussed above,
video coprocessors are optimized
for a single task:
manipulating
graphics on a video screen. They
can't access a disk drive or read
a
keyboard, but they
sure
can
move pixels around!
Unlike the simple
interface be-
tween the microprocessor and
the
math coprocessor, video
coprocessors usually
require a
good deal of support
circuitry.
Commercial products include
the Pepper 1600 from Number
Nine Computer and the Ultra
2000 CAD from INI Computer
Systems.
Unlike typical
IBM -style video
controllers
(Hercules,
EGA.
etc.),
those
video
adapters
use
video
coprocessor
IC's that
receive
dis-
play
instructions
directly
from
applications
software.
In that
sit-
uation. the
CPU
is relegated
sim-
ply to coordinating
screen
opera-
tions
with keyboard
input,
disk
access,
and other
housekeeping
system
functions.
Parallel
processors
Parallel
processors
are
more
versatile
than
dedicated
coprocessors
because
they can
be
programmed
to do more
than one
thing,
be it number
crunching,
video graphics,
or just
plain sys-
tem
routines.
Like a coprocessor,
a parallel
processor works
alongside
the
main CPU and
receives
its in-
structions
from the software
ap-
plication rather
than from
the
CPU. Generally,
each parallel
pro-
cessor is an
island unto
itself. It
is assigned
a specific task.
it does
it, and then it
waits for another
assignment.
A crude analogy
to parallel
pro-
cessing is
a local -area
network
(LAN). in which information
is
accessed by an
individual
node
from a shared data
bank, and
is
then
processed according
to the
user's needs. Then the
results
are
returned to the
data bank so
that
others
may make use of
it.
Parallel processors
handle
things
on a
more basic
level.
Whereas the
LAN may have
more
than one application
running on
the
network, with each
user
using
a different piece
of soft-
ware, the
parallel processor
is
driven by one
all- encompassing
application that
takes total, con-
trol of each processor.
The transputer
An increasingly
popular paral-
lel processor
for PC applications
is called a transputer,
an acro-
nym
derived
from TRANSistor
comPUTER. The term
was coined
by
INMOS Corp. of Bristol,
En-
gland, (the same
company
that
makes the
DAC used in
most
VGA adapters);
it basically
means a computer
on a chip.
In fact, INMOS makes two
transputers: the
T414 and the
T800. As shown
in Fig. 1. the
T414 is a 32 -bit CPU with
2K of
83
Comentários a estes Manuais