[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference hydra::amiga_v1

Title:AMIGA NOTES
Notice:Join us in the *NEW* conference - HYDRA::AMIGA_V2
Moderator:HYDRA::MOORE
Created:Sat Apr 26 1986
Last Modified:Wed Feb 05 1992
Last Successful Update:Fri Jun 06 1997
Number of topics:5378
Total number of notes:38326

3549.0. "Digital Signal Processor (DSP) for Amiga" by EUCLID::OWEN (It's just a madhouse anyway) Mon Mar 05 1990 18:50

I grabbed this off the usenet... reposted w/o permission (of course).

Steve

---------------------------------------------------------------------- 
I have a BIG SCOOP for everyone.
 
At the February JAUG (Jersey Amiga Users Group) meeting we were
witness to a new product.  I think this will be a real ground-breaking
product for the Amiga; as I explain at the end of this article.
 
The product was "The Bonsai 2000 DSP board" for the Amiga 2000 by John
Cadona of Cadona Research & Engineering (CRE...  pronounced like CRAY).
After 2 years of research & development, this was the first public
showing.
 
Nutshell Review:
----------------
This board gives you either one or two DSPs in your Amiga 2000 computer
(or A2500, or A2500/30, you get the hint).  The DSP is the AT&T WE DSP
32C at 50MHz.  You can have 0-256K RAM per DSP.  (each DSP has 6K local
RAM).  The hardware interface is really well thought-out.  The software
interface is especially good.  This should make the scientific and
audio communities very interested in the Amiga.  The price performance
made this especially interesting.  (see the end of this article)
 
DSPs In General:
----------------
The talk began with an explanation of DSPs in general.  A DSP (Digital
Signal Processor) is a board that can do the high-speed calculations
required to do sound manipulation.  They were originally designed for
telecommunications (digital phone systems, etc.).  Of course, anything
that can do such calculations can also be programmed to do calculations
for other applications.  So, a DSP in an Amiga (instead of a phone
system) can be a next generation math-coprocessor.  Though, that's
really an insult to the chip.
 
A "signal" is a sound-wave traveling down a wire.  A computer translates
that into a series of numbers to represent what the signal did over a
period of time.  If you can manipulate those numbers fast enough, you
could manipulate the sound as it's happening (real-time manipulation).
 
Amiga sound stores 8-bits for every "number" when it digitizes or plays
sound.  A CD player uses 16-bits (or 18 ...depending on how you think
about it) per number.  A CD player plays at 44.1MHz; this means that it
records 44.1 million numbers per second!  Now you understand why a DSP
has to be able to do fast math.  If you have an equation that will
lower a sound 2 octaves and you want to do it in real-time; you need to
do that equation 44.1 million times per second.
 
DSPs store their numbers in different formats.  The DSP used in the
NeXT machine uses fixed-point numbers.  The AT&T chip uses floating
point (FP) numbers.  This gives this chip an advantage because it can
use a wider range of numbers.
 
For example, 0dB is a whisper that can be barely heard by the human
ear.  120dB is painful to humans.  The floating point format that the
AT&T WE 32C can hear BETTER than a human ear.  It can hear a larger
range and it can hear at a better resolution.
 
How does a DSP do what it does?  Here is a simplified description:  If
you have a 10-second sample you may have thousands of numbers.  A DSP
can perform the same calculation over each and every number at a
blinding rate.  This calculation can modify the signal or reach some
aggregate statistic, etc.  Most all wave theory requires such
calculations as do other mathematical work.
 
The AT&T WE 32c DSP Chip:
-------------------------
 Uses floating point (can "hear" better than a human).
 6K of internal RAM right on the chip.  External RAM helps performance.
 2 internal processes can happen at once.  If you want to multiply one
sample by 5 and add the previous sample's value the DSP can start the
next multiply while the current "add" is being performed.
 12.5 MIPS.
 25 MFLOPS.
 16 Meg of RAM max.
 16-bit parallel port.
 16 megabaud serial port.
 Internal/external communication between different parts of the chip.
 Special FP format (single precision) but can convert to/from IEEE
single precision on the fly.
 
Since this chip is the choice of chips for scientists; and since it is
*the* DSP chip used by AT&T/Bell-foo/Bell-frob researchers it has a C
compiler, assembler, linker.  It has application libraries,
simulators/debuggers, etc.  The application libraries are nice.  Need
to process an echo (reverb effect)?  Just link to the correct
application library can call the routine!
 
Applications:
-------------
John had a long list of possible applications.  I'll try to list them
here and explain some of the more interesting ones.
 
 Audio processing & manipulation (could be a VoiceMail processor).
 Sound synthesis (voice or music) -- Roland synth. quality.
 Spectral estimation & detection.
 Echo cancellation.
 Modulation/demodulation -- It can be a high-speed modem or FAX.
 Encryption.
 Filtering.
 MIDI Integration -- Play keyboard & have DSP work on the sound.
 VERY high-end audio -- More than CD-quality, so use an Amiga with a
lot of RAM and a large hard drive as your audio studio.  When you are
done, submit it to a CD-manufacturer for pressing.
 Signal generator -- Generate any tone you want.
 Speech recognition (do the audio processing for your software).
 Imagine processing.
 Pattern Recognition.
 Servo-control/robotics.
 HAM Radio.
 Radar processing -- Could sample AM radio right off the antenna!
 Sonar processing.
 Personal super computing (more on that later).
 Mathematics modeling -- a current fad in the scientific industry.
 Data collection.
 Research.
 Education -- Use it's power to draw something that a student is trying
to visualize.  He gave one example of something that is difficult to
understand, but if you see it animated in real-time it all comes clear.
The animation requires the horse-power of a DSP.
 
Graphics Applications:
----------------------
 -- For use with graphics, the Bonsai2000 could do translation,
scaling, rotations, shading, perspective and other data manipulations.
These all require a lot of mathematical calculations that the DSP can
do.
 -- Realistic scene generation will be more possible.  Ray-tracing,
radiosity ("glow"), and mapping & projectional methods are all
math-intensive.  Imaging wrapping an image around a 3D shape in an
instant.
 -- Scientific Visualization often requires a CRAY.  How many people
would settle for 1/2 a CRAY all to themselves at this price instead of
millions of dollars for a CRAY that they have to share with other
researchers?
 -- ALMOST REAL-TIME RAY-TRACING.  Ray-tracing is slow because it
requires so much floating point math to be performed.  This is what the
DSP does well!
 
The Card Itself:
----------------
John offered to pass the card around.  After hearing it's price no body
wanted to take responsibility of passing it person-to-person.  It was
pretty funny.
 
I was impressed when I got hold of the card.  The hardware is finished;
no patches were showing on the board.  I remember when the first RAM
cards for the Amiga were released and AmigaWorld did a review of a
bunch of them.  They all had hand-soldered last-minute patches all over
them except for the ASDG board, which looked completely professional.
This card looked the same way.  It looked like it had been in
production for a while and was done by a very serious developer (which
describes CRE).
 
Configurations are flexible:  You can vary the kind of performance you
want based on how much you can afford to pay:
 -- 1 or 2 DSPs running at 50MHz.
 -- 0, 32, 64, 128, 256K RAM per DSP.
 -- Using PALs, you can reconfigure the memory map of the DSP.
 -- The less-expensive "DSP32" can also be used (slower performance).
 -- 8Kx8 or 32Kx8 SRAM (8- or 32-bit bus)
 
16-bit parallel interface to Amiga w/DMA access to all RAM memory and
control registers.  In other words, you can go "straight to the chip".
 
Uses static RAM, 35ns. (as a comparison, 150ns RAM is used in the
Amiga 1000).
 
Runs in parallel with the Amiga.
 
High-speed 32-bit bus on the card for communications between the
components.  You can plug boards into this 32-bit bus to (1) give the
first DSP more memory (2) parallel access to other devices (3) add
even more DSPs.  An example of (2) was to add some kind of medical
device to the bus so that the DSP can monitor an experiment or a
patient.
 
General I/O support via daughter boards:  Personality modules on each
DSP.  Attach samplers, CDs, etc.
 
Tightly couples sync & control via flexible interrupt system between
all Bonsai subsystems and the Amiga.
 
Demos:
------
He showed a sound-wave graph of his voice and his wife's voice (each
person had said the same words) and we were able to see some
differences that make voice recognition difficult.  Another example
(not shown, but talked about) was the voice print of "How do you reck a
nice beach?" and "How do you recognize speech?" and how similar they
are.  That's why computer voice-recognition is a long time away.
 
He held a microphone to one end of a slinky and dropped the other end.
The graph generated by the Amiga showed how the slinky...  ummm...
slinked.  When you use your ear for this experiment, so many sounds
happen at once that you can't tell what's going on.  This was all plain
as day when you could see the individual waves graphed on the screen.
 
He spoke into a microphone and had the DSP output his voice with echo
(the room had a nice audio system so it was interesting).  John used
the DSP to generate a couple different tones.  At one point he started
getting a lot of feedback but it wasn't harsh feedback and it
eventually evened out and became one steady tone.  He explained how the
DSP had just found the perfect pitch of the room and how that can be
used to design "the perfect audio system" for that room's acoustics.  He
then shut it off before the ceiling would start to crumble (Was that
live or Memorex?)
 
Price:
------
Native Development Kit is $500 for the software.  This does everything
on-chip.  Results are echoed back to the Amiga.
 
Amiga Development Kit's price is unknown.  This lets you compile, etc.
on the Amiga and download code to the DSP.  It will include a .library
for accessing the DSP so that all languages can access it.  1/2 a CRAY
accessible from AmigaBASIC!
 
The boards themselves will run from less than $1500 (1 DSP with no
external RAM) to $4000-$5000 (2 DSPs with 512K RAM).  No prices on
"personality modules" or "daughter boards" yet.
 
Implications:
-------------
The NeXT has a Motorola DSP 56001; which is inferior to the AT&T model.
The AT&T chip is the chip-of-choice for serious research.  It's faster
and it's floating-point.  This means the NeXT has one sophomoric DSP
while your Amiga could have 2 superior DSPs.
 
AT&T/Bell-frobs can buy Bonsai boards because AT&T/Bell-frobs must all
use the AT&T DSP chips.  This means it has a large potential market.
 
This is the big one:  John claims "Best price/performance ratio of any
FP DSP in the industry" with this board.  I believe him.
 
Summary:
--------
Wow!  This board can make an Amiga 2500/30 more capable than a NeXT for
many applications.  It is the least expensive price/performance ratio
that can be found.  The software support is excellent.  The design is
great too.  It was shown at EuroDevCon '90 and it made many people
interested.  I predict that if this board is marketed correctly it can
really make the scientific community take notice of the Amiga.  All
this comes at a time when C-A is stepping up its college/
university-orientented marketing.
 
-Tom
(John Cadona is on the network and might be reading this, hopefully
he'll correct any factual errors.)
 
P.S.  Also at this JAUG meeting I was supposed to give a review/
presentation about AmigaTeX from Radical Eye Software.  Due to the
length of John's presentation it was postponed to the next meeting
(March 30th).  No hard feelings, John!
 
---------------------------------------------------------------------
This article is 100% Copyright 1990 Tom Limoncelli.  Re-print and
copying permission granted only from me.  Permission granted to copy
throughout Usenet/Bitnet/Internet/FidoNet.  I can be contacted at
+1 201 408 5389 or as tlimonce@Drew.edu or tlimonce@Drew.Bitnet or
limonce@pilot.njin.net or "Tom Limoncelli / Drew University / PO Box
802 / CM 1060 / Madison, NJ 07940".
     
Sorry for the long disclaimer but the last time I wrote a long article
like this (different topic, similar magnitude of importance) I found
various "rumors" columns had printed parts of it without permission.
--
Tom Limoncelli       The computer industry should spend more time in front of
tlimonce@drew.uucp              their computers.  Remember when "Look & Feel"
tlimonce@drew.Bitnet                      was what you tried to do on a date?
limonce@pilot.njin.net                                                    -Me
T.RTitleUserPersonal
Name
DateLines
3549.1nitMILKWY::JANZENNoting is a privilege not a rightMon Mar 05 1990 19:528
    nit: CD's do not have a 44.1 MHz sampling rate.  It is approximately
    44.1kHz, a factor of a thousand slower.
    
    Well, I can't afford $2500 for the total kit and don't have a 2000
    anyway.  Too bad, woulda had some fun.
    I wish he'd quoted word lengths.  I don't
     have any data on the AT&T chip, as I do for TI and other DSP chips.
    Tom
3549.2human potentialMILKWY::JANZENNoting is a privilege not a rightMon Mar 05 1990 19:552
    Humans hear about a 20-bit effetive dynamic range.
    Tom