NetNews Usenet Archive 1993 #1/NN_1993_1.iso/spool/comp/sys/dec/6757

Internet Message Format | 1993-01-08 | 3.7 KB

Xref: sparky comp.sys.dec:6757 comp.dsp:2951
Path: sparky!uunet!europa.asd.contel.com!gatech!usenet.ins.cwru.edu!agate!forney.berkeley.edu!jbuck
From: [e-mail] (Joe Buck)
Newsgroups: comp.sys.dec,comp.dsp
Subject: "Are DSP Chips Obsolete?" (was re: Alpha fft performance)
Date: 8 Jan 1993 20:22:30 GMT
Organization: U. C. Berkeley
Lines: 59
Message-ID: <1iknq6$[e-mail]>
References: <[e-mail]> <[e-mail]>
NNTP-Posting-Host: forney.berkeley.edu

In article <[e-mail]> [e-mail] (Michel Fingerhut) writes:
>For some more information on signal processing on Alpha, see our technical
>report "Are DSP Chips Obsolete?" [disponible chez nous].

I have read this report.  I am not impressed.  In fact, if it were turned
in as a term paper for a computer architecture class, I'd give it a B
minus.  If it were turned into a digital signal processing class, I'd
give it a D.

Why?  Because it suffers from "all the world's a Unix workstation" disease.
(On the other hand, many CS grad students also suffer from this disease,
so if the grader knew nothing about the digital signal processing
marketplace your report might fool him or her).

Your paper is correct in the narrow sense that DSPs are not the best
devices to use for raw, open-loop number crunching with no timing
requirements (or timing requirements loose enough so that one second
more or less does not matter).  But DSPs are not being used in such
applications to any significant degree.

The authors did no market research.  Your paper reveals total ignorance
about just what DSP chips are being used for in industry.  Most of these
applications are in the telephone industry.  The authors know what a modem is
because you use one to dial into their workstation, but because they do not
understand the algorithms involved, they are unaware that high-speed modems
do echo cancellation and channel equalization, and these algorithms
require precise timing and have tight feedback loops.  You can't just
buffer up a second's worth of signal at a time as this paper implies.

When you make a long distance call that goes through a satellite link,
your voice is being processed in real time by an echo cancellation
algorithm, which works by adaptive filtering.  Again, real-time
is essential.  The key is that the most demanding DSP algorithms
have feedback: what happens this millisecond depends on what the
world did a millisecond ago.

The paper ignores the issue of cost, power consumption, and weight.
Cellular phones are going digital.  These require real time speech
compression algorithms in a unit small enough to be hand-held and
run for hours on battery power.  For that reason, the future of the
DSP industry is going in the direction of "DSP cores" -- the DSP
is only part of the chip, and the customer can add whatever other
logic is needed.  The goal is NOT to get the maximum number-crunching
power; it is to have just enough performance to fit the application
and its memory on one chip and have it run in real time.  This
means that the engineer has a hardware/software codesign problem:
produce a combined solution that gets the job done quickly and
for low cost.  Sorry, I don't think that Alpha chips are going to
be the answer to the problems the DSP industry currently faces,
though I expect that it will make a very nice workstation CPU.

The distinction between DSPs and conventional processors is not
how fast they do a multiply.  It is their memory bandwidth.  Modern
DSPs have multiple memory banks and typically do three I/O operations
per cycle.  They are best used in embedded systems, systems with
small amounts of memory and very tight timing requirements.

--
Joe Buck    [e-mail]