DSP - a historical perspective

[DrJustice]

I found the the following bit of amazing fact form an interview with mosc (I take the liberty of quoting Play, the interviewer):
"In 1981 you began a 15 year career at Bell Labs where you participated in the design of some of the first DSP chips."
Go here for the full story: http://noiseusse.org/interviews.htm?id=2

Holy tubestage!
Our own Howard Moscovitz - One of the Fathers of DSPs!

I'm not worthy

What an adventure it must have been... Please dearest mosc, can we gather around the fire and hear the story about the first DSPs? Are there any papers on this?

DJ
--

[mosc]

Yikes, calling me the Father of DSPs is a huge overstatement and a disservice to many other people who worked on the first chips. I was just one member of a team. The people who deserve that title are Jim Bodie and R. A. (RAP) Pedersen. Jim worked in Holmdel, New Jersey and came up with the idea of a microprocessor with an architecture and instruction set built around a fast multiplier/accumulator, the essential operation in all digital signal processing. He searched for the best circuit designer in Bell Labs and found Pedersen in Allentown, Pennsylvania, who devoted his design group to the silicon design problem. They both received the IEEE Lieberman Award for this achievement.

The first chip was called the DSP. It was a big hit at Bell Labs and found its way into many products from electronic switching systems to digital answering machines. The biggest application in volume was as a touch tone decoder. Echo cancellors were also a big use for DSPs. It was almostly immediately replaced by the DSP-2. This one was similar to the DSP, but the silicon was much improved. The original chip became know as DSP-1. DSP-1 was partially designed with standard cell logic, a VLSI design technique that uses predesigned circuits connected up with channel routers. The software that did this was LTX - another Bell Labs first. Standard cells make for fast designs, but slow circuts. DSP-2 was totally a custom design. Every circuit was designed by a human and the layout was too. Those days are long gone.

There is some dispute about who came up with the first DSP chip. At one time NEC claimed this honor. I think they indeed had the first chip that was sold commercially. In the late seventies, the Bells Labs chips were made by Western Electric which was the in-house manufacturer for AT&T. They weren't sold commercially, but many were given to researchers world-wide as was the custom in those days.

There are many technical papers in the literature. I'd check the IEEE archives. They aren't too interesting.

While I worked on the final design work on DSP-1, and DSP-2, most of my time was spent working on what was then called the DSP-3, or later DSP-32, a 32 bit floating point chip. We presented a paper at the 1985 IEEE ISSCC (International Solid State Circuits Conference). It won the coveted best presentation honor.

The DSP-32 was at the time the largest non-memory chip made, with about 450,000 transistors. These records get broken weekly, but it was a ground breaking development nonetheless. The circuits were very innovative and we got many patents from this project. The DSP-32 was used in many Bell Labs projects, but it was expensive and overkill for most telecommunications applications which are most efficiently served by a 16 bit fixed-point processor. This is what Texas Instruments hit the streets with in the mid-eighties, and they built a multi-billion dollar business out of it. By the time AT&T, then Lucent, then Agere Systems, got into the commercial DSP business, it was too little to late. Last I heard, Agere is totally out of the DSP business.

Looking back at my career at Bell Labs, and it's subsequent demise, I'm impressed by the brilliant people I had the chance to work with, and by the subsequent rise of the stupid ones. I think you'll find that in a lifeboat exercise, the brilliant ones are fed to the sharks.

[DrJustice]

Thanks for the lore!

I have edited my posting to correctly say "One of the Fathers". I do get carried away some times. Of course there was a team effort and Jim Bodie and R. A. Pedersen are hereby duly mentioned. There is also W. Patrick Hays who worked on the DSP-32 design. Interestingly, Lucent still offers descendants of the DSP32 chip.

I suppose that most readers are aware of how DSPs has penetrated our digital society - which is quite natural as DSPs are concerned with processing real (and unreal!) world phenomena in real time, IMHO the most interesting form of computing. They are everywhere, not least in music where they're now practically driving all things, except pure analogue. It's hard to overstate the importance of them.

DJ
--

[mosc]

Yes, Pat Hayes was one of the architects on the DSP-32 project. My guess is that if you can still get DSP-32's it is from inventory. I don't think they are made any more.

[jksuperstar]

[DrJustice]

I agree heartily. Although I understand that people in the audio industry hang on to the 56k for familiarity reasons, there are now much more suited and high powered offerings. My favourite being the SHARC despite its (only) two buses. I wonder why Motorola haven't at least made a "72k" series, that's 32bit wide instead of 24bit - wouldn't that be dead simple?

I suspect that a mistake on ADs part is the very high price of their tools. If they started sell them at a resonable price (or just give them away) there would possiblybe a quicker shift away from the 'old' 24 bit devices. Ah, and there's TI, not overly popular in audio - but the PPG Realizer used a heap of floating point TI dsps AFAIK.

DJ
--

[mosc]

Sounds like discussions in our old DSP marketing group. It turns out you make more money selling chips to big volume customers. So rather than focus on cheap development tools, or good general purpose architectures, you end up custom designing special chips for big applications, like cell phones, telephone answering machines, automobiles, etc. The marketing people would always say one big customer is worth 1,000,000 small ones.

I used to counter that big customers grow out of small customers. Guys that hack together a hobby project on your DSP go to work for big companies.

Anyway, the ways of IC architecture are strange and mysterious, and stupidity is everywhere. You are right about FPGA's with DSP cores. That would be great for small runners, but when you get the big volumes, the costs will not be attractive.

Hardware doesn't excite me too much anymore. The action is in software. I got into software developing CAD tools for our DSP design department. Those were the days of silicon compilers and I was in the thick of it. The game was to write a program does a DSP function and have a compiler lay out a chip. We had great success and wrote a lot of IEEE papers. Then the lifeboat exercises started. The circuit generator work was cancelled because the time frame was too long. The focus was shifted to chips for cell phones. Big customers are great, but when you loose their business for one reason or another, then you are out of business. Now Agere makes no DSPs, and no CAD tools. If they would have kept the research going they would have something to sell now. Like much of the American high tech industry, its the Titanic all over again.

The people who make all the bad decisions usually leave to go to better jobs in other companies before the fruits of their actions are understood. They don't even learn from their mistakes. New even stupider people are hired to fix the problems. Mediocrity is contagious.

[DrJustice]

[mosc]

[jksuperstar]

You'd be happy to know there is still much work, especially in FPGAs, in silicon compilation. There are now tools that work with Java...which is far better in describing hardware than "C" is, since it's a multi-threaded language (ie- simultaneous things). Still, Verilog and VHDL are the workhorse languages, simply because the tools for other languages do not compare. Once *someone* develops the tools, the rest will follow, and this is more true now, given the system-on-chip and "softcore" processors, meaning it is very easy to blur the line of what goes in software, and what in hardware, without dealing with inter-chip boundaries & timing.

As for AD. GNU tools are growing fast for the blackfin, since Linux was ported to that platform. I bought a devel kit for $99US, which included both board & software, has a executable size limit, but no functional limits.

[mosc]

I figure the work wouldn't stop and I'm glad to hear about new developments. I'm working own web site technologies now and have more than enough to keep up with in this field. There is immediate gratification that I like.