Professional Documents
Culture Documents
(0 - 0) (CV) The - First - Microprocessor - An - Interview - With - Marcian
(0 - 0) (CV) The - First - Microprocessor - An - Interview - With - Marcian
net/publication/3321686
CITATIONS READS
0 632
2 authors, including:
George S. Moschytz
ETH Zurich
321 PUBLICATIONS 4,323 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by George S. Moschytz on 18 June 2014.
SPM: In this issue, creative thinking is the Dr. Hoff: My first love was chemistry. high school, I studied both electronics
main thread of our discussion with Dr. Ted My parents gave me a chemistry set as a and chemistry, and our high school,
Marcian Hoff, Jr. Welcome! Informally, gift when I was about nine. My father’s although small (about 35 in the graduat-
how would you define a creative person? brother, John, who is only 12 years older ing class of Churchville, New York), had
Dr. Hoff: I believe all humans are than me, became a chemical engineer excellent science teachers who offered an
inherently creative, but some are dis- after World War II and gave me many of after-school science club. While I contin-
couraged from believing in themselves, his chemistry textbooks. I loved the idea ued to study chemistry, my interest in
while others have more opportunity for of finding out how things worked, and electronics grew, and I mail-ordered a
their creations to be appreciated. chemistry seemed like magic. When I cathode-ray tube from a New York City
Perhaps the best definition would be “a was about 11, my uncle John gave me a surplus store. I used that tube to build a
person who is curious about how things subscription to Popular Science fairly simple oscilloscope.
work, and then takes action to make Magazine as a Christmas gift. When I
things work better.” saw an ad for a free Allied Radio catalog, I SPM: Soon your interest in technology
mailed a request for it thinking that it went beyond a simple hobby. What hap-
SPM: You were an imaginative kid: might be fun to learn how a radio pened next?
chemistry and electronics were your worked. The following Christmas, my Dr. Hoff: When it came time to
playground from early years. Would parents gave me a short-wave radio kit attend college, based on job prospects,
you tell us a bit more about that time? ordered from that catalog. Throughout my uncle recommended either electrical
or chemical engineering. I chose electri-
cal engineering. The summer after I
EDITORS’ INTRODUCTION graduated high school, my father got me
Our guest in this issue is best known as the inventor of the first microprocessor (the an interview in an electronics lab at the
Intel 4004) and a coauthor of the LMS adaptive algorithm. Marcian (Ted) Hoff, Jr. was company where he worked: General
born on 28 October 1937 in Rochester, New York. He received the B.S. degree in 1958 Railway Signal Company (GRS), in
from the Rensselaer Polytechnic Institute, Troy, New York, and the M.S. and Ph.D. Rochester, New York. They gave me a
degrees from Stanford University in 1959 and 1962, respectively. He was a research summer job as a lab technician, wiring
associate with Stanford University in 1962–1968, manager of applications research in
up test circuits. (I was told that having
1968–1980 and Intel Fellow in 1980–1983 at Intel Corporation, and vice-president of
built an oscilloscope really helped me get
corporate research at Atari in 1983–1984. Dr. Hoff also worked as an independent con-
the job.) One of my first projects at GRS
sultant in 1984–1990. He is currently chief technologist with the consulting company
FTI/Teklicon, San Jose. In addition to focusing on the µP and the LMS, Dr. Hoff’s work involved the design of an audio fre-
has spanned various aspects of digital signal processing that resulted in the first com- quency track circuit. The basic concepts
mercial monolithic telephone CODEC, the first commercially available switched-capaci- of the circuit had already been developed
tor filter, an early signal processing chip (the Intel 2920), and speech recognition by the engineers I was working for.
hardware. He is the 1997 recipient of the prestigious Kyoto Prize and the 1980 IEEE However, in the course of building a pro-
Cledo Brunetti Award. Dr. Hoff was inducted into the National Inventors Hall of Fame totype, I made some suggestions for
in 1996. He remembers that the latter event was “quite a performance, with a song improving the reliability and fail-safety of
and dance number including dancing chips.” the system. When it came time to patent
With his collaborators, Dr. Hoff shares a passionate interest in technology. With the
the system, the engineers insisted that I
DSP History column editors, on behalf of IEEE Signal Processing Magazine (SPM), and
be included as a named inventor. When I
our readers, he shares the 35th anniversary of the microprocessor, his thoughts on cre-
returned to GRS the next summer, it was
ativity, and his work-as-a-hobby philosophy.
—Adriana Dumitras and George Moschytz suggested that I team up with an old-
“DSP History” Column Editors time lightning protection engineer in an
adrianad@ieee.org effort to improve the audio track circuit.
moschytz@isi.ee.ethz.ch Frank Reese had left school after the
eighth grade, was self-taught, and at the
age of 60 had some 65 patents. He was a was connected to the rails. Trains at a Processing Magazine) together with
truly remarkable character and had a further distance would not be detected Prof. Bernard Widrow. What can you
wonderful sense of humor. (Among other because of the inductance of the rails. It tell us about the creative process that
stories, he told the story of having an occurred to me that we might get a bet- led to the algorithm?
idea at a lunch, writing up a patent dis- ter response characteristic by making a Dr. Hoff: Prof. Widrow described to
closure on a paper napkin, having it series tuned resonant circuit using that me some previous work he had done in
signed and witnessed, and using the inductance. The numbers indicated that pattern recognition: a process for reduc-
bonus from the resulting patent to buy a a reasonable capacitance value in series ing the overall error in recognizing a
car for cash, which car his son crashed; would work, and sure enough, when set of patterns by presenting the pat-
he then realized that, having paid cash tested on real rails, the response charac- terns in rapid succession and making
for the car, he had neglected to get insur- teristic was closer to the ideal. parameter adjustments to reduce an
ance!) Since systems used on the rail- Another such moment was later on at aggregate measure of error. Again, in an
roads are subject to some pretty violent Intel, when I wondered about how well effort to understand what was happen-
abuse, such as being hit by lightning components would match in our silicon- ing in such a system, I tried to analyze
bolts, Frank had a test setup that was gate metal-oxide-semiconductor (MOS) the behavior mathematically. The
used to simulate lightning bolts. That process. I knew that bipolar transistors resulting equations seemed to indicate
summer I came up with a technique that printed next to each other matched pret- that similar error reduction could be
allowed the audio track circuit to absorb ty well, but I didn’t know about our MOS obtained by making parameter adjust-
large amounts of energy without dam- components. I had figured I could make ments one pattern at a time, rather than
age, which resulted in my second patent. a digital-to-analog (D/A) converter by require the patterns to be applied as a
using a string of resistors to divide a ref- group. We then built a model that
SPM: Starting with your Stanford years erence voltage and use MOS transistors allowed us to test the algorithm, and it
as a Ph.D. candidate and then as a post- as analog switches to connect to taps on worked. Despite being quite simple, the
doc, your work produced results that the resistor string. However, I needed to algorithm seems to work very well in
became well known. If you were to look know how good a string of resistors I many applications. In addition, the con-
back, which were the “aha” moments could make, so I asked my boss, who was cept of reducing mean-squared error is
(aka the “Eureka!” moments) in your an expert on MOS, how well resistors relatively easy to understand and seems
professional life to date? would match. When he told me “not very to be quite effective for improving sys-
Dr. Hoff: I believe most of my “aha” well,” I complained, saying that I needed tem performance.
moments have come from wondering a quantitative measure, a number. Then
about how something worked and then he admitted that he didn’t know the SPM: After receiving your Ph.D. degree
seeing if it could be made to work better. number. I decided to try to find out for at Stanford in 1962, you continued to
For instance, one such moment was myself. I located some wafers that con- work there. How did your transition to
related to the audio track circuit. One of sisted of nothing but test devices. Intel happen?
its goals was the following: should Unfortunately each test device had only Dr. Hoff: I stayed on as a research
the circuit fail, it should not indicate the one of each type of resistor, so I could associate at Stanford working with Prof.
presence of a train. This goal was the not test adjacent devices. However, I fig- Widrow. Much of our work was supported
opposite of most track circuits. It ured it would be interesting to see how by government contracts. Another faculty
occurred to me that if the receiver had values varied over the wafer. I plotted the member working with us was Prof. Jim
an amplifier included, then that amplifier data, and there was quite a variation over Angell, who also consulted for Fairchild.
might break into oscillation and give a the wafer. However, I noticed that the He apparently gave my name to Bob
false train-presence indication. The solu- variation wasn’t random but rather a Noyce about the time when Intel was
tion seemed to be having a passive nice smooth gradient over the wafer. A being formed. I had met Bob once before
receiver. That required a sensitive relay little mathematical analysis indicated when we demonstrated some speech
in the detector and sufficient audio that had two identical resistors been recognition results. One day, in the sum-
power into the rails to be able to operate printed side by side, they would match to mer of 1968, I got a phone call from Dr.
that relay when detected. GRS had relays a fraction of 1%. Using this information Noyce, who asked if I might be interested
that could be operated with one-tenth of about matching and gradients, I was able in a new company he was starting. I
a watt, and we were able to fairly easily to design a D/A converter that was small think I was ready to try something differ-
generate about a watt of audio power to and accurate. ent than government sponsored
apply to the rails. The numbers looked research, so I replied affirmatively. I was
good, so a passive receiver proved feasi- SPM: In 1960, you created and imple- interviewed at Bob Noyce’s home. Talk
ble. Another goal of the audio track cir- mented the least mean square (LMS) about luck; but I had been interested in
cuit was that it should only detect trains algorithm (presented in the January developments in integrated circuits
that were very close to the place where it 2005 issue of the IEEE Signal while at Stanford and had conversations
device for a wide variety of applications nal processing and analog-to-digital our group was the first to offer a com-
and urged our marketing staff to renego- (A/D) and D/A conversion. mercially available single-chip CODEC.
tiate the contract. When Busicom asked The best process for analog work at For every CODEC, two filters are needed:
for some price concessions, our market- Intel seemed to be the one we used for one to band-limit the analog signal
ing staff succeeded in getting the rights EPROM memory, because it could take before it is digitized, and another one to
to sell to others. However, even then, relatively large voltages for program- eliminate digital noise from the analog
there were major concerns within Intel. ming yet operate with normal logic lev- signal derived from a digital input. I
Selling computers might be seen as com- els. When we started performing believe our group was also the first to
peting with potential memory cus- reliability tests on some of our circuits, produce a commercially available
tomers, and support requirements for we found some unexplained failures, switched capacitor filter to perform those
computers were perceived as formidable. even when operating at well below the switching functions.
Finally, a decision to offer the devices programming voltages. I finally got some We also recognized that there were
was made, and they were officially help from Intel’s process experts, who other functions performed in telephone
announced for sale in November of 1971. used scanning electron microscopy to systems, such as decoding the touch-
We suggested many applications such as find out what was going wrong. A simple tone signals used in dialing and generat-
process control, elevator control, high- process tweak eliminated the problem. ing a dial tone. We designed an
way signal control, computer peripheral So I consider the overall challenge to EPROM-programmed DSP chip that
device control, cash registers, point of find a way to use Intel’s capabilities for included A/D and D/A conversion on the
sale terminals, and medical electronics. analog applications to have been the same chip in the hopes of performing
Our customers soon found many more. most difficult. (Overall, the higher density many of those functions with one chip
of the MOS processes over bipolar design. Our hope was to offer the analog
SPM: Has the 4004 microprocessor been processes gave us many advantages, once designer the type of flexibility that the
your most challenging technical assign- problems on noise and amplifier design microprocessor offered to the digital
ment so far in terms of problem analysis had been solved. MOS processes were designer. However, our design was not
(or what is sometimes called “making the considered noisier, but the nature of the very successful. I believe there were a
strange familiar” in creative thinking)? noise was such that there were ways to number of reasons. We probably made
Dr. Hoff: Once the microprocessor compensate.) too many compromises in combining
was introduced, I was asked by Dr. Noyce The other difficult aspect of the proj- the conversions with the digital proces-
to see if there were any opportunities for ect was that the telephone industry at sor, resulting in a digital processor that
Intel in the telephone industry. That the time was rather closed. We had sev- was less than spectacular in its perform-
seemed almost impossible for two rea- eral contacts at telephone equipment ance. We also found that there seemed to
sons. First, a lot of signal levels in the manufacturing companies, and they be relatively few analog engineers at that
telephone industry were incompatible made it quite clear to us that they time (around 1979) that were really com-
with Intel’s processes. Second, the tele- resented our invading their territory. We fortable with DSP. Lastly, switched capac-
phone industry at the time was pretty were told by one of these companies that itor techniques allowed relatively
much a regulated monopoly. Let me they would be the ones to design prod- inexpensive custom designs for many of
explain each. Many of the signals were ucts, and once those designs were speci- the functions we had expected our digital
analog, and it seemed almost a given that fied, we might try to qualify as a foundry. chip to perform.
anything involving analog signal pro-
cessing required a bipolar process. SPM: Between 1975–1980 you headed a SPM: Later, from 1980 to 1983, you
However, Intel’s bipolar process was not group that produced the first commercial headed a group that developed speech
really well suited to analog designs. I monolithic telephone (CODEC), the first recognition hardware at Intel. What was
hoped to find a way to use one of Intel’s commercially available switched capaci- the main idea of this work and where
MOS processes for our analog work. tor filter, and an early signal processing was it applied?
Then my group at Intel got very fortu- chip known as the Intel 2920. What can Dr. Hoff: We built a speech recogni-
nate in that we persuaded Prof. Paul you tell us about these products? tion system that used our digital signal
Gray of the University of California at Dr. Hoff: After studying the tele- processor for the analog input and con-
Berkeley to come and consult for us. He phone industry, it seemed to me that our ventional microprocessors for the subse-
was also very interested in analog appli- best hope was to offer products to pro- quent recognition. It included a type of
cations of MOS processes, and he had mote more use of digital technology in state machine to reduce the amount of
done some outstanding pioneering work the industry. Our first basic design was vocabulary search needed at each point
toward that end at Berkeley. We had a for a telephone CODEC, a device to con- in the recognition process. Some units
number of his students join our group at vert telephone quality analog signals to had speech synthesizers, so they could
Intel and solved many of the problems of the standard telephone digital formats in carry on a type of conversation with the
applying MOS technology to analog sig- use at the time, and vice versa. I believe user. My understanding is that one of the
dispersion and PMD [9]. A feedback could be represented with a digital filter REFERENCES
[1] S. Cao, J. Chen, J.N. Damask, C.R. Doerr, L.
scheme (that we do not display in Figure model as described in this article. Use of Guiziou, G. Harvey, Y. Hibino, H. Li, S. Suzuki, K.-Y.
4) could be used to track system changes well-developed signal processing tech- Wu, and P. Xie, “Interleaver technology: Comparisons
and applications requirements,” J. Lightwave
to optimize the equalizer performance. niques and algorithms to design these Technol., vol. 22, no. 1, pp. 281–289, Jan. 2004.
Although a straightforward and ele- optical devices is a wise use of existing [2] C.K. Madsen and J.H. Zhao, Optical Filter Design
and Analysis: A Signal Processing Approach. New
gant approach, the electronic equalizer technology. As demonstrated in this arti- York: Wiley, 1999.
has an inherent shortcoming. Some cle, signal processing could play an [3] C.-H. Cheng, L. Lin, T. Towe, T.J. Xia, C. Wong,
information about the optical signal such important role in the development of K.-Y. Wu, and J.-C. Chiao, “Spectral slicers without
cascaded dispersion penalty,” in Proc. 28th European
as polarization and phase is lost during advanced optical communication sys- Conf. Optical Communication, Sept. 2002.
the O/E conversion. As a result, the elec- tems. However, as demonstrated in the [4] B.B. Dingel and T. Aruga, “Properties of a novel
tronic equalizer illustrated in Figure 4 case of an electronic equalizer, some noncascaded type, easy-to-design, ripple-free optical
bandpass filter,” J. Lightwave Technol., vol. 17, no. 8,
has a limitation on its performance. A optical system characteristics may pp. 1461–1469, Aug. 1999.
special modulation scheme could be require special attention if signal pro- [5] P. Yeh, Optical Waves in Layered Media, New
used to solve this problem to some cessing techniques are to be applied suc- York: Wiley, 1991.
extent. It has been demonstrated that cessfully. Therefore, interdisciplinary [6] C. Chen, D.Z. Chen, D.L.P. Peterson, Jr., C.-H.
Cheng, S. Lin, H. Fang, T. Xia, K.-Y. Wu, and G.
when a single side band modulation cooperation between researchers in Zhou, “A newly developed dispersion compensation
module for ULH/metro applications,” in Proc. CIIT
scheme is applied in optical systems, the optics and signal processing will be cru- 2003, Nov. 2003, pp. 221–225.
performance of an electronic equalizer cial for optical communications to fully [7] C.K. Madsen and G. Lenz, “Optical all-pass filters
can be improved significantly [10]. benefit from signal processing. for phase response design with applications for
dispersion compensation,” IEEE Photon. Technol.
Lett., vol. 10, no. 7, pp. 994–996, July 1998.
CONCLUSION ACKNOWLEDGEMENTS [8] C.K. Madsen, J.A. Walker, J.E. Ford, K.W.
Optical communication plays a signifi- The author would like to thank Dr. Goossen, T.N. Nielsen, and G. Lenz, “A tunable dis-
persion compensating MEMS all-pass filter,” IEEE
cant and increasing role in our society. James Kiper of Miami University for his Photon. Technol. Lett., vol. 12, no. 6, pp. 651-653,
June 2000.
The public demand for higher network comments and help.
[9] H.F. Haunstein and R. Urbansky, “Application of
speed requires an optical backbone net- electronic equalization and error correction in light-
work with larger capacity. Accompanying AUTHOR wave systems,” in Proc. 30th European Conf. Optical
Communications (ECOC), Sept. 2004.
high transmission-rate optical commu- Chi-Hao Cheng is with the Department
[10] P.M. Watts, V. Mikhailov, M. Glick, P. Bayvel, and
nications system are severe technical of Electrical and Computer Engineering, R.I. Killey, “Single sideband optical signal generation
specifications for optical devices and sys- Miami University. He is a Senior Member and chromatic dispersion, compensation using digi-
tal filters,” Electron. Lett., vol. 40, no. 15, pp.
tems. Many popular optical devices of the IEEE. 958–960, 22 July 2004. [SP]
SPM: Let us wrap up by looking in the applications we will find. When we com- lation. In particular, fast and accurate
past with a thought for the future: your bine traditional DSP with other logical language translation would seem to
invention of the microprocessor was a processing such as data encryption and offer a huge potential for improving
major breakthrough. Where do you addition or elimination of redundancy, human communication and coopera-
consider that breakthroughs are needed we can expect to improve reliability and tion, and better machine understanding
now in DSP? security of all of our communication of language should help make comput-
Dr. Hoff: The more speed we can channels. I would also like to see more ers even more useful.
offer in both the DSP and the associated natural language processing, including
A/D and D/A conversions, the more recognition, understanding, and trans- SPM: Thank you. [SP]