Question 1: Tell us about yourself. What is your background, and what current projects are you working on?
I graduated from Caltech in 1973, with a B.S. in Physics. I did my Ph.D. work at the University of Colorado, under Herb Kroemer (2000 Nobel laureate in Physics). We developed the first microscopic theory for the electronic structure of semiconductor heterojunctions. I spent an additional year with Herb at the University of California at Santa Barbara, and then joined the Central Research Laboratories at Texas Instruments in 1977. At TI I initially worked on a number of different compound semiconductor devices: power GaAs MeSFETs, heterojunction bipolar transistors, and vertical MeSFETs.
In 1985 I joined the Nanoelectronics program at TI, which was
established by Bob Bate. The objective of the program was to develop
a follow-on technology to silicon MOS. Bob had plotted the intersection
of the Moore's Law trend line with the projected physical limit of
MOS transistors (then estimated to be about 0.2 micron gate length);
the intersection occurred in about 1998. The experimental work
was focussed on quantum-well and quantum-dot devices in semiconductor
heterostructure technologies, which were the only nano-scale
technologies that offered any possibility of signal gain at the
time (using resonant tunneling as a current control mechanism). I
concentrated on developing the theoretical technologies required to
design devices in which quantum effects like tunneling coexist in
about equal magnitude with more classical transport phenomena. This
work, (which after 1990 was joined by several colleagues) involved the
development of a series of transport computations that by about 1995
resulted in a comprehensive understanding of the detailed physics of
quantum-well electronic devices
(see http://www-hpc.jpl.nasa.gov/PEP/gekco/nemo).
The Nanoelectronics Branch was included in the sale of TI's government
electronics business to Raytheon in 1997. It had been supported by
federal research funding for a number of years, and was concentrating on
compound semiconductor technologies.
Pursuing beyond-the-roadmap R&D at TI during this period required us
to pay attention to a couple of points that I do not see reflected in
more recent nanoelectronics work. The first was that we could not ignore
the properties and capabilities of the incumbent technology (silicon MOS).
The second is that we had contact with workers who had actively
participated in the creation of solid-state electronics and computing
in the 1960s (some of whom were in the management chain!). Thus, we
could neither ignore the history of technology. We examined a great many
more ideas for post-silicon technologies than just resonant-tunneling
transistors. But in the end, all of the alternatives have turned out to
be rather minor variations on ideas which had already been tried by the
mid 1960s. The principles established then are still true: One can do
no more than a very few stages of logic without restoring the logic
levels. (Consider diode logic or magnetic-core logic.) Restoring logic
levels requires amplification (gain). The key to any successful digital
technology is gain.
Question 2: How much longer do you believe that Moore's law will continue? Do you believe that the predictions of the International Technology Roadmap
for Semiconductors (ITRS) are accurate?
First, I think we need to understand precisely what Moore's law is.
Scientists and engineers look at an exponential growth curve and imagine
that there is some first-order differential equation behind it. That
was true of Moore's law during the period when the cost of IC fabrication
was only very weakly dependent on the feature size, that is when
full-wafer contact lithography was the dominant technology (up to about
the mid 1980s). When it became necessary to switch to step-and-repeat
projection lithography, Moore's law ceased to be a natural consequence.
It continued because it had become the collective wisdom of the IC
industry. It was the ideal management tool, because it supplied a
product development strategy and schedule that required no further
inputs (except by the technologists charged with making it happen).
At present, I think Moore's law is most accurately viewed as a
business strategy. It will be abandoned when it produces losses
rather than profits. A necessary condition for continued scaling
is that the cost of production of an IC has to scale sublinealry with
the functionality. Many workers in "nanotechnology" research claim
that making things smaller will make them cheaper. This is simply not
true. Making things smaller is always more expensive. (Show me a
researcher who has cut his budget request because he is working on
smaller-dimensioned systems.) IC scaling has continued because the
costs increase at a slower rate than the net functionality. Another
necessary condition is an exponentially growing total market to absorb
the increasing development costs. (This last requirement is going to
cause a great deal of trouble for programs like NASA's nano-satellite
initiatives.)
I think that a useful historical analogy is the aircraft industry.
From, say, 1940 to 1965 the capabilities of both military and civilian
aircraft developed very rapidly, culminating in Mach 3 military craft
(SR-71 and XB-70) and Mach 2 civilian craft (Concorde and TU-144).
In the end, the benefits of supersonic flight have not been adequate
to justify the costs involved. The costs in this case are directly
traceable to the rapidly increasing energy input required for speeds
above about Mach 0.9. In the case of IC scaling, the question is which
element of the cost is going to diverge first? (Lithography is the most
likely candidate.) The analogy to the aircraft industry also predicts
that some semiconductor companies are going to overshoot the end of
Moore's law, and lose a great deal of money in the process. I think we
may already be seeing the first signs of desperation, in gamesmanship
involving processor clock frequencies, and in the size of Intel's
McKinley chip.
Having made the point that the real limits are economic, I can tell
you that what I am hearing now is that there are real problems making
transistors work at 50 nm gate length. I suspect that a significant
redesign of the MOSFET might squeeze one or two more generations at
this point, but that will be about it. The consensus among those
with whom I have been in contact is that exponential growth will be
essentially over by the end of this decade (to finally answer the
original question).
Question 3: The ITRS seems preoccupied with current silicon technology. Do
you believe that the electronics industry will ever abandon silicon in favor
of some other semiconducting material (such as Gallium Arsenide?)
I don't think the semiconductor industry will ever abandon silicon for
large-scale digital circuits, because such a switch would produce only a
one-time improvement of a factor of a few in speed (and perhaps some tens
of percent in power dissipation). It is not a scaling strategy for more
than one generation. On the other hand, I expect that when simple scaling
of monolithic ICs ceases to be a viable strategy, we will see much more
creative work on hybrid assembly processes. These would permit designers
to choose optimum semiconductor technologies for different subsystems
(millimeter-wave transmitters and receivers, for example) and place them
in a single package. This will result in an increased demand for compound
semiconductor components in the high-speed digital, rf, and
opto-electronic subsystems.
Now, one thing that will happen is that fabrication processes developed
for GaAs monolithic microwave ICs (MMICs) will be adopted (though never
acknowledged) by the mainstream silicon IC industry. I recall an
interaction some years ago when TI's leading silicon process development
engineers were briefed on the GaAs processes. They were quite amused
that electroplating was used to form interconnect metals. (I belive the
quote was something like: "That can never be a high-volume IC process.")
Well, how they are depositing copper in the much-vaunted state-of-the-art
interconnect process? Electroplating! Following this line of reasoning,
I am confident in predicting that the low-K inter-metal dielectric
problems will be solved by removing the dielectric completely (known as
"air bridge technology" in the GaAs business), and that about ten years
from now, a big innovation in silicon will be through-the-wafer metal
vias for low-inductance grounding.
Question 4: Recently there has been enormous interest in quantum computing.
Yet critics claim that quantum computing would only be useful in decryption.
Do you believe that quantum computing will ever become a mainstream
computing paradigm?
I am not at all convinced that quantum computing will even be useful
for decryption. One question we have to ask about any alternative
computational technology is: "What is the scaling strategy, and why
should we believe that it will work?" In the case of quantum
computation, any scaling strategy is manifestly self-defeating. Quantum
computation requires that a pure quantum state be maintained and
manipulated by time evolution under some suitably constructed Hamiltonian
operator, without any irreversible (or phase-breaking) interaction
occurring until the computation is completed and the results measured.
Now those irreversible interactions can be prevented in small systems
(like a few atoms) which are well isolated from their environment. When
we try to scale any such system up to a much larger size (say around a
thousand atoms), it will become a classical system, in which the
irreversible interactions cannot be avoided. Proponents of quantum
computing will point to superconductors and Bose-Einstein condensates
as macroscopic systems which are in a coherent quantum state. That state
is always a ground state, and ground states will not perform
computations.
There is another little-discussed aspect which is inherited from
quantum computing's intellectual progenitor, reversible computing.
Reversible computing proceeds from a mapping between boolean operations
and unitary (invertible) operators, and is proved universal in the sense
that any boolean function can be evaluated. However, no one examines
how efficient (in terms of hardware or execution time) this approach
really is. The answer is that, in comparison with modern digital
electronics, reversible computing is hopelessly inefficient, with the
required hardware scaling approximately exponentially as conventional
hardware is scaled linearly. The reason is that digital electronics
enables structures that are not strictly evaluators of boolean
expressions, but perform a higher level of functionality. The simplest
examples are regenerative digital circuits (flip-flops, counters,
registers, jam latches, etc.); other examples include three-state
circuits that enable the bus architecture, and memory itself. The
reversible computing paradigm effectively requires that each bit of
memory be replaced by the entire boolean logic tree that produced the
result that it would have stored.
Now, I think that there will be at least one benefit of quantum
computer research. That will be an experimental test and critical
reevaluation of the "projection axiom" of quantum theory. That is the
assumption that a measurement causes a collapse of the wavefunction onto
a pure state in the basis of the measured dynamical variable. This has
always been something of a "deus ex machina" in the theory, marking an
apparently unbridgeable divide between a quantum system and the classical
measuring apparatus (which, being made of atoms, really ought to be a
quantum system, too). It is the way that irreversible interactions are
included within elementary quantum theory. So far as I can tell, the
projection axiom is a necessary ingredient of proposed quantum
computation schemes, as it is how the qubits that have been corrupted by
an irreversible interaction can be identified. If, on the other hand,
interactions with the external environment are more subtle and continuous,
then quantum computation faces precisely the same limitations of signal-
to-noise ratio that classical analog computation did. (For, example,
suppose the quantum computer consists of nuclear spins in some
appropriately designed molecules. If the chemical environment caused
small shifts in the energy levels of the spin states, this would be
reflected in shifts in the frequency of oscillation of the wavefunction,
producing phase errors at the end of the computation.) I strongly
suspect that this scenario will be found to hold in the end.
Question 5: What about optoelectronics? There are companies that claim to
have created interconnects out of fiber optics. Could fiber optics ever be
placed on an integrated circuit? Will we ever see an all optical computer?
There are really three separate issues here: 1. Is there any benefit to
doing computations optically? 2. Is there any benefit to using optics to
move information within an IC? 3. Is there any benefit to using optics to
move information between ICs? First, let's look at the basic physics.
Photons are massless bosons. That means that they obey a linear
superposition principle, and in the first order of approximation do not
interact with each other. That makes them an ideal medium of
communication, because you do not want your message to be degraded
because it bounced off someone else's message. To get photons to interact
with each other, we have to employ nonlinearities in material systems,
which are typically weak, requiring large optical powers to achieve any
useful nonlinear effects. Computation, as we typically know it, is an
intrinsically nonlinear process. (Think of an AND gate: you can't get
its output as a linear combination of its inputs.) Now, massive fermions
interact nonlinearly in the lowest order, via the Pauli exclusion
principle if nothing else. Electrons are the lightest known massive,
charged fermions. It is no accident that our dominant computational
technology is called "electronics".
The work of Alan Huang at Bell Laboratories in 1990 pretty accurately
demonstrated the potential of optical computing. His system had a
capacity of 32 bits, switching at 1 MHz, and required lasers that I
suspect were near the kilowatt range to power it. With that kind of
competition, electronics wins, hands down.
The question of the usefulness of optical interconnect (across a
chip) or optical pinout (chip to chip) is more complex. The two
things that determine the usefulness of optical communication are
(i) photons interact only very weakly with non-metallic materials,
and (ii) the wavelengths of light that are useful for optical
communication are of the order of a micron. The result is that the
optical components are at least an order of magnitude larger than
the corresponding electronic components (lasers and detectors vs.
transistors, fibers vs. metal wires). If an IC designer chooses
to use interconnect wires and transmission lines of the same
dimension as a proposed optical interconnect, he will achieve at
least the same performance, at a much lower cost.
For pinout, the comparison is not quite so straightforward.
The transit time across a chip is, and will remain, a small fraction
of the clock period. The consequence of this is that one does not
need to worry very much about matching the impedance of the
transmission lines crossing a chip to the components driving them.
For chip-to-chip communications, we can depend less and less on
such an assumption. Now, there is a big problem in that the impedance
of the transistors used in the digital circuitry is at least in the
kilohm range, whereas physically realizable transmission lines have
impedances of a few tens of ohms. This necessitates devoting a
significant amount of chip area to the buffers required to drive
the off-chip lines. Thus, the potential of optical pinout is to
provide a better impedance-matching solution, and it must be judged on
this basis. One can calculate an effective impedance for the optical
connection by dividing the threshold voltage of the laser by its
threshold current, and discounting this by the efficiencies of the
electron-to-photon and photon-to-electron conversion processes. I
don't have the data to make this comparison, but the key is obviously
the development of low-threshold-current-density lasers that are
integrable on a Si chip.
Question 6: Have you done or seen any research on carbon nanotubes? They
have been proposed as interconnects, as memory devices, and even as
switching devices.
I have been following nanotube research mostly at the level of reading
the press releases. One thing is pretty clear: Sorting the components of
soot is not the route to a competitive data processing technology. Nor
are such "breakthroughs" as reinventing the concept of a fusible link to
blow out the undesired connections
(http://www.hp.com/hpinfo/newsroom/press/23jan02b.htm).
Self-assembly of macromolecules is not going to lead to technologies
that employ mechanisms (whether mechanical, electronic, or optical)
unless techniques are developed to accurately and discretionarily
place and orient the macromolecules. In the absence of such techniques,
carbon nanotubes have a bright future in structural materials, catalytic
systems, and other traditional applications of chemical technologies, not
in the applications dominated by electronics.
Computers must be designed, and the design has to be communicated
in some fashion to the physical system that will implement that design.
Now, it is perhaps possible that one could fabricate arrays of
nano-manipulators that could place and align nanotubes or other
self-assembled nanostructures. However, one must remember that
for electronic applications, the competing technology is lithography,
and the winner will be determined by the resulting cost of fabrication.
It will be very hard to beat the cost of lithography.
(Nano-manipulators may work at smaller scales than lithography, but
there will always be the question: Why make things smaller if
that turns out to be significantly more expensive?) On the other
hand, if one follow's Richard Feynman's vision of making small
machines that fabricate and assemble smaller machines, that means
that one will have developed micron and hundreds-of-micron sized
manipulators along the way to nanomanipulation. I suspect that
these latter systems will find more useful employment in assembling
hybrid microelectronic systems.
Also, if a suitable nano-manipulation technology were available, it
is far from clear that carbon nanotubes are what you want to use for
active devices. I think we could make useful amplifiers from atomic
clusters (nanocrystals) of ordinary semiconductor materials. From
what I can tell of nanotube chemistry, I would conclude that
nanoscale semiconductor heterostructures have many more degrees of
freedom (design options) than nanotubes. The
unique thing that I see in nanotubes is that, from the point of view
of electron transport, they impose no surface, such as a nanocrystal
would have. I think this is more significant than the usually cited
properties (like the relationship between the chirality of the
molecule and the electronic band structure). Surfaces, when their
electronic properties are uncontrolled, have been a problem in
semiconductor electronics since the days of the cat-whisker detector
on a galena crystal. On the other hand, interfaces with well
controlled properties are necessary components of all known active
devices. It remains to be seen what the tradeoffs between different
materials will be.
Question 7: Certain microbiologists are claiming that the future of
computing lies in cell-based media. Some are even claiming that DNA could be
used in computing. Is there any truth to these assertions?
If those computers required even a small fraction of the overhead that
ordinary biological systems require (for structural integrity, energy
conversion, material transport, information transfer, waste-product
disposal, etc.) then a present Intel processor has already won that
competition by at least a couple of orders of magnitude. From an
engineering point of view, the only virtue of biological systems is
that they are self-replicating. There may be, however, a limited role
for DNA chemistry in addressing certain types of problem. In those
problems where the complete solution requires one to examine all
possible permutations (NP-complete problems) the idea of encoding
those permutations in DNA and examining the results by chemical
processes is an attractive concept. This brings Avagadro's number
to bear on the solution. Systems which solve problems in this
fashion are going to look a lot more like biochemistry laboratories
than desktop computers, however.
Question 8: How much research have you done on molecular electronics?
Researchers in the field of molecular electronics claim that the rate of
progress in the field is enormous. Do you agree?
Again, I have been following molecular electronics mostly from a
distance, since Forrest Carter was promoting his version of it in the
mid-1980s. As to the progress, I think the field advanced from
complete speculation to "molecular conduction" with the demonstration
of single-molecule conduction by Mark Reed and Jim Tour. I would
reserve the term "electronics" for a technology in which a device with
signal gain has been demonstrated. I know of no such demonstration for
single or few molecule systems. Now, about a year ago, I heard an
informal report on the DARPA Molecular Electronics program. The
statement was made that a number of the investigators (chemists,
obviously) had pointed to vials of solutions and claimed that these
contained single molecule switches and/or transistors, but of course no
one knew how to hook them up to test them, a transistor having three
terminals and the state of the art of molecular connection allowing
only two contacts. (This situation simply cries out for a
discipline-chauvinist joke, but I will restrain myself.)
If we look a bit closer at Jim Tour's molecules I think we can see
some of the complexities of the problems of molecular electronics.
Jim puts thiol groups at the ends of his polymers, to cause the ends
to bind to gold electrodes and thus encourage the molecules to assemble
themselves across the gap between two electrodes. Now, this binding
is accomplished by a transfer of charge. When a similar interaction
between a metal and a semiconductor occurs we call it a "Schottky
barrier," and it is not a desirable type of contact. (What we are
looking for is an "ohmic contact.") This illustrates what I fear will
be a fundamental conflict in the design of molecular electronic
components, between structural and electrical requirements. In
semiconductor technology, we are accustomed to not having to worry
about the crystal holding itself together as the transistor goes
about its business. Fundamentally, this is due to the fact that there
are many more chemical bonds in the crystal than the number of
electrons we are moving around. In single-molecule electronics, (at
least with the relatively simple molecules that are currently in play)
this ceases to be the case. When we move electrons around, we are
fiddling with the bonds that hold the molecule together. If you think
of electrical conduction as electrons moving along the network of
chemical bonds, this suggests that one ought to look at larger
molecules, where there are many more paths through network of bonds
(as, for instance, in a carbon nanotube as opposed to a polymer).
But again, the question is: When we figure out how big a molecule
has to be to be an effective transistor, how much smaller is it than
a semiconductor transistor?
Question 9: Eric Drexler and Ralph Merkle have proposed computing designs
based on rod logic. They claim that these techniques could result in a
computer the size of a bacterium. Are you familiar with rod logic?
I actually own a rod computer! It is made of plastic and I acquired it
in 1965, when I was in junior high school. It is called a Digi-Comp I,
and was sold as an educational toy starting in about 1963. It is a
three-bit programmable state machine. The manuals for this device can
be found on the web at: http://fallout.tjs.org/galena/digicomp/ .
I recently relocated it during the archaeological excavation of my
parents' basement. The scaling limitations of this device are
pretty obvious: the size and complexity of this architecture goes
as the square of the number of bits, and the structural limitations
of the materials would enforce a limit of about five bits.
Concerning Drexler and Merkle's proposal, where to begin? Suppose we
accept the proposition that the molecules are perfectly rigid, that
there will be no loss of signal level that would require amplification
to achieve logic level restoration. (And we had better make such an
assumption, as the only effective amplification mechanisms of which I
am aware operate on fluids because they are intrinsically valves [apart
from stimulated emission in the laser]. For all our talk of solid-state
electronics, the solids are really only containers for the electron
gas.) Now, in a state-of-the-art IC, a bit is represented by perhaps
5000 electrons, or a total mass of about three atomic mass units.
In Drexler and Merkle's molecular rods, the molecular weight of a bit
would be at least of the order of 1000. In electronics, the mass of
the electron is not yet a limiting factor (in terms of frequency, its
limit is about three orders of magnitude higher than the current limits
due to dissipative effects) but the mass of the molecules will
certainly limit the speed of molecular rod computers. Why should we
believe that there is an advantage in switching over to much more
massive particles?
Then there will be the matter of cost. As I've already noted, there
is an inherent conflict between structural stability and computational
functionality. The beauty of semiconductor electronics is that these
two functions are addressed by completely different parts of the total
system. But with molecular rods, we obviously have to have molecules
forming the structural framework and molecules sliding back and forth
doing the computation. If we depend upon the structure to assemble
itself, the rods will almost certainly end up locked in place (or
otherwise, placed at random in response to the dictates of the 2nd law
of thermodynamics). Thus they will have to be manipulated into place.
It is here that the three-dimensional structure, far from being an
advantage, becomes a fatal weakness. Three-dimensional structures,
whether built block-by-block or layer-by-layer, necessarily require a
labor that is directly proportional to the number of components in at
least one of the linear dimensions. (Compare that to the nearly
constant fabrication cost of planar strucures in the early days of
ICs.) Now, one can still, in principle, craft a scaling strategy that
satisfies the prerequisite of Moore's law that cost scale sub linearly
with functionality. But the chance of actually implementing such a
strategy would require some extremely effective innovation in placing
and orienting large numbers of molecules in a parallel operation. And
then you still have the problem of starting from an orders-of-magnitude
higher cost than the mature IC process, and working your way down the
learning curve to an actually competitive technology. (Compare the
problems flat-panel displays have had in trying to replace the cathode
ray tube.)
Question 10: Writers such as Ray Kurzweil argue that the fields of computing
and communication are advancing exponentially, and that these exponential
advances will continue for the foreseeable future. Do you agree?
It is well-known to those involved in marketing ICs that the
leading edge of a product life-cycle curve resembles an exponential.
We are obviously looking at a life-cycle curve, but the question is
what are the time constants? The exponential growth in the number
of features on an IC is coming to an end in the next few years.
The opportunities for growth in communications bandwidth will continue
for some time after that. But, more to the point, we have been extremely
wasteful of the resources that the Moore's law growth of ICs has given
us (think of the Windows operating system). Thus, opportunities for
innovation in hardware and software design will persist for many years
to come. What we must hope for is that the business and investment
climate will not stifle those opportunities.
Question 11: Should the electronics industry devote more effort towards the
post-silicon future?
Well, I hope that by now I have dispelled the notion that the
short-sighted semiconductor companies are just sitting around ignoring
all these marvelous opportunities that the physics and chemistry
communities are providing. The truth of the matter is that there
has been a serious concern over these issues for nearly 20 years, that
the industry has had (at least until the past few years) a great
reservoir of technical wisdom, and that the lack of action reflects
the quality of the opportunities.
If there is a problem illuminated by the electronics industry's
failure to fully embrace "nanotechnology," the problem is the
collective failure of the research community to cultivate and
communicate technical wisdom (which would have severely dampened
the enthusiasm for nanotechnology as a replacement for electronics).
The last place one would go to find out how to design and fabricate
a complex system is the refereed scientific literature. Why? Because
the literature contains many more examples of the wrong way to do
something than the right way. This is a consequence of asking only
for novelty and correctness in our results. The wisdom that is
accumulated from years of experience in high-tech development is what
is required to evaluate competing ideas. It is not easily rendered in
the form of ink on paper.
Question 12: What are your plans for the future?
I am trying to address the problem of communicating technical wisdom.
The specific project is something I call a "HyperTextBook" on electron
device theory. This is a work that is viewable only on a computer, and
looks like the printed page, but it incorporates interactive, graphical
simulation programs. The objective is to incorporate a large element of
play (in the sense of unstructured interaction with the simulations)
into the learning of a complex technical subject. Also, I must stress
that these are really simulations, with the solution process performed
by the computer, rather than animations of analytic approximations. A
subsidiary objective is to demonstrate the limits of validity of those
analytic approximations.
It is my belief that with this medium I can demonstrate many aspects
of device behavior and design principles that are not demonstrable with
analytic mathematics, by inviting the reader/student to try it for
himself/herself. Such material of course does not appear in the
refereed technical literature.
This project requires that I redevelop all of my professional
simulation codes in Java, and to develop a number of new ones. Thus the
task will occupy me for at least the next few years.
