Bernard Cole
URL: http://www.embedded.com/showArticle.jhtml;jsessionid=GCSOE4PC1YQDGQSNDBCCKHSCJUMEYJVN?articleID=13100886
Semiconductor companies will be
conferring over the next few months, as part of the International Technology
Roadmap for Semiconductors (ITRC) to develop a process technology roadmap for
future development in connected computing systems.
I hope they remember their past and
take a second look at some of the ideas and concepts that were once considered
but then thrown into the trash basket on the way to gigahertz clock rate,
multimillion transistor SoCs. They weren't bad ideas, just ideas that didn't fit
in with the needs of the marketplace at the time. But the times - and the
market dynamics - have changed.
One of those discarded ideas was to
move beyond binary logic to circuits based on multi-valued logic in which the
information density and processing efficiency of a circuit could theoretically
be increased substantially without any further expensive
"improvements" to the underlying fabrication technology.
Among the major alterations likely
to be approved are the addition of technologies for wireless communications,
including silicon-germanium, gallium arsenide and indium phosphide, as ways in
which wireless devices can be pushed to clock rates approaching 100 GHz. These
changes are important to the future of wireless technology if it is to achieve
data transmission bandwidths equivalent to or exceeding those on the wired
Internet.I find the references to
silicon-germanium one of the most interesting and tantalizing aspects of this
roadmap, considering how compatible it is with the use of multi-valued logic. SiGe's,GaAs, GaAsP, InP and other exotic combinations, transistors built with it are
heterojunction devices which are inherently capable of producing multiple
threshold levels.
Theoretically, SiGe could be used
to build devices that move beyond simple 0/1, on-off based binary logic. Such
structures can reliably generate multiple signal levels that are easily
discriminated. They could be used to build 3-base, 4-base, and higher logic
functions, effectively increasing a device's information density without
further shrinking the transistor structure. This option is something that
should at least be considered as we move ultimately into the sub-nanometer
range, where we are already facing problems relating to the cost of the fabrication
equipment.
Way back when the industry was
moving to two-to-four-micron geometries, most of the mainstream semiconductor
companies started investigating logic circuitry because
they felt that the shift in manufacturing equipment required to move to smaller
geometries would be too expensive.
Those who were building mainstream 16-bit
microcontrollers and microprocessors were especially interested after doing a
few quick calculations. For example, according to my calculations a 16-bit
microcomputer with on-board memory has access to no more than 216
bits memory (about 65k bits), while that same microcomputer with memory based
on ternary logic would have direct access to 316 or 43 Mbits of
memory. But
there were stumbling blocks in the way. For one thing, because they were using
homogenous silicon structures such as silicon, they had to come up with all
sorts of circuit tricks to express multi-valued logic using binary structures.
But work-arounds were found, and Intel, Fairchild, National Semiconductor,
Signetics (now Philips), and Motorola all had products on the market that had
ternary or quaternary logic hidden inside.
At
about the same time, the industry was looking for higher speed alternatives to
silicon transistors to boost clock rates. As they looked at various
combinations of gallium arsenide and other compounds, they realized that what
gave such non-silicon transistors their speed was their heterojunction nature.
It was then that early work began on trying to create some silicon-friendly
hetero-junction structure that could achieve comparable performance.
But
researchers at IBM, Motorola, TI and several universities also noticed that
heterojunction devices, silicon-based or not, had another interesting feature
-- they were inherently multi-threshold, able to discriminate and generate
multiple signal levels. This ability overcame several problems with earlier
attempts to use binary gate structures.
Ðreviously, to get around the inability to reliably generate and detect
multiple thresholds in binary silicon gates, it was necessary to come up with
silicon-greedy logic structures that could express multi-valued logic. But the
designs were problematic because the state of the industry at the time was good
only at discriminating two logic levels and it would have required a lot of
work to discriminate three or four.
Now,not only do we have SiGe transistor structures that are inherently friendly to
multi-valued logic, the industry has become quite good at generating and
discriminating multiple voltage and current thresholds. But
there is a third problem, not in the silicon itself, but in engineers'
willingness to move beyond the binary logic way of thinking that has become
second nature to them. The way that Intel and other companies got beyond that
problem in the past was to keep the engineer away from the ternary and
quaternary logic. That was very costly in silicon area at the time.
Now that extra silicon need not be so costly. But
will engineers be willing to move away from the safe world of binary logic,
even if the multi-valued logic is well hidden? Despite the fact that all of the
theoretical work on multi-valued logic that is available, we may be faced with
a situation similar to that of the missionary in an apocryphal story I heard in
an undergrad class in anthropology.
It
seems that after a year or so of trying to teach natives deep in the jungle how
to count using the decimal system, the missionary was meeting with total
failure. His pupils either did not get it or did not want to get it. When he
asked an anthropologist in the village why, the response was that this
particular tribe had a numbering system that consisted of zero, one and many.
The tribe members in their ordinary life up to then had no reason to consider a
numbering system that offered them more choices. They were not aware of all the
complications of modern life that would require a more sophisticated numbering
system.
Unlike
these mythical aborigines, I think that the economics of semiconductor
manufacturing now is forcing us to move beyond zero and one. We are already
considering other old ideas previously thrown into the wastebasket of history.
Shouldn't we take another look at multi-valued logic?
Binary logic is like driving through
Manhattan and only to be able to drive
straight and make right turns. Ternary logic is being able to drive straight
and turn left and right. Not only can you get somewhere potentially faster, in
a one-way grid, you now can reach places you couldn't reach before.
We will have to work with 729
commutative functions in ternary logic as opposed to 8 in binary logic. So more
is probably not always advantageous.
One of the handicaps in ternary and
multi-value logic is that mostly arithmetical examples are used to demonstrate
the benefits. Though of course of great importance, many engineers will not
design arithmetical circuits.
Synchronous methodologies are simply
a way to break down real world asynchronous problems into smaller and easily
digestible bits, to be given out to the many less talented engineers to solve,
who required tools to assist them. Or rather for the general engineering
community to understand and use easily.
Asynchronous design probably require
a very different mindset which few people can master. Synchronous will never be fully optimised. Asynchronous design will never be
easily mastered by most engineers, which lead me to predict that the future is
of a mixed-hybrid one, especially for large design. A fully asynchronous
design, would probably be more possible for small-scale designs which on the
other don't quite require the benefits of such design, which still must work
with the synchronous-dominated digital world.