henry@utzoo.UUCP (Henry Spencer) (12/10/86)
Some time ago I alluded to a paper I'd seen which discussed what a good
basic computer technology should look like. I promised to dig it up and
post a precise reference. Finally, here it is: "What Makes a Good Computer
Device?", Robert W. Keyes, Science, vol 230, p 138, 11 Oct 1985. Since a
lot of folks expressed interest, herewith a summary of Keyes' comments.
(Keyes is at IBM Thomas J. Watson Research Center.)
A good computer device must be amenable to mass fabrication, i.e. putting
thousands of devices on a single chip. This appears to be essential to
building complex systems of reasonable (by modern standards) cost and size.
A subtle implication of this is that the technology must be tolerant of
modest variations in device characteristics, because mass fabrication does
not lend itself to great precision and reproducibility in the characteristics
of individual devices. Things like aging and temperature variations also
alter device characteristics, again requiring tolerance.
Note especially that novel technologies which impose restrictions on circuit
elements often propose to use "threshold logic", in which a three-input AND
(for example) is implemented by adding the inputs and comparing the result
against a threshold. Threshold logic puts great demands on the accuracy
of components and signal levels and so is ill-suited to mass fabrication.
All circuits degrade signals. The essence of digital systems is that small
degradations don't matter, because digital circuits consider a 0.99 signal
equivalent to a 1 and a 0.01 signal equivalent to a 0. Buildup of small
degradations will eventually become unacceptable, however, so it is vital
to re-standardize signals frequently. The basic operation required to do
this is amplification. And here's where the mass-fabrication requirement
bites hard: to get reliable re-standardization in the presence of variation
in device characteristics, we need a high-gain amplifier device so that minor
variations in gain won't matter.
Other important characteristics, some of them seemingly trivial or obvious
but all of them important, are: good isolation of input from output (i.e.
no built-in feedback), amplification of both charge and voltage, good
fan-out (ability to drive many inputs from one output), adequate fan-in
(ability to build gates with several inputs), ability to switch state in
both directions so that no separate "reset" operation is needed, and the
availability of the inversion operation which is vital for logic.
Things like high speed (the usual driving factor for unorthodox devices),
small size (for low speed-of-light lag), and low power consumption are
also obvious necessities.
Silicon transistors do all these things quite well. Moreover, since they
are the constant subject of intense development efforts, they steadily
get better.
Germanium transistors were the first transistors, and are faster than
silicon. Being transistors, they share most of the good properties of
silicon transistors. There have been periodic attempts to revive them.
They have two severe practical problems. One is that simply oxidizing
the surface of silicon forms a tough, stable film of silicon dioxide,
which is a superb insulator with excellent dielectric properties and is
also a good protective coating for finished chips; germanium oxide has none
of these virtues, and no adequate substitute is known. Second, the low
energy gap of germanium tends to turn it from a semiconductor into a
conductor when it gets hot, which would mean that dense germanium ICs would
need much better cooling than silicon.
Gallium arsenide shares germanium's problem of the lack of an equivalent
for silicon dioxide, and as a result GaAs transistors are built in rather
different ways from silicon transistors. They work all right, though, and
germanium's conductivity problem is absent. However, there are serious
materials problems. The threshold voltage, a crucial parameter, is quite
sensitive to both the composition and thickness of a specific layer, which
requires that both be controlled carefully. It is difficult to make good
low-resistance GaAs, which makes contacts in particular difficult. GaAs
is not as stable as silicon, which restricts fabrication methods. Finally,
there are saturation phenomena that reduce GaAs's speed advantage as the
individual devices are miniaturized. GaAs already has a secure niche in
specialized applications like RF amplifiers, but it's not clear that it is
a viable competitor to silicon in general.
Tunnel diodes switch very rapidly because of fundamentally small size.
Unfortunately, tunnel-diode logic uses threshold logic, a particular wart
because tunnel-diode characteristics vary a lot. Inputs are poorly isolated
from outputs. Gain is poor. And switching is unidirectional, requiring
a separate reset system.
Josephson junctions and related devices are fast because of very low voltages
and hence very small stored charges to slow down switching. But their gain
is inherently low, and their characteristics are sensitive to a number of
things, so they need precise fabrication. "In fact, a refined fabrication
technology is needed to make devices consistently that can be switched
at all." Equally severe are their cooling problems: while they generate
much less heat than silicon, effective heat removal is very difficult at
liquid-helium temperatures. If the fabrication problems can be solved --
IBM didn't think so -- the cooling problems limit their speed advantage
so severely that it doesn't seem worthwhile.
Optical logic was first proposed over twenty years ago, and has made little
progress since. Existing devices have poor gain, require precise fabrication
to work at all, tend to have rather variable output characteristics when they
do work, have low fan-out, don't isolate input and output well, tend to use
threshold logic, and usually show only unidirectional switching. "It is not
surprising that few attempts to construct even something as simple as a ring
oscillator have been made..."
--
Henry Spencer @ U of Toronto Zoology
{allegra,ihnp4,decvax,pyramid}!utzoo!henrymangler@cit-vax.Caltech.Edu (System Mangler) (12/12/86)
In article <7396@utzoo.UUCP>, henry@utzoo.UUCP (Henry Spencer) writes: > we need a high-gain amplifier device so that minor > variations in gain won't matter. Even in fairly well-behaved technologies like MOS, the gain is often not very high. In PMOS, which used enhancement-mode pullups, a gain of -3 was considered good. 4-transistor dynamic RAMs amplify on the same principle and have a tough time getting a gain of -1.5. Even CMOS may have gains as poor as -4 due to short-channel effects; it gets worse at high supply voltages and small feature sizes. Some forms of GaAs logic allow the pulldown transistor to conduct quite a bit when it is nominally off (due to limitations on where the threshold can be set), and these too have trouble getting much gain. Don Speck speck@vlsi.caltech.edu {seismo,rutgers,ames}!cit-vax!speck