Robert W Keyes. American Scientist. Volume 97, Issue 2. Mar/Apr 2009.
On October 22, 1985, the “Science Times” section of the New York Times ran an article titled “Light, Not Electricity, Key to New Computers.” No progress in computing with light was announced for several years, but in 1990 a report of an optical computer put together by AT&T’s research laboratory inspired the headline “Is Optical Computing The Next Frontier Or Just a Nutty Idea?” in the January 30 Wall Street Journal. “Nutty” may be the wrong word; there are some reasons why researchers have pursued optical computers. Photons travel faster than electrons. But the AT&T effort could not measure up to well-established electrical methods and was the last gasp of serious commercial efforts into optical computing, although some theoretical, academic research continues. InfoWorld Media named the concept one of its “crackpot technologies that could shake up IT” in February 2008. Optical technology may eventually find some associated uses, such as for transferring huge data files, but it doesn’t look like it will ever be a viable option to replace the transistor in the computing process itself. Why? What is needed to make a good computer device? Why do some things work as computing elements and others fail, in spite of being the focus of massive research and development efforts?
Computers of any kind-electrical, optical or otherwise-are constructed by connecting a lot of simple circuits. Many millions of elementary electronic devices are in turn needed to form these circuits. The large number of components means that the base-level devices must be provided at a very low cost; they cannot be individually tested or adjusted.
The basic component of the modern computer is the integrated circuit, or chip, which is itself made up of several types of devices, including enormous numbers of transistors. Circuits are made today by processing circular silicon wafers, currently about 12 inches in diameter, that will be divided into a few hundred chips when processing is completed. The manufacturing process involves a long series of operations that exposes a group of wafers to chemical reagents and modifications of their surfaces, and subjects them to cycles of heating and cooling.
Conditions within the processing chambers all over a large wafer inevitably vary somewhat and result in small differences from wafer to wafer and between devices on different parts of a wafer. The devices produced are therefore not perfectly reproducible objects that meet precisely defined specifications. Further, devices are subjected to heating and cooling and large electric currents during use that can cause additional changes over time. Computing circuits thus must be able to function in the presence of significant differences between nominally identical devices.
Potential transistor replacements often seem to offer promising results under controlled laboratory conditions, but none as yet has managed to survive to the stringent requirements of mass production. Limitations on the type of computing signals needed form an even higher barrier for new devices.
In the computing process itself, data may pass through a long series of circuits on the way to producing a result. For example, computers are frequently used to simulate the evolution of a real-world physical system over time, such as studying the sequence of stages in an explosion or predicting climate change. Such simulations take the result of one step as the input for the next one. Small errors, such as might be caused by electrical noise or by an aberrant device, must not be allowed to accumulate during the execution of a lengthy program, as this may produce an error in the end results.
The way to prevent the propagation of errors through a series of operations is well known: digital representation of data. Information in digital form has been used since ancient times in such records as marks on a stick or stones in a jar, and in calculating for a thousand years with the abacus. Each bead on a wire of the abacus has two possible positions, which is what “digital” means: a small set of potential values of some physical quantity that is used to represent information. The two positions of a bead of the abacus are fixed by contact with the bars that hold the wires or with another bead. As in the two meaningful positions of an abacus bead, contemporary computers also work with information in the form of digits that may have two different values, called binary information or a “bit.”
The two possible values of a bit may be called zero (0) and one (1), or true and false. Computer logic executes functions defined by tables of zeros and ones. Logic functions include actions such as “AND” and “NOR:” In the first case, the function will return a “true” or “1” only if both the inputs are present, in the second case, it only returns such a result if neither input is present. Binary results then can be translated into other forms, such as decimal digits and alphabetic characters, for presentation to users.
The standard signal values in an electrical computer are two voltages that are the same throughout the machine. One voltage represents a zero and the other a one, for example. The result of an operation by a logic circuit is sent to a following circuit by setting a switch to make a connection to one of the two voltage standards. The signal that is sent does not depend on the characteristics of the circuit that activated the switch.
It is therefore vital that any device designed to replace a transistor in a logic circuit must be able to transmit data in discrete, digital form. Ambiguous signals will lead to errors and potentially the failure of the whole system. Transistors work with digital data because they only have two states, basically “on” and “off.” They are essentially switches, and indeed they are descended from such a physical apparatus.
The Evolution of the Switch
Early electrical computing machines used mechanical relays-small levers controlled by electromagnets-as the switches. Although relays worked, they were slow in comparison to electronic devices and were a source of frequent failures. Contacts could stick together or become dirty.
Relays were replaced in large systems by vacuum tubes, glass cylinders pumped free of air with a heated filament in one end and a plate at the other end. The filament emits electrons that are attracted to the plate by a positive voltage there. A sparse plane of wires called a grid intervenes between the filament and the plate. A small positive charge on the grid electrode controls the motion of a much larger charge, in the form of negatively charged electrons flowing from the cathode through the grid to the plate. And a negative voltage applied to the grid can repel the negatively charged electrons coming from the filament, cutting off the current.
The use of vacuum tubes permitted much faster operation than was possible with the mechanical motion in a relay, although heating their filaments consumed a large amount of power. Early, large one-of-a-kind computers, such as Colossus in Britain and ENIAC in the U.S., that convincingly demonstrated the usefulness of large computers, were based on vacuum tubes. The computers in the SAGE air defense network each contained 55,000 vacuum tubes and used three megawatts of electrical power. Vacuum-tube computers also spawned a significant computer industry, represented by, for example, the Sperry-Rand UNTVAC and the IBM 700 series in the early 1950s.
The invention of a solid-state switch, the transistor, in 1947 transformed electronics again. The small size, low power demand and very high reliability of transistors led to their rapid adoption by the computer industry and a variety of other electronics applications. Their quick, widespread implementation led Fortune magazine to designate 1953 as “The Year of the Transistor.” The transistor’s combination of desired properties made it feasible to assemble large numbers of them into a single system. The IBM 7030. “Stretch” computer, delivered in 1961, used 17,000 transistors.
Transistors are made with the class of materials called semiconductors, which conduct electricity under some conditions but prevent its flow in other circumstances. That property is needed for a switch to have on and off states. The nature of the quantum-mechanical energy bands that bind atoms together to form solids is such that it is possible to view currents in semiconductors as being carried by both positively charged and negatively charged particles. The positively charged particles are called “holes,” implying a place where an electron is missing. Electrons and holes are donated by small amounts of another material, called an impurity or a dopant, added to the semiconductor. Whether the semiconductor predominately transmits electrons or holes depends on the type of impurity. Materials that conduct electrons are called ?-type (for negative-type) and ones that conduct holes are called p-type (for positive-type).
Both n-type and p-type materials are needed in a transistor to make it work. The first type of commercial transistor, called a bipolar junction transistor, used a layer of one type of semiconductor between two layers of the other, creating two junctions of n- and p-types. A current applied to the center section by a thin wire, called a gate electrode, controls whether or not a current will pass through the transistor.
The field-effect transistor (FET), developed after the bipolar junction type, is the one that is commonly used in computers. A FET uses only one intersection of n-type and p-type materials. The gate electrode is separated from the semiconductor surface by a thin insulating layer. A current of electrons passing from a source electrode to a drain wire through the transistor can be controlled by the amount of charge on the gate. In an n-type FET, for instance, a positive charge on the gate electrode attracts electrons to the surface, creating a conducting channel in the semiconductor that current flows through.
Commercial transistors were initially discrete devices, which had to be individually wired into a circuit. The invention of the integrated circuit in 1960 further revolutionized the computer industry By allowing many transistors to be fabricated and interconnected on a single piece of silicon, a chip, the integrated circuit drastically reduced the cost of manufacturing transistors. The steady advance of the degree of integration after 1960 gained fame as “Moore’s Law.” As early as 1965, Gordon Moore noted that there had been an annual doubling of the number of devices on an integrated chip. The trend continued for many years, paving the way to increasingly affordable and powerful computers. A chip can now contain about a billion transistors.
Much of the growth in the content of a chip was achieved by making everything on the chip smaller. The cost of processing a wafer depends only weakly on what is on the wafer; the majority of the manufacturing expense is bound up in the number of wafers. So the increased number of devices per unit of silicon surface area led to a rapid decrease of the cost per device. This cost aspect of modern integratedcircuit technology led Gordon Moore to another remarkable generalization: the price of integrated circuitry has long been a constant one billion dollars per acre, in spite of the increasing number of transistors on that acre. The current price per transistor on an integrated chip is about 0.002 cents. A staple used for fastening together sheets of paper costs 10 times as much as a transistor.
Once the fabrication of a few thousand transistors on a single chip became possible, the microprocessor, a small computer-on-a-chip, was invented. The microprocessor made it possible to incorporate rather complex functions into various kinds of machinery, such as automobile engines and washing machines. Daily experience illustrates the broad impact of miniaturization and integration and the resulting low cost per device of integrated electronics. Human lives are enhanced by hearing aids and implantable devices that owe their existence to the miniaturization of electronics. (In fact, the hearing aid was one of the first commercial applications of the transistor, several years before its use in computing was realized.) The watch on my wrist was made possible by the development of compact circuits that could count down from a megacycle electronic oscillator to the response times of liquid-crystal displays and mechanical components. Replacement of cash by credit cards is feasible because of electronics that rapidly authorizes transactions. That’s not to mention the obvious devices, such as laptops, GPS units, MP3 players and handheld computing devices such as iPhones that connect wirelessly to the Internet.
Nevertheless, ever since the introduction of computers using transistors as switches, there have been attempts to replace the transistors with devices based on other physical principles. The quest for a better device continues in spite of the spectacular success of transistorized computing systems. Alternative systems have the potential to offer some advantages, such as increased computing speeds or greater resilience in harsh environments. But this potential has as yet proved impossible to reach.
The tunnel diode, a semiconductor device invented a decade after the advent of the transistor, was the subject of an early proposal for faster logic functions, as it could operate at high frequencies. These diodes have high levels ‘ of dopants at the junction between n-type and p-type materials, which aligns electrons and holes on either side of the junction. Due to quantum-mechanical effects, electrons tunnel through the barrier. The unusual feature of the tunnel diode is its negative resistance, meaning that, unlike ordinary conductors of electricity, when a voltage is applied to it, there is a range in which the current decreases as the voltage is increased. With such a negative-resistance element it is not difficult to construct an electrical circuit that can be in either of two stable states, a circuit that is said to be bistable. When the negative-resistance device is connected to a battery (or other source of electrical power), the circuit can be in either a low-voltage state or a high-voltage one. Any state in between is unstable, like a ball balanced on a knife blade, so the circuit will spontaneously go to one of the stable states.
It has been suggested more than once that the two stable states created by a circuit with a negative-resistance element could be used to represent zeros and ones for the purposes of computation. Logical operations would require that the states of two or more bistable circuits control the state of another circuit. Indeed, in laboratory demonstrations, very small tunnel diodes could carry a substantial current that could charge the capacitance of another tunnel-diode circuit quickly. These results inspired governments and large companies to support major projects to develop computers based on bistable negative-resistance circuits, hoping to realize machines that operated much faster than computers that used transistors. However, the lack of a way to produce a true digital standard signal demanded an attempt to use digital methods with accurate reproduction of the current-voltage characteristic from one circuit to another. The required accuracy was too difficult to achieve in even a few tunnel diodes, and attempts to use the tunnel diode in logic circuits were abandoned after 1966. Similar projects that used other kinds of negative-resistance devices met the same fate.
In the 1970s, much attention was devoted to the possibility of computing with bistable circuits that used super- conducting devices known as Josep son junctions. At a critical temperature, usually close to absolute zero, superconducting materials offer no resistance to electrical flow. Josephson junctions use two layers of superconducting material with a very thin layer of nonsuperconducting material sandwiched between them. The junctions could be in two different states, a zero-voltage superconducting state or a normal resistive state. In addition, the state of a junction could also be influenced by a magnetic field that might be produced by current in an- other wire controlled by another Josephson junction. These interactions between currents provided considerable flexibility in the design of bistable circuits. Very fast switching was demonstrated in this case as well, which encouraged massive development projects. The devices were found to be particularly sensitive detectors for tiny levels of voltage and magnetism. However, again there was no reference to digital signal standards, and the accurate control of device characteristics that was required to produce a standardized signal could not be achieved. In the early 1980s disillusionment set in and attempts to develop Josephson-junction computers were terminated.
For several decades, there have been numerous attempts to use quantum mechanics in computing. Electrons have an intrinsic spin and an associated magnetic moment. The spin can be in two directions, up and down, and methods have been developed to “polarize” elections into only one of these states. However, controlling electron spin takes large amounts of energy, and spin states naturally decay over time. In addition, detection of spin states remains difficult.
In the last few years, several research groups have tried to use microscopic lever-type switches to replace transistors. One approach at Hewlett-Packard uses nano-electrical-mechanical systems, devices on the nanometer scale that are also etched out of silicon, but which are designed for mechanical movement. Another experiment at Cambridge University uses carbon nanotubes (CNTs), hollow cylinders of carbon 1 to 2 nanometers in diameter and only an atom thick, that bend back and forth to close an electrical connection. These devices are essentially a miniaturized return to relays, the switches that started the computing revolution. Such approaches are in very early research and will require much development before they can even be explored as a commercial option. Only time will tell if these devices will fare any better than their early ancestors in computing.
As mentioned earlier, another major well-supported quest for faster logic in computing was based on the use of optical rather than electrical signals. Several approaches to optical computing were pursued in various laboratories. A few efforts were inspired by the discovery of optical bistability, which led to devices based on materials that could be in either of two states that have different degrees of transparency. Such bistability could be realized with a material whose optical properties depend on the intensity of the light traveling within it. Strong temperature dependence of the optical properties of favorable materials and a need for high accuracy in the physical dimensions of devices made for a difficult technology, and there was no way to refer a signal to a standard set of values. Some “optical” computing used the interaction of light with a semiconductor to actually perform logic operations, with light only serving to carry signals from place to place.
The 1990 demonstration of optical logic by AT&T depended on an electro-optic device that was difficult to implement and could not compete with purely electrical logic. Other ideas for using light in various ways as part of computing systems have been promoted and called “optical computing,” but none have been able to demonstrate any commercially feasible advantage over transistor-based computing.
The isolation of the switching signal from the output signal was an essential part of computing with relays. Nothing about the recipient of the signal sent on by a logic circuit influenced the circuit that produced the signal. In contrast, the input signal is shared with the output terminal in a bistable device. The connection of the device to its output can influence its action because the signal sent by the switch to the next logic circuit is generated by passing on the difference in currents between stable states, a quantity that can vary considerably from device to device. The result of the operation of the bistable circuit depends on the characteristics of the device itself, rather than a standard digital signal.
What of the future of transistors? Moore’s Law now represents a synergism between increasing computing power per dollar and new powerful applications made feasible by the reduced cost of circuits. It also is frequently viewed as a promise of what the electronic industry will deliver. Participants in the growth of the semiconductor industry have long asked themselves how far the trends that have been so successful until now can be extended. Gordon Moore identified three contributions to the steadily increasing number of components per chip. Obviously, increasing chip size and making everything on a chip smaller allow more to be placed on a chip. The chip content has increased faster than can be attributed to these factors, however. The third contributor to growth was modifications in device design that caused a device to occupy less space, such as putting parts on top of one another rather than side-by-side. This last component was dubbed “cleverness.”
A possible limit to present trends is obvious in miniaturization: Smaller devices are made from fewer atoms. How many atoms does it take to make a device? Surely, more than one, but no convincing answer has been offered. Currently chips are etched using light and photosensitive materials, but the size of the etched results is limited by the wavelength of the light. Going to extremely short ultraviolet wavelengths has helped to extend this manufacturing process. To take it further, fabricators are looking at using electron beams instead of light beams, but the electron-beam method is still very slow.
The difficult problems that the industry is facing right now, however, concern the more macroscopic properties of matter. As devices are reduced in size, a point is being reached at which electrons are passing through the everthinning barriers that are intended to prevent current flow by quantum tunneling. This problem is being attacked by a search for more effective barriers.
Also, the presence of more and more devices in a small space and the unwanted tunneling currents are producing increasing amounts of heat per unit of area, and removing the heat is emerging as a practical difficulty. Present-day high-performance microprocessors dissipate 50 to 100 watts on a chip with an area of a few square centimeters, typically at a rate of 30 to 50 watts per square centimeter. For comparison, hot plates and other cooking surfaces produce 3 to 5 watts per square centimeter. Cooling the chips is presently straining heat-transfer technology. Liquid cooling of chips is possible, at the cost of greatly complicating the physical hardware of the computer.
Programs may involve millions of operations, each of which takes a finite amount of time, and working speed is another aspect of computer performance that affects a system’s utility. Miniaturization in itself improves speed of operation: Electrical signals have less distance to travel and use less electrical charge to switch a device. Improved device designs also reduce the amount of charge needed to operate a device and facilitate the rapid motion of electrical charge.
The importance of the semiconductor industry to national economies has led governments and the world-wide industry to form the International Technology Roadmap for Semiconductors, a biannual report produced by teams of experts from five countries, which tries to anticipate future semiconductor advances. The Roadmaps describe improvements in materials and manufacturing methods that will be needed to realize the projections, aiming to provide guidance to research and development activities and to help equipment suppliers anticipate the needs of the semiconductor industry. The Roadmaps provide a glimpse of upcoming transistor technology.
The most recent Roadmap from 2007, and an update released in 2008, both stress the need for not just reducing component size, as per Moore’s Law, but also making chips more efficient and versatile. The reports list energy use as a major concern, as well as a movement to increase and improve performance in an equivalent amount of space.
An important way to fulfill these goals is the development of new materials that will replace silicon and allow chips to further diversify the tasks for which they can function. One success story has been integrated circuits printed on thin films of plastic that are then incorporated into light-emitting diode (LED) display screens and televisions. In another example, in 2008 Weixiao Huang of Rensselear Polytechnic Institute demonstrated the first working chip built on gallium nitride instead of silicon. Gallium nitride is the material on which blue LEDs are made, the technology responsible for the high-resolution Blu-ray video format. Gallium nitride works at hotter temperatures and has low sensitivity to radiation, so it can be used in more extreme environments. It also can be used in the microwave frequency region, opening up the use for such chips in wireless data transmission.
An advance much further out in the Roadmap’s projections is transistors that incorporate carbon nanotubes (CNTs). Phaedon Avouris and his colleagues at IBM Thomas J. Watson Research Center envision CNTs taking the place of a silicon channel in field-effect transistors. As transistors shrink in size, so do their channels, and at a certain point quantum tunneling comes into play. But CNTs, whose shape would allow a gate electrode to wrap completely around them, could provide strong coupling and control. CNTs conduct well and can produce the large on /off current ratios needed to effectively propagate computing signals. The problems with CNTs, however, remain numerous. Notably, at this point it’s hard to produce batches of CNTs with uniform properties, such as tube diameter. It may be some years before CNT technology advances to the point that they can be a part of massproduced transistors.
The Sole Survivor
Vacuum tubes and transistors succeeded as devices for computers because they can emulate a relay This is no coincidence, since the transistor was invented at the Bell Telephone Laboratories by a group that was formed specifically to find a solid-state replacement for the many relays used in telephone operations. As in a relay, a small signal can control a large signal in another circuit. This ability is called “gain” in electronics. The high gain of tubes and transistors (the large change in output produced by a small change in the input) makes it possible for them to act as switches, transmitting the result of a logic operation with a true digital signal obtained from one of a set of standard voltages available throughout the computer. The restoration of a signal to its intended digital value does not depend on accurate device-to-device reproduction of the switching characteristic signal. Considerable variability of the transistor’s output can be tolerated without affecting its ability to restore a signal to its intended value.
The attraction between positive and negative charges, and the repulsive force between like charges, make the gain of both transistors and vacuum tubes possible. In each case a small amount of charge on one element controls the movement of a far larger quantity of charge. This one phenomenon has been the sole physical effect capable of supporting electronic amplification since it was introduced with the invention of the vacuum tube that inaugurated the electronic age a century ago. The basic defect in proposed alternatives to the transistor in computing is that there is no other reliable way to supply the high gain needed to perform the switching action that connects a logic signal to a set of standard values that implement truly digital computing.
For more than half a century, many fruitless attempts have been made to find novel device technologies for computing. Methods that could be demonstrated in the ideal conditions prevailing in a laboratory were unable to cope with the hostile environment offered by large systems, where devices are packed close together to take advantage of the economies of mass fabrication and to minimize the time taken for signals to propagate between them. The performance of devices is influenced by heat produced by their neighbors. They must endure high current densities and large temperature grathents and must function for many years. In addition, signals traversing wires in a computer are subject to attenuation and distortion during transmission and may be contaminated with crosstalk induced by nearby wires.
In the face of all these imperfections, logic circuits must still make reliable binary decisions concerning the meaning of signals received from other devices. They must process signals and forward information without feedback. Success in preventing signal deterioration and loss in such an environment has only been achieved with digital technology, the establishment of standard signal levels throughout a system and the resetting of signals to the correct digital value at each step. Resetting is accomplished by connecting to standard terminal devices with high gain that can act as switches—transistors. The prominent defect in novel devices that have been seriously considered as alternatives is a lack of the gain needed to make the connections that restore signals to standard values. As a solid-state device that provides high gain, the transistor is unique and essential to computational electronics. It has changed much in size, material and fabrication method since its inception, and it may continue to shift with the times, but it will not soon be replaced.