The first new passive circuit element since the 1830s might transform computer hardware
Switching in Time
The transistor is a three-terminal device, with three connections to the rest of the circuit. It acts as a switch or amplifier, with a voltage applied to one terminal controlling a current flowing between the other two terminals. No such design is possible with memristors, which have just two terminals. But memristors can nonetheless be used to build both memory and digital logic. The key is to exploit the memristor’s built-in sense of history: A signal applied at one instant can affect another signal that later travels the same path. The first signal exerts this control by setting the internal state of the memristor to either high or low resistance.
The favored layout for memristor memory is a crossbar structure, where perpendicular rows and columns of fine metal conductors are separated by a thin, partially doped layer of TiO2. In this way a memristor is formed at every point where a column crosses a row. Each bit in the memory is individually addressable by selecting the correct combination of column and row conductors. A signal pulse applied to these conductors can write information by setting the resistive state of the TiO2 junction. A later pulse on the same pair of conductors reads the recorded information by measuring the resistance.
A near-term role for memristor crossbar arrays might be as a competitor for “flash” memory, the nonvolatile storage technology used in cell phones, cameras and many other devices. Each cell in a flash memory is a single transistor, modified for long-term storage of electric charge. The memristor structure is simpler and requires only two connections, so it might be made smaller than the flash-memory transistor. Thus there’s the possibility of higher density and lower cost.
Building logic circuits out of memristors would be a somewhat greater departure from current practice. In the early years of solid-state electronics, a technology called resistor-transistor logic had a brief vogue; the idea was to minimize the number of expensive transistors and maximize the number of cheap resistors. But with the coming of integrated circuits the economic incentives changed. With potential advantages in size and power consumption, memristors might shift the balance back toward a technology that combines active and passive devices. Williams and his colleagues have demonstrated a set of memristor logic gates that are computationally complete—they can implement any Boolean logic function.
Active components such as transistors would still be needed even if most information processing were done by memristors. The reason is that signals are reduced in amplitude by every passive circuit element, and at some point they must be restored to full strength. This requires a transistor or some other active device. Methods of fabricating hybrid circuits that combine transistors and memristors on the same substrate are an active area of investigation.
In binary digital circuits, memristors would operate as switches, toggling between maximum and minimum resistance. In this mode, the state of a memristor encodes one bit of information. If several intermediate resistances could be distinguished reliably, then the information density could be raised to two or three bits per device. The writing and reading processes would have to be calibrated to resolve four or eight levels of resistance. (Some flash memory chips already achieve this.) The end point of this evolution is to let the resistance vary continuously, and operate the memristor as an analog device.
One intriguing way to exploit analog memristors would be to build a machine modeled on the nervous system. In biological neural networks, each nerve cell communicates with other cells through thousands of synapses; adjustments to the strength of the synaptic connections is thought to be one mechanism of learning. In an artificial neural network, synapses must be small, simple structures if they are to be provided in realistic numbers. The memristor meets those requirements. Moreover, its native mode of operation—changing its resistance in response to the currents that flow through it—suggests a direct way of modeling the adjustment of synaptic strength.