A memory chip that can compute
Conjuguer processeur et mémoire, c’est possible !
Dans un ordinateur, on dissocie normalement le processeur et la mémoire. Tout cela pourrait bien changer. Une équipe internationale de chercheurs a trouvé une méthode pour que la « ReRAM », de la mémoire conservant les données en l'absence d'alimentation électrique, effectue des tâches informatiques initialement dévolues aux processeurs. Un gain de temps et d’espace certain !
Electronics has long relied on a division of labour. At the heart of myriad devices, from computers and smartphones to drones and dishwashers, a microprocessor can be found busily crunching data. Switch the power off, though, and this chip will forget everything. Devices therefore contain other, different sorts of chips that work as a memory. That is inefficient, because shuffling data between the two types of chip costs time and energy. Now, though, a group of researchers working in Singapore and Germany think they have found a way to make a single chip work as both a processor and a memory.
HOW MEMORIES ARE MADE
2. Both sorts of existing chip rely on transistors. These are tiny electronic switches, the ons and offs of which represent the ones and zeroes of the digital age. In the quest for speed, a processor’s transistors need to be able to flip rapidly between those two states. This speed is bought, however, at the cost of the forgetfulness that makes a separate memory essential. Meanwhile, the non-forgetful transistors used in a computer’s permanent form of memory are too slow to make useful processors. To make a chip which can do both has led some scientists to look at abandoning transistors altogether.
3. Among those scientists are Anupam Chattopadhyay of Nanyang Technological Uni-
versity, in Singapore; Rainer Waser of RWTH Aachen University, in Germany; and Vikas Rana of the Jülich Research Centre, also in Germany. The chips they are interested in are made of tiny “cells” instead of transistors. Each cell has two electrodes (a transistor has three), and these sandwich a layer of metal oxide. This oxide (commonly of tantalum or hafnium) changes its state of electrical resistance in response to pulses of charge passed through it by the electrodes. The change in resistance is caused by the movement within the oxide of some of the oxygen ions which make up its crystal lattice.
4.In a simple version of such a cell, a high state of resistance is read as a digital “one” and a low resistance as a digital “zero”. Crucially, the relocated oxygen ions stay put when the power is switched off. This means the arrangement can act as a data store, known as a resistive random-access memory, or ReRAM. Several chipmakers, including Panasonic, Fujitsu, HP, SanDisk and Crossbar (a Californian startup), have begun manufacturing ReRAM chips, and many in the industry think that, memorywise, they are the wave of the future.
GREAT POTENTIAL
5. Drs Chattopadhyay Waser and Rana, however, believe that to focus on memory is to undersell the new chips. They note that, though not as fast as a top-flight microprocessor, ReRAM nevertheless switches states much faster than conventional memory—fast enough, they think, for it to do computing as well as data storage. Moreover, ReRAM has other features that might make it a good processor.
6.With two instead of three electrodes, ReRAMs should be easier to manufacture and allow lots of cells to be packed tightly into a small space. Of particular significance is that, unlike a transistor, a ReRAM cell can be designed to do more than just switch “on” and “off”. It can, if built correctly, have multiple levels of resistance, each representing a number. Such a system would be able to store more data in a given space.
7.On top of that, it might not be confined to doing binary arithmetic. This matters, because certain computations which are hard and slow in binary logic might be managed easily and quickly in arithmetical systems of higher base.
DO THE MATH
8. So far, the three researchers have managed to construct a tantalum-based ReRAM with seven states of resistance. Eight are possible, and perhaps more, with more research. Eight levels is a good initial target, because it would permit the representation in a single cell of all possible three-digit binary numbers (ie, 000, 001, 010, 011, 101, 111, 110 and 100). A conventional chip would need three transistors to do this.
9. Sticking with binary arithmetic would make it easier to use existing software with such a system. But eight states of resistance could also, in principle, be used to do arithmetic directly in base eight. And, because eight is an exact power of two, swapping between the two bases in response to the requirements of the software involved could be done efficiently. Drs Chattopadhyay, Waser and Rana have not yet got that far.