The fourth basic element in electronics, the memristor promises a lot. But is it living up to expectations, asks E&T.
No one has yet suggested that the memristor might halt global warming but it can only be a matter of time. Ever since a team at Hewlett Packard Laboratories in California demonstrated the first one two years ago, there have been some very excited claims for the 'resistor-with-a-memory' including a way of saving energy in server farms, a way of extending Moore's Law and the route to new kinds of brain-type computers.
Memristors represent a new class of nanoscale passive device characterised by a resistance that depends on factors such as the magnitude and polarity of the voltage applied or, in the case of spintronics, the degree of magnetisation. Turn off the voltage or the magnetising source and the memristor remembers its most recent resistance without any source of power until you turn it on again. Some research devices have held their values for at least three years.
As synaptic connections between neurons in the brain behave like this too - becoming stronger or weaker depending on the characteristics of the chemical or electrical signal - scientists have speculated that memristors can be used to build computers that work like brains - the official term is neuromorphic. If not brains, then at least something that responds to sensory input in a similar way.
Realistically, the first commercial memristors will probably be in non-volatile memories because these are relatively simple to build and would benefit from the ability to scale memory cells to 10nm and below. HP, which is working on devices with switching speeds of one nanosecond, recently said it could have a competitor to flash memory in three years with a capacity of 20GB/cm2.
Memristance theory, courtesy of Leon Chua, preceded practical implementation by nearly 40 years, not least because the hysteretic behaviour Chua described is only significant at the nanoscale. Similar effects have, in fact, been observed in various materials and structures over the years. In organic field effect transistors, for instance, they arise from the movements of trapped charge-carrying ions, eroding the reproducibility of the FET's current-voltage characteristics.
Memristance can be described in a number of ways, one being a circuit component based on the relationship between charge and magnetic flux. A wider definition links a memristor's voltage with current and a state variable, such as the thickness of a titanium dioxide layer. In turn, the changing state variable depends on the amount of charge flowing through it.
HP Labs' first memristor was based on a titanium dioxide film, with a layer depleted of oxygen atoms (TiO2-x) creating the conductivity gradient. The charge-carrying vacancies behave like n-type dopants, lowering the depleted layer's resistance. When an electric field is applied, the vacancies move and change the boundary between the different layers, altering the resistance of the film as a whole.
Themis Prodromakis and colleagues from the Institute of Biomedical Engineering at Imperial College London more recently built complementary devices with an excess of oxygen in one layer (TiO2+x), which acts as a p-type dopant, so the enhanced layer's resistance increases when an electric field is applied. This group has also developed two fabrication methods based on standard lab processes that can make both n-type and p-type devices with a consistent reproducible electrical response, with dimensions controlled down to 5nm in thickness and 70nm laterally. They presented these results at the IEEE International Symposium on Circuits and Systems (ISCAS) in June 2010.
The ability to make complementary devices means memristor circuits can respond in both an inhibitory and excitatory way, either strengthening or weakening resistance when exposed to various inputs. 'Different polarities and reverse biased devices give you greater flexibility when designing any kind of circuit. But these kinds of devices are particularly interesting for building neuromorphic computers because they can fully implement neuron/synapse brain function,' Prodromakis told E&T.
Other research teams including Xiaobin Wang and Yiran Chen at the disk drive maker Seagate Technology are experimenting with spintronic memristors with the magnetic tunnel junctions (MTJ) used in disk drive read heads and MRAM memories.
As Wang and Chen explained at the Design, Automation and Test in Europe (DATE) conference in March, an MTJ cell comprises two or more ferromagnetic layers separated by nanometer thin oxide barrier layers. By fixing the state of one magnetic layer as a reference and changing the magnetisation of an adjacent 'free' layer using spin-polarised electrons, resistance becomes dependent upon the cumulative effects of the spin excitations. The free layer has two magnetic segments; one parallel and the other anti-parallel to the magnetic direction of the reference layer.
The key to a successful spintronic memristor, say Wang and Chen, is the ability to precisely control the movement of the interface between the segments, known as the 'domain wall'. For multi-bit MTJ cells, if the domain wall can be controlled to move and stop continuously across the free layer, the memory device can, at least theoretically, store continuous information through continuous resistance change.
Claims that memristors can help extend Moore's Law rest on combining CMOS transistors with crosspoint arrays of memristive devices to make configurable architectures that can perform logic functions, routing, an electronic analogue of a synapse or store information.
Last year in the Proceedings of the US National Academy of Sciences, HP gave details of a hybrid integrated test circuit that interconnected two 21x21 nanoscale memristor crossbars and several conventional silicon field-effect transistors (FETs). Using this test chip, they simultaneously routed multiple signals through the crossbar from and to the FETs and carried out a Boolean sum-of-product operation.
In a different experiment based on a portion of the same circuit, HP showed how the voltage from an operation in the integrated circuit could be used to reprogramme a memristor inside the crossbar array to have a new function, such as memory, interconnect or logic.
More recently, in a paper in Nature in April 2010, the same research team demonstrated how memristors could perform a novel type of logic operation called material implication (the 'if condition is true then consequence' construction) that requires devices to act simultaneously as logic and memory elements using resistance instead of voltage or charge as the physical state variable.
A former member of the HP team, Dmitri Strukov is now ploughing a similar but separate research furrow at the computer-engineering department of UC Santa Barbara, combining CMOS technology with stackable memristive crosspoint devices to develop a dense field-programmable gate array (FPGA).
'In today's commercial FPGAs, only 10 per cent of the circuit area is used for computation, the rest is interconnect and configuration circuitry with up to 50 per cent of the area taken by configuration bits usually held in SRAM,' Strukov told E&T at the DATE conference.
With further reductions in chip scaling, says Strukov, the FPGA real-estate problem will only get worse. 'In ASICs, you can add more metal layers but in FPGAs, since all the active routing circuitry, such as the configuration memory and multiplexers, are typically located on the silicon substrate, it doesn't help.'
In Strukov's approach, the FPGA configuration-bits and most of the routing are moved out of the silicon substrate and into a kind of club sandwich of multiple upper memristive and metallisation layers. The CMOS substrate provides high-density interconnects to the multiple crossbar layers through a single set of lithographically defined vertical via plugs. Connections between overlapping wires in adjacent metallisation layers are programmed by changing the resistance of the corresponding crosspoint memristor.
'One of the nice things about this is that the devices will have very tiny footprints. You can make the metallisation wires with CMOS technology and the very simple memristor structures with sub-CMOS scale features,' he says.
Strukov and his colleagues have developed a prototype place and route tool and a 100-gate scale demonstration chip, although integrating memristors on top of CMOS was a challenge.
Compared to CMOS FPGAs, Strukov estimates that his devices could be 100 times denser assuming the use of diode logic with 5-20 per cent defects. Diode logic is a way of using diodes - or memristors in this case - to construct Boolean logic gates. Only non-inverting functions may be implemented and because the output of any gate is further from a perfect 0 or 1 than the input only a few gates can be linked together.
Because memristors are passive devices and need transistors to drive them, efficient integration with CMOS is fundamental to making them useful. Spintronic memristors can apparently be combined with CMOS relatively easily using magnetic random access memory (MRAM) processes. But the technology HP is pursuing may be more demanding. 'We need a very flat, smooth substrate although we think that will be possible using something like a damascene process. A more challenging aspect is device variability,' explains Gilberto Ribeiro, a senior scientist at HP Labs.
An intriguing subject under discussion at HP Labs is whether the formulas used to describe variability in transistors are applicable to memristors. 'It is uncharted territory not only for us but for everyone in this field,' says Ribeiro. 'It is important to think beyond zeros and ones and to consider what else one can accomplish with these devices.'
Digital (two state) types of storage will inevitably be the first prototype demonstrations, he says. 'Coding more bits into a single cell is still an area of research. The off/on resistance ratio in our devices can exceed 1,000, which means that one can pack a large number of bits into a single cell. However, the variability in the way the devices respond to the external stimuli will delay these types of implementations.'
Even if non-volatile memristive memories sound a bit of a let down when compared to computers that behave like brains, there will be novel features. At DATE this year, Seagate researcher Wang outlined how spintronic memristors could provide a new form of information security. In order to access the data stored in a spintronic memristor, the user must excite the device electrically, an activity that can be memorised in a memristor and later revealed. 'When a user reads the data information stored in the device, the administrator who wrote the data knows immediately if they check the device,' says Wang.
But if you're waiting for a neuromorphic-computer, you may need to be a little patient.