Teams from Binghamton University in New York and Georgia Tech have been working on switching devices that mimic the operation of human neurons called neuristors. Instead of using niobium dioxide (NbO2) to replicate the switching behaviour observed in ion channels within biological neurons, the team have used a metal-Nb2O5−x-metal structure they call a memdiode. This provides the rectification, hysteresis, and capacitance necessary for high density neuristor circuitry and is much esasier to make on standard CMOS process technology as it does not need to electroform a conducting filament or a large external capacitor.
This could lead to cheaper, more energy-efficient, and high-density neuristor circuits than previously possible, accelerating the way to more energy efficient and adaptable computing, says Louis Piper, associate professor of physics and director of materials science and engineering at Binghamton University (above).
The memdiode operates as a voltage threshold triggered switch so that the neuristor oscillates when the circuit is excited by a current. Initially, a capacitor is charged in parallel with one switch, and that switch turns on when a certain threshold voltage is reached. This switch will also turn off at a voltage lower than the turn-on voltage. When this first switch turns on, the voltage increases on a second memdiode/capacitor pair which causes the second memdiode to switch. This time delayed switching raises then lowers the voltage, creating an oscillatory pulse.
The memdiodes act in the same way as previous NbO2 threshold switches (both act as voltage threshold switches) with the added benefits of integrated capacitance and as-deposited switching. In this way two such memdiodes can form the basis of a neuristor by incorporating both the switching element and the capacitance into one device that does not require electroforming. Replicating a human brain requires 104 synapses per neuron and 106 neurons/cm2, so electroforming each individual NbO2 memristor becomes unfeasible.
The memdiodes have distinct advantage as the native capacitance allows the memdiode to be scaled to nanometer sizes without large external capacitors for application in the neuristor circuit.
“You could put 5G and 6G everywhere and assume that you have a reliable internet connection all the time, or you could address the problem with hardware processing, which is what we’re doing,” said Piper. “The idea is we want to have these chips that can do all the functioning in the chip, rather than messages back and forth with some sort of large server. It should be more efficient that way.”
“One of the main problems we have with trying to make these systems is the fact that you have to do this electroforming step,” said Piper. “Like with Frankenstein’s monster, you basically pulse a large amount of electricity through the material, and suddenly it becomes an active element. That’s not very reliable for an engineering step with fabrication. That’s not how we do things with silicon transistors. We like to fab them all and then they work right away.”
Researchers at Georgia Tech created a range of Nb2O5−x devices with thicknesses ranging from 10 nm to 300 nm, which shows that the transition voltage scales as the logarithm of thickness with thicker devices requiring a larger voltage to transition to the conducting state. This contrasts with a traditional MIM diode based on Fowler-Nordheim tunneling where the turn-on voltage scales nearly linearly with thickness.
“We want to have materials that inherently have some sort of switching operation themselves, which we can then use at the same dimensions where we’re meeting the end with silicon,” said Piper. “The ability to scale and the ability to remove some sort of alchemy with regards to this electroforming process really makes it more in line with how we do semiconducting processing nowadays; this makes it more reliable. You can build a neuristor out of this, and because you don’t need the electroforming, it’s more reliable and what you can build an industry on.”
“The real effort at Binghamton has been toward trying to model, from an atomic point of view, the nature of these states, how they arise from physics and chemistry, and also instead of just looking at the inert materials and then correlating it with the device performance, can we actually see how these states evolve as we operate the device?” he added.
The paper, “Scalable Memdiodes Exhibiting Rectification and Hysteresis for Neuromorphic Computing,” was published in Scientific Reports.
- COMPLETE MACHINE LEARNING FRAMEWORK TARGETS NEUROMORPHIC SOC
- APPLIED, ARM TO DEVELOP CERAM FOR NEUROMORPHIC APPLICATIONS