A memristor is a device whose resistance remembers the voltage it has seen, so the same spot both stores a number and multiplies by it. Chips built from arrays of them work more like a brain than like a processor with memory bolted on the side – which is what neuromorphic means.
Going deeper
A memristor’s resistance depends on the voltages it has already seen, so one device both stores a number and multiplies by it. Arranged in a crossbar, an array of them performs a whole matrix multiplication where the data already sits, instead of moving it to a processor.
A resistance that remembers
Apply a large enough voltage across certain thin and their resistance changes – and stays changed when the voltage is removed. The usual microscopic cause is motion: oxygen drifting into a conducting filament, or metal atoms migrating from an electrode. Sweeping the voltage traces a pinched loop, current against voltage, passing through the origin: the signature of a memristive device.
In the switching layer can be a single . A monolayer between two metals switches non-volatilely – the “atomristor” result – with the switching attributed to metal atoms substituting into . That is about as thin as a memory element can be, and the layer can be grown or at low temperature, which matters for stacking memory above logic.
Why this suits neural networks
The dominant operation in a neural network is multiply-and-accumulate: many inputs, each scaled by a weight, summed. In a digital processor, weights are fetched from memory to the arithmetic units, and that movement costs more energy than the arithmetic. A crossbar inverts the arrangement: weights are stored as conductances at the crossing points, input voltages are applied to the rows, and Ohm’s law multiplies while Kirchhoff’s law sums the currents down each column. One matrix-vector product happens in one step, in the analogue domain, where the data lives.
Devices that change their conductance gradually can also imitate a synapse directly, strengthening or weakening with pulses – which is where the word neuromorphic comes from.
What stands in the way
Analogue accuracy is the hard part. Conductance states drift, vary between devices and between cycles, and are not perfectly linear in the programming pulse; sneak currents flow through unselected paths unless each cell has a selector; and converting between analogue and digital at the edges of the array consumes much of the energy saved. Networks therefore have to be trained to tolerate the hardware, or trained on it.
For 2D materials specifically, the attractions are thinness, low switching energy and back-end-compatible integration; the open questions are endurance, retention at operating temperature, and whether uniformity can be reached. Published figures usually come from a handful of devices, and arrays large enough to run a real network remain rare.
For specialists
Non-volatile resistive switching used as memory and as an analogue synaptic weight, so that a crossbar performs multiply–accumulate where the data already sits. Two-dimensional materials contribute switching layers under a nanometre thick – vacancy or metal-atom migration in a monolayer TMDC gives non-volatile switching – along with synaptic and back-end-compatible integration above logic.