In March 2036, a three-author team at imec in Leuven published a review in IEEE Transactions on Electron Devices consolidating twelve years of work on MORSe arrays (metal-oxide resistive switching elements) as on-chip weight stores for neuromorphic computation. The review is the device community’s first systematic statement of what the class can and cannot do, and its central negative finding, that no training rule native to the substrate has been demonstrated to approach gradient-based training, became the constraint that the following decade of cognitive system design was organized around.
The Device
A MORSe cell is a thin metal oxide sandwiched between two electrodes. An applied voltage drives oxygen vacancies through the oxide and assembles them into a conductive filament spanning the gap; reversing the bias ruptures the filament. The conductance left behind after a switching event is the stored value, and it persists with no power applied. The acronym compresses the element’s construction, and the review records that the field settled on it early over the alternatives precisely because it names what the device does rather than what it was once theorized to be. Three decades of materials work since the first reliable thin-film switching demonstrations have gone into making the switching repeatable, and the review’s early sections are a candid accounting of how much of that work remains outstanding.
The Array, and Why the Distinction Matters
Cells are fabricated at the intersections of a crossed-wire grid. A voltage applied across a row produces, at each intersection, a current equal to the product of the applied voltage and that cell’s conductance, and the currents sum along the column. A vector-matrix product therefore executes in a single read, at the location where the weights are stored.
The review’s central architectural claim is a distinction it argues is routinely elided in the neuromorphic literature. Digital neuromorphic processors of the TrueNorth, SpiNNaker and Loihi generation co-locate computation and memory physically, and are correctly described as avoiding the cost of shuttling activations to a separate memory die. They nevertheless store weights as values that are read out of a memory array before they can participate in an arithmetic operation. A MORSe array does not do this. The conductance is the weight and the multiplication is a consequence of the cell’s physics rather than of a fetch-and-multiply sequence. The review states the consequence directly:
“In a MORSe crossbar there is no instruction that fetches a weight, because there is no operation that could be performed on a weight that the weight is not already performing.”
The review notes that this is the only sense in which the device class earns the claim of eliminating the von Neumann bottleneck, and that claims made on behalf of digital neuromorphic architectures using the same language are, on inspection, about packaging rather than about computation.
Array-Level Constraints
The review devotes its longest section to limits, and its tone throughout is that these are properties of the physics rather than defects awaiting a better process.
- Device-to-device variability. Programmed conductances scatter across a fabricated array. Effective precision after programming is 4 to 6 bits per cell in the best arrays reported at the time of writing, against the 8 to 10 bits that single-device measurements in the materials literature frequently imply.
- Stuck-at faults. Cells that fail to switch, or that remain permanently conductive. At the array dimensions then achievable, fault rates required write-time masking or redundant columns to keep a programmed matrix usable.
- Read disturb and retention drift. Conductance shifts under repeated read, and decays over months without refresh.
- Sneak paths and IR drop. In a passive crossbar, current from a selected cell can return through unselected cells, and wire resistance attenuates the signal along a row. Both corrupt the vector-matrix product, and both worsen non-linearly with array dimension.
- Scale. Reported demonstrations sat between 10³ and 10⁶ cells. The review is explicit that the shortfall against biological scale is not a manufacturing problem awaiting a better fabrication line. The sneak-path and variability effects that begin to bind at 10³ cells bind harder at 10⁶.
The Absence of a Native Training Rule
The review’s most consequential section concerns not what the arrays cannot do but what nobody has made them do.
An array can be programmed to hold a chosen matrix by measuring each cell and applying write pulses until its conductance matches a target. This works, and it is the entire commercial case for the device class. It requires the target to come from somewhere else, and in every commercial and near-commercial use the review surveys, that somewhere else is conventional digital hardware running gradient-based training:
“The array is a deployment medium for a model trained elsewhere. It accelerates inference and contributes nothing to learning, which means that its economics are inherited from the paradigm it was expected to displace.”
Work on rules native to the substrate is surveyed at length and reported as unsuccessful, for three reasons the review treats as connected. Potentiation and depression are not symmetric in real devices, so a rule that strengthens and weakens by the same mechanism does not return the array to where it started. Write nonlinearity means a fixed pulse produces a different conductance change depending on the cell’s current state, so local updates cannot be composed reliably. Endurance is finite, so any rule that updates frequently wears out the array it is learning into. The review gives measured figures for all three.
It then states the conclusion that the rest of the field spent the following decade answering:
“The binding constraint on this field is not the absence of a manufacturable device. It is the absence of a training rule native to the substrate.”
The review closes by naming a possibility it declines to pursue. Every learning rule the field has attempted to port to these arrays is a rule for assigning credit for error against an objective. The authors observe that this may be a property of the algorithms rather than a property of learning, and that a substrate which stores its own weights at the point of use may be better matched to a rule with no objective to assign credit against. They state the hypothesis without endorsing it, and note that no such rule had been proposed.
Institutional Context
The work was supported by imec core funding and a European Research Council programme grant, with no industrial partner. The review observes in passing that MORSe research survived the 2031 contraction of commercial artificial intelligence comparatively undamaged, on the grounds that it had never been financed by a scaling thesis and therefore had no scaling thesis to lose.
Its section on relevance addresses one argument current at the time of writing. Following the 2026 announcements of domestic sub-5 nanometre production in China, it was briefly argued that continued conventional scaling would absorb the demand for alternative substrates before those substrates became necessary. The review rejects the argument on its own terms: “Feature size is not the quantity this paper is about.” The constraints it catalogues are set by read noise, device variability and sneak-path current, none of which improve as feature size shrinks.
The 2045 Connection
The neural core’s 2045 division of labour, with node state held in a digital control plane and edge weights in a MORSe array, is the arrangement this review identifies as the only one in which the device class’s efficiency claim survives its own limits. The 2045 specification’s node register set and its claim that the architecture has no von Neumann bottleneck are separate consequences of that single decision.
An editorial note on authorship: P. Bhaskaran, the review’s lead author, appears as a co-author on the Fennimore et al. (2045) neural core preprint, the only author on either paper to hold a device-physics rather than a cognition appointment.