The emulation ladder, rung by rung, toward tissue
A new survey chapter walks through phase-change memory, resistive RAM, and spintronic devices as artificial neurons and synapses, from filament physics to skyrmion racetracks. Taken seriously, it reads two ways at once: as a status report on the silicon competition to biological computing, and as an accidental spec sheet for exactly the properties a living synapse delivers for free.
Source: Non-Volatile Memory Devices for Neuromorphic Computing (book chapter), arXiv:2610.39625 (cond-mat.mes-hall), 29 October 2026; published version DOI 10.1002/9781394390274. Author: Debasis Das, BITS Pilani, K K Birla Goa Campus. Primary source. Read: the full arXiv HTML of the chapter, including the device-physics sections, the cited demonstrations, and the concluding comparison.
What the work claims
This is a survey chapter, and it should be weighted as one: it curates and explains other groups' device demonstrations rather than reporting new primary results. Its claim is organizational. Emerging non-volatile memory devices, meaning PCM (phase-change memory), RRAM (resistive RAM), and spintronic devices, can each emulate the two primitives neuromorphic hardware needs: spiking neurons, through integration-and-fire dynamics built from device physics, and plastic synapses, through multilevel conductance states updated by pulse timing. The chapter's conclusion is a pragmatist's ranking: PCM offers multilevel storage and mature fabrication but suffers drift and high programming current; RRAM offers low power and fast switching but struggles with device variability, endurance, and nonlinear asymmetric weight updates; spintronics, especially domain-wall and skyrmion devices, promises ultra-low energy and linear symmetric updates but remains early-stage.1
What makes the chapter worth a careful read is not novelty but compression. In about twenty pages it assembles the key demonstrations an organoid-intelligence reader needs to benchmark silicon against tissue: the device physics, the energy numbers, and the failure modes, all in one place.
How it works
PCM stores information in the phase of a chalcogenide glass: amorphous regions are highly resistive, crystalline regions conductive, with a resistance ratio up to roughly 10^3 between them. A short high-amplitude pulse melts and quenches the material amorphous (RESET); a longer lower-amplitude pulse crystallizes it (SET). Two emulation tricks follow. As a neuron, repeated weak crystallizing pulses thin the amorphous region and raise conductance until the cell snaps crystalline and fires, a physical integrate-and-fire that Tuma and colleagues demonstrated in phase-change devices.2 As a synapse, partial SET and RESET pulses walk the resistance through many intermediate states, and timed pre- and post-synaptic pulses potentiate or depress the cell, emulating spike-timing-dependent plasticity. The cited PCM synapse demonstration achieved about 10^7 switching cycles of endurance and, with 10 ns pulses, roughly 50 pJ of energy per RESET and 0.675 pJ per SET.3
RRAM switches by forming and rupturing nanoscale conductive filaments of oxygen vacancies in oxides such as HfO2, with a compliance current preventing damage during SET. The same physics emulates integrate-and-fire: one demonstrated PCMO device integrated sub-threshold pulses until conductance jumped sharply past 50 microsiemens, then reset, and a small network of such neurons classified Fisher's Iris at 95.33 percent against 96 percent for an ideal simulated leaky integrate-and-fire neuron. As synapses, HfOx devices have shown around 128 distinct conductance levels. The chapter is honest about the central problem: state-of-the-art RRAM updates are nonlinear and asymmetric under identical pulses, which degrades learning, and online training of even MNIST demands at least about 6-bit weight precision.1
Spintronics replaces structural change with magnetic state. A magnetic tunnel junction's resistance depends on the relative orientation of two ferromagnetic layers, and short current pulses that fail to cross the anisotropy energy barrier produce partial, relaxing rotation of the free layer, which is a physical leak; full switching is the spike. Domain-wall and skyrmion devices go further: a wall or skyrmion's position along a nanotrack can represent a synaptic weight continuously, giving linear and symmetric potentiation and depression under identical pulses, the property RRAM lacks. A bilayer-skyrmion demonstration implemented 3-bit synapses with per-bit switching energies on the order of a few femtojoules. The chapter also flags the obstacle: skyrmions drift off-axis under the skyrmion Hall effect, which bilayer structures with antiferromagnetic coupling can mitigate.1
Where a skeptic should push
First, provenance. Many of the strongest demonstrations the chapter cites come from the author's own collaboration network, and the survey includes no comparative benchmarking of the three device families on a common task. The energy figures, 0.675 pJ per SET, a few femtojoules per skyrmion bit, are single-device laboratory numbers, not array-level figures including addressing, periphery, and write-verify overhead, which historically erode device-level energy claims by orders of magnitude. Weight the chapter as a map of what has been demonstrated, not of what ships.
Second, internal accuracy. The chapter's STDP section states that potentiation occurs when the post-synaptic neuron fires before the pre-synaptic one, and it defines the timing interval as t_pre minus t_post. That is the reverse of the canonical spike-timing window in the hippocampal literature the text invokes, where pre-before-post pairing potentiates, and the interval convention is the opposite sign.4 The chapter is internally consistent but conventionally inverted, and a reader comparing curves across papers could misread every plot. Similarly, the framing comparison in the introduction, a megawatt-class AlphaGo against a 20-watt human, is a popular estimate repeated for rhetoric, not a measured quantity. These are correctable flaws, but they are exactly the kind a trusting reader absorbs uncritically.
Third, the missing benchmark. Every demonstration is scored on how well it imitates a biological synapse, yet no cited device reproduces the full spec at once: multilevel, linear, symmetric, endurance above 10^7 cycles, sub-picojoule writes, and decades of retention simultaneously. The gap between the device that does all of these and today's best is the honest measure of the field, and the chapter does not state it that way.
What device synapses teach the organoid case
Read against organoid intelligence, the chapter inverts. It is written as an account of silicon imitating biology, but it functions as a benchmark of biology against silicon, because the persistent failure modes it catalogs, nonlinear weight updates, drift, endurance limits, write variability, are precisely the properties a stem-cell-derived neural culture exhibits natively in their solved form: linear graded plasticity at real synapses, self-repair on cellular turnover, analog storage in molecular concentrations, and energy per synaptic event in the femtojoule range. The emulation ladder has a top rung, and the top rung is the thing being imitated. For anyone arguing that organoid computing's advantage is efficiency, this cuts both ways, which is why the chapter deserves attention from the wetware side rather than only the device side.1
The opportunity is hybrid and specific. A cultured network's weaknesses are the silicon strengths the chapter documents: non-volatile, drift-free weight storage for long-term task memory (PCM and spintronics retain for years; an organoid's electrophysiological state drifts over days), deterministic readout interfaces, and 10^7-cycle write endurance that no culture survives. The architecture this suggests is a division of labor the field has already half-adopted: living tissue as the adaptive, power-frugal dynamical front end; eNVM as the non-volatile associative memory bolted on at the interface. The chapter's device-physics detail, particularly the femtojoule skyrmion numbers and the STDP pulse-timing schemes, is directly usable engineering input for the CMOS side of an organoid co-processor.
The threat is the trajectory, and it is concrete. The energy argument for wetware has rested on biology operating near thermodynamic floor per synaptic event; the chapter's own numbers, sub-picojoule PCM SETs and few-femtojoule skyrmion bit writes, show artificial synapses entering the same decade of energy while beating tissue on endurance, retention, and manufacturability. If the near-term device families close the linearity gap that RRAM still has, the remaining wetware advantages narrow to adaptivity and three-dimensional self-assembly, which are real but harder to monetize and slower to improve. There is also an underappreciated second threat: the field's benchmarks. Because every eNVM demo is graded against biological synapses, the community keeps refining the comparison target; organoid intelligence should take that seriously and publish its own hard numbers on plasticity linearity, write endurance, and energy per task, before the comparison happens without it.1
The bottom line
Established, as surveyed facts: phase-change, resistive, and spintronic devices have each demonstrated physical mechanisms for integrate-and-fire neurons and plastic multilevel synapses, with cited laboratory figures including 10^7-cycle PCM endurance, sub-picojoule PCM SET energy, about 128 RRAM conductance levels, Iris classification at 95.33 percent with a PCMO neuron, and few-femtojoule skyrmion bit writes. Established, as honest survey conclusions: RRAM updates are nonlinear and asymmetric, PCM drifts and draws high programming current, and spintronics is early-stage. Not established anywhere, in this chapter or the field: a single device family delivering the full biological synapse spec at array scale. For organoid intelligence, the chapter's value is diagnostic. It shows the competition's scorecard, and the scorecard says the silicon imitation of a synapse remains decades of device physics short of the original, while the energy margin that once justified computing on living tissue is narrowing on the strongest single metric. The field should answer with its own published numbers, not with the assumption that biology wins by default.
Frequently asked questions
What are the three device families the chapter compares?
Phase-change memory (PCM), which stores data in the amorphous versus crystalline phase of a chalcogenide glass; resistive RAM (RRAM), which switches by forming and rupturing conductive filaments in oxides such as HfO2; and spintronic devices, which store information in magnetic orientation, including magnetic tunnel junctions, domain-wall racetracks, and skyrmions.
How can a memory device act like a neuron?
By exploiting the physics of gradual switching. In PCM, repeated weak crystallizing pulses slowly raise conductance until the device snaps to the crystalline phase, emulating integration and firing; in magnetic tunnel junctions, pulses too weak to cross the anisotropy barrier cause partial, relaxing magnetization rotation, which emulates leak. A demonstrated PCMO RRAM neuron classified Fisher's Iris at 95.33 percent, close to 96 percent for an ideal simulated leaky integrate-and-fire neuron.
What is the biggest unsolved problem for artificial synapses?
Simultaneously meeting the full biological spec. No single device family today combines multilevel states, linear and symmetric updates, high endurance, low write energy, and long retention at once. RRAM, for example, offers about 128 conductance levels but nonlinear, asymmetric updates that degrade learning; PCM drifts and draws high programming current; spintronics has excellent update symmetry but remains early-stage.
Is the chapter's spike-timing description standard?
No, and readers should be careful. The chapter states that potentiation occurs when the post-synaptic neuron fires before the pre-synaptic one, and defines the timing interval as t_pre minus t_post. That is internally consistent but the reverse of the canonical STDP window in the hippocampal literature, where pre-before-post pairing potentiates, and it uses the opposite sign convention. Comparisons across papers require checking which convention each plot uses.
Why does a device survey matter for organoid intelligence?
Because it documents the competing substrate's progress in numbers. Sub-picojoule PCM writes and few-femtojoule skyrmion bit writes show artificial synapses entering biology's energy decade, while endurance of 10^7 cycles and years of retention are things no culture matches. The chapter doubles as a scorecard for how much of wetware's claimed advantage remains, and which parts, adaptivity and self-assembly, survive the comparison.
What hybrid architecture does the chapter suggest?
A division of labor: living tissue supplies adaptive, energy-frugal dynamics and native plasticity, while eNVM devices supply what cultures lack, non-volatile weight storage that does not drift over days, deterministic interfaces, and write endurance no biological substrate survives. The chapter's device physics and pulse-timing schemes are direct engineering input for the silicon half of such a hybrid.
References
- Das D. Non-Volatile Memory Devices for Neuromorphic Computing. Book chapter, DOI 10.1002/9781394390274; arXiv:2610.39625 (cond-mat.mes-hall). 2026. https://arxiv.org/abs/2610.39625. Accessed 2026-11-04.
- Tuma T, Pantazi A, Le Gallo M, Sebastian A, Eleftheriou E. Stochastic phase-change neurons. Nature Nanotechnology 11: 693-699. 2016. https://doi.org/10.1038/nnano.2016.70. Accessed 2026-11-04.
- Kuzum D, Jeyasingh RGD, Lee B, Wong H-SP. Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing. Nano Letters 12(5): 2179-2186. 2012. https://doi.org/10.1021/nl201040y. Accessed 2026-11-04.
- Bi G-Q, Poo M-M. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience 18(24): 10464-10472. 1998. https://doi.org/10.1523/JNEUROSCI.18-24-10464.1998. Accessed 2026-11-04.