All-Magnonic Neurons

• All-magnonic neurons are neuromorphic elements that use spin waves for weighted summation, nonlinear activation, and inter-neuron signaling entirely within the magnon domain.
• Experiments on YIG and Ga:YIG platforms demonstrate deterministic cascading, programmable thresholds, and physical pattern recognition using integrated magnonic circuits.
• They offer low Joule heating, GHz–THz operation, and scalability via reconfigurable coupling and inverse-designed components, paving the way for advanced neuromorphic architectures.

Searching arXiv for the cited papers to ground the response in current literature. All-magnonic neurons are neuromorphic elements in which inputs, internal processing, nonlinear activation, and outputs are all carried by magnons, or spin waves, in a magnetic medium, without conversion to electronics between neuronal stages. In this sense they differ from hybrid magnonic-electronic schemes that read spin-wave signals electrically and then re-drive subsequent stages. Recent work has moved the topic from conceptual building blocks such as analog adders, resonators, and inverse-designed switches toward experimentally integrated neural circuits in nanoscale yttrium iron garnet (YIG) and gallium-substituted YIG (Ga:YIG), including deterministic cascading, programmable thresholds, reconfigurable weighting, fading memory, and physical pattern recognition (Guo et al., 10 Jun 2026).

1. Definition and conceptual boundaries

An all-magnonic neuron performs weighted summation and nonlinear activation in the spin-wave domain and emits a spin-wave output that directly drives downstream magnonic nodes. The defining requirement is therefore not the absence of microwave or optical instrumentation at the experimental boundary, but the preservation of a magnon-to-magnon signal path inside the computational fabric. This distinction is explicit in both the integrated YIG threshold-neuron circuits and the broader magnonic-network perspective: the on-chip signal remains magnonic even when excitation is provided by microwave antennas and readout is performed by micro-focused Brillouin light scattering or related probes (Guo et al., 10 Jun 2026, Wang et al., 2023).

The motivation follows from four properties repeatedly emphasized in the literature: no charge transport along waveguides and thus no Joule heating in the propagation medium, GHz–THz carrier frequencies, coherent wave interference for summation and routing, and reconfigurability through fields, anisotropy control, domain walls, or spin–orbit effects. Within neuromorphic computing, these properties support several paradigms: traditional real-space magnonic artificial neural networks, reservoir computing in real space or $k$-space, and stochastic or probabilistic computing based on phase-bistable or thermal-magnon dynamics (Wang et al., 2023).

A recurrent misconception is that all-magnonic neurons are synonymous with phase-encoded majority gates. Earlier magnonic logic indeed relied heavily on linear interference and phase-sensitive encoding, which made cascading difficult and often required external amplification or precise phase control. The more recent threshold-neuron work instead emphasizes cascadable nonlinearity, signal regeneration, and phase-robust operation, while still remaining fully magnonic between neuronal stages (Guo et al., 10 Jun 2026).

2. Dynamical basis and activation physics

Across the literature, the common starting point is Landau–Lifshitz–Gilbert dynamics,

$$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$

with material-specific exchange, dipolar, anisotropy, and damping terms setting the spin-wave dispersion and attenuation. In out-of-plane-magnetized YIG threshold neurons, the relevant mode family is forward-volume spin waves with strong intrinsic nonlinearity and isotropic in-plane dispersion, while the slow-envelope dynamics can be written in nonlinear Schrödinger/Ginzburg–Landau form,

$$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$

In this description, the cubic term produces nonlinear frequency shifts and phase self-adjustment, and the local gain parameter $\eta$ becomes positive only when the pump flips into its high-emission state (Guo et al., 10 Jun 2026).

Several nonlinear activation mechanisms have been demonstrated or proposed. In integrated YIG threshold neurons, a pump-biased bistable region defines a firing threshold $I_{\mathrm{th}}(P_{\mathrm{pump}})$, and once activated the output is clamped:

$$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$

$$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$

This realizes self-normalized outputs largely independent of detailed input amplitudes or phases. In Ga:YIG analog neurons, the activation instead arises from an amplitude-dependent magnon frequency shift that feeds back on coplanar-waveguide excitation efficiency, giving a sharp trigger response, tunable fading memory, and a differentiable activation used in neural-network simulations,

$$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$

The same paper describes the decay envelope as $h(t)=e^{-t/\tau}$, with $\tau$ increasing by three orders of magnitude for a $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$0 pump-power change and reaching several microseconds (Breitbach et al., 22 Sep 2025).

Other implementations use resonant nonlinearities. The nanoscale magnonic nano-ring resonator employs a Kerr-like self-phase modulation and power-dependent detuning of a critically coupled notch resonance, producing threshold-like activation, power limiting, and bistability (Wang et al., 2020). Magnonic Fabry–Pérot resonators composed of YIG and CoFeB nanostripes exhibit a negative effective nonlinear shift within the cavity, causing transmission gaps to downshift with increasing power and enabling both threshold activation and transmission suppression (Lutsenko et al., 11 Feb 2026). Chiral two-dimensional resonators based on a permalloy disk above YIG use a nonlinear blue shift of an edge mode to modulate the amplitude and phase of wide-angle scattered spin waves (Fripp et al., 28 Nov 2025). A different theoretical route maps antiferromagnetic domain-wall dynamics onto the leaky integrate-and-fire form,

$$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$1

with the domain-wall displacement as the effective membrane potential and polarized antiferromagnetic magnons supplying the input current (Brehm et al., 2022).

3. Architectures and signal pathways

The experimentally demonstrated integrated threshold neurons in nanoscale YIG use high-quality liquid-phase-epitaxy films, most often $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$2-thick YIG, with three $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$3-wide waveguides merging into a central combining region linked to a pump-controlled output waveguide. A uniform out-of-plane field $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$4 establishes forward-volume spin waves, and microwave pulses at $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$5 are applied to both inputs and pump. Weighted summation is implemented physically by spin-wave combination under the pump antenna, while individual weights are tuned continuously by a DC current on the input antenna that modifies transmission through the asymmetric Oersted field and suppresses the effective weight $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$6 from approximately $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$7 to $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$8 (Guo et al., 10 Jun 2026).

Ga:YIG analog neurons use coplanar-waveguide antennas patterned on a $$\frac{d\mathbf{M}}{dt}=-\gamma \mu_0 \mathbf{M}\times \mathbf{H}_{\mathrm{eff}}+\frac{\alpha}{M_s}\mathbf{M}\times \frac{d\mathbf{M}}{dt},$$9-thick film with strong perpendicular magnetic anisotropy and in-plane magnetization. In this case each CPW acts as a neuron, and synaptic connectivity is provided by propagating magnons between neurons, in triangular many-to-one geometries for multi-input integration and in linear chains on a $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$0-wide waveguide for cascades (Breitbach et al., 22 Sep 2025).

The device space is broader than these two platforms. Brächer and Pirro’s analog magnon adder used a loss-compensated YIG delay-line resonator as a linear prequel to an all-magnonic neuron, integrating phase-coherent input pulses until a controlled nonlinearity is reached (Brächer et al., 2018). Inverse-design magnonics introduced a unified route to compact de-/multiplexers, a nonlinear switch, and a circulator on nanometer-thick YIG films, thereby assembling weighted summation, activation, and directional routing natively in the magnon domain (Wang et al., 2020).

Realization	Material platform	Reported neuron-relevant function
Integrated threshold neuron	Nanoscale YIG waveguides	Weighted summation, programmable threshold, self-normalized output, deterministic cascading
Analog trigger neuron	Ga:YIG thin film with CPWs	Sharp triggering, tunable fading memory, multi-input integration, cascadability
Resonator-based pre-neuron	YIG delay-line resonator	Analog leaky integration with phase-coherent summation
Inverse-designed blocks	Patterned YIG	Summation, nonlinear switching, unidirectional routing
Fabry–Pérot resonator	YIG/CoFeB	Threshold activation and transmission suppression
Chiral 2D resonator	Py disk above YIG	Wide-angle scattering and activation of secondary neurons
AFM domain-wall proposal	AFM insulator waveguide	Leaky integrate-and-fire dynamics with magnonic input and output

A second misconception is that all-magnonic necessarily means “no electromagnetic elements anywhere.” The literature instead draws a sharper systems boundary: inside the chip, neuron-to-neuron communication is purely magnonic; at the periphery, microwave antennas, optical probes, or future inductive or spintronic detectors may still be used without altering the all-magnonic nature of the internal signal path (Guo et al., 10 Jun 2026, Wang et al., 2023).

4. Cascadability, phase robustness, and network function

The central systems problem in wave-native neural hardware is not merely nonlinear activation, but cascadable activation with signal regeneration. The integrated YIG threshold-neuron circuits directly target this issue. Once activated, the neuron re-emits a standardized spin wave at the drive frequency with amplitude set by the nonlinear state rather than by the amplitude of the preceding signal, so each stage thresholds and normalizes simultaneously. In a two-stage Y-shaped circuit, the first neuron’s regenerated output can travel through a curved waveguide to become an input of the second neuron; when this combined signal exceeds the second threshold, the second neuron fires deterministically, and when the first neuron remains below threshold the second remains off (Guo et al., 10 Jun 2026).

Phase robustness is equally central. In the two-input threshold-neuron experiment, the output intensity is flat as the input phase difference sweeps from $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$1 to $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$2, and the output varies by less than a few percent while activation increases the intensity by about $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$3. The proposed mechanism is that the neuron responds to $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$4 through thresholding, while nonlinear frequency shift and magnon–magnon scattering drive the combined envelope toward a fixed point with clamped amplitude and a phase set by the local pump-driven oscillator (Guo et al., 10 Jun 2026).

Ga:YIG analog neurons realize a related but distinct notion of cascadability. In the three-neuron chain $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$5, nonreciprocal Damon–Eshbach emission ensures forward-only propagation, $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$6 fires only when both $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$7 and $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$8 are pumped and $$i\partial_t A+v_g\partial_x A+\frac{D}{2}\partial_x^2A+\Gamma |A|^2A=i\eta A.$$9 triggers the chain, and the resulting output increase is approximately $\eta$0 over sub-threshold excitation. The same platform also demonstrates time-domain coincidence and many-to-one fan-in with a two-input coincidence window of approximately $\eta$1, as well as frequency interoperability when neurons in the chain operate at different frequencies (Breitbach et al., 22 Sep 2025).

The broader literature extends these concepts beyond one-dimensional feed-forward chains. The perspective on nanoscaled magnonic networks emphasizes frequency-division multiplexing, nonreciprocal connectors, domain-wall synapses, and reservoir architectures, while the chiral resonator work reports reciprocal-space scattering over approximately $\eta$2, compressed to about $\eta$3 in real-space group-velocity directions, giving a two-dimensional fan-out that can activate multiple downstream neurons. This suggests that cascadability in magnonic neural hardware need not be restricted to serial repeater chains; it can also be implemented through directional routing, multiplexing, and geometrically weighted fan-out (Wang et al., 2023, Fripp et al., 28 Nov 2025).

5. Experimental demonstrations and quantitative performance

The most explicit integrated neural-circuit demonstration to date is the seven-neuron YIG circuit performing physical recognition of binary $\eta$4 letter patterns corresponding to “HUST.” The architecture comprises two three-input neurons and five two-input neurons in four layers, with $\eta$5 antennas total. Thresholds are programmed per neuron via pump powers, and the reported $\eta$6 classification matrix shows strong diagonal responses and more than $\eta$7 contrast in cases differing by a single input feature. In the same platform, a single three-input neuron can be tuned continuously through threshold regimes $\eta$8, $\eta$9, $I_{\mathrm{th}}(P_{\mathrm{pump}})$0, and $I_{\mathrm{th}}(P_{\mathrm{pump}})$1, corresponding respectively to “all three inputs required,” “any two inputs suffice,” “a single input suffices,” and “pump alone emits” (Guo et al., 10 Jun 2026).

Weighted classification has also been demonstrated directly in hardware. With equal weights $I_{\mathrm{th}}(P_{\mathrm{pump}})$2 and $I_{\mathrm{th}}(P_{\mathrm{pump}})$3, the three-input YIG neuron realizes a majority function with high outputs for “110,” “101,” and “011.” Suppressing one input weight to approximately zero yields selective recognition of patterns containing only the two weighted inputs. The same measurements report standard deviations over three repeats, indicating repeated consistent operation (Guo et al., 10 Jun 2026).

In Ga:YIG analog neurons, the measured self-activation curve exhibits a $I_{\mathrm{th}}(P_{\mathrm{pump}})$4 output increase within a narrow $I_{\mathrm{th}}(P_{\mathrm{pump}})$5 power window, with self-activation power $I_{\mathrm{th}}(P_{\mathrm{pump}})$6. Pump-controlled decay times reach several microseconds, far longer than the intrinsic magnon lifetime, and can be tuned by both pump power and pump frequency. When the experimentally validated activation function is inserted into a constrained PyTorch convolutional network with weights limited to $I_{\mathrm{th}}(P_{\mathrm{pump}})$7 and no learnable biases, the reported average test accuracies are $I_{\mathrm{th}}(P_{\mathrm{pump}})$8 on MNIST and $I_{\mathrm{th}}(P_{\mathrm{pump}})$9 on Fashion-MNIST. Under a bottleneck constraint in which the second-to-last layer is reduced from $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$0 to $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$1 neurons, trainable activation parameters increase Fashion-MNIST accuracy from $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$2 to $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$3 (Breitbach et al., 22 Sep 2025).

The resonator literature supplies complementary quantitative evidence. The analog magnon adder integrates pulse amplitudes from “1” to at least “1500,” with a fitted slope $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$4, a roundtrip time of approximately $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$5, and an intrinsic spin-wave lifetime of approximately $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$6 (Brächer et al., 2018). The magnonic Fabry–Pérot resonator shows transmission-gap downshifts up to approximately $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$7 between $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$8 and $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})=0 \quad \text{for } I_{\mathrm{in}}<I_{\mathrm{th}}(P_{\mathrm{pump}}),$$9, activation near $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$0, and suppression near $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$1, with simulations indicating approximately $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$2 amplitude suppression at the $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$3 resonance (Lutsenko et al., 11 Feb 2026). The chiral 2D resonator operates at an edge-mode resonance $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$4 with linewidth $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$5, and the scattered field can induce secondary nonlinear frequency shifts exceeding $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$6–$$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$7, sufficient to activate downstream neurons in the reported geometry (Fripp et al., 28 Nov 2025).

6. Scalability, limitations, and research directions

The case for scalability rests on several distinct arguments. Integrated YIG threshold neurons demonstrate fan-out and multi-stage cascading without the need for global phase alignment, and their nanoscale wavelengths favor compact layouts. The perspective paper adds system-level benchmarks for the magnonic channel: $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$8 for YIG $$I_{\mathrm{out}}(I_{\mathrm{in}},P_{\mathrm{pump}})\approx I_{\mathrm{sat}}(P_{\mathrm{pump}}) \quad \text{for } I_{\mathrm{in}}\ge I_{\mathrm{th}}(P_{\mathrm{pump}}).$$9, $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$0 for YIG $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$1, $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$2 for CoFeB $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$3, and approximately $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$4 for Ga:YIG $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$5, compared with a $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$6 CMOS reference of approximately $$c_{\mathrm{out}}=\sqrt{\sigma(c_{\mathrm{in}}^2;P_{\mathrm{Pump}},f_{\mathrm{Pump}})}.$$7. The same source stresses, however, that these benchmarks account for the magnonic channel only and that Ohmic losses of metal antennas can dominate unless magnetoelectric transducers are used (Wang et al., 2023).

Current limitations are explicit in the experimental papers. In the integrated YIG threshold-neuron circuits, explicit energy per operation and signal-to-noise ratio were not reported, and outputs were measured optically by micro-focused Brillouin light scattering. Threshold control relies on pump power, which incurs energy overhead, and more precise quantitative models of the bistable threshold and its dependence on geometry and materials are described as desirable for design automation (Guo et al., 10 Jun 2026). In Ga:YIG neurons, device-to-device variations and inhomogeneities influence foldover hysteresis and thresholds, while practical large-scale systems still require on-chip transduction or purely magnonic downstream stages rather than micro-focused Brillouin light scattering (Breitbach et al., 22 Sep 2025).

The device-physics literature adds further caveats. Resonator neurons balance selectivity against bandwidth, and their response depends sensitively on coupling gaps, linewidths, or stripe dimensions (Wang et al., 2020, Lutsenko et al., 11 Feb 2026). Chiral resonators require careful control of edge quality and higher-amplitude nonlinear processes (Fripp et al., 28 Nov 2025). Antiferromagnetic domain-wall neurons remain proposals, albeit proposals that offer THz operation, zero stray fields, and helicity-enabled inhibition (Brehm et al., 2022).

Several directions recur across the field: co-integrated microwave drivers and detectors, energy-optimized pump schemes, inverse-designed combining regions, programmable synapses based on domain walls or anisotropy tuning, larger two-dimensional and three-dimensional meshes, and on-chip learning. A plausible implication is that “all-magnonic neuron” should now be understood less as a single device archetype than as a design class united by one systems criterion: weighted summation, nonlinear activation, and inter-neuron communication remain in the magnon domain throughout the network. Within that class, recent integrated threshold circuits provide the clearest experimental evidence that signal regeneration, phase robustness, and deterministic cascading can coexist in a compact magnonic hardware platform (Wang et al., 2020, Guo et al., 10 Jun 2026).