A wafer of multibit magnetic synapses, and the gap between one device and a population
Magnetic RAM has been the binary synapse: fast, enduring, nonvolatile, but stuck at two states while analog rivals offered dozens. Zogbi, Giuliano, Kateel, Nguyen, Chatterjee, Palomino, Carpenter, Rao and Incorvia, from UT Austin, IMEC and KU Leuven, show that elongating the tunnel junction to a five-to-one aspect ratio unlocks up to eight resistance levels at the 300 mm wafer scale, with the in-plane magnetic field acting as a dial between deterministic binary switching and stochastic multilevel switching. Then they report the number that matters: when the measured device-to-device variation of that wafer is fed into a simulated crossbar, MNIST accuracy falls from 96.1% to 48.6%. The state count is a single-device result. The accuracy is a population result, and the two are separated by almost fifty points.
Source: Synaptic Multi-Weight Wafer-Scale Domain Wall Magnetic Tunnel Junctions Driven by Spin Orbit Torque, arXiv:2611.08405, 5 November 2026. Primary source. Read: the full 18-page PDF including device structures, the state-count analysis, the micromagnetic simulations, Table 2, the crossbar evaluation, and the conclusions.
What the work claims
This is an experimental device paper with a simulation-based network evaluation, and the split matters when weighing the two headline numbers. On the experimental side, the claim is that industrially relevant, 300 mm wafer-scale magnetic tunnel junctions can be multilevel synapses. Conventional wisdom has been that wafer-scale MRAM is limited to binary states, which is why existing multilevel spintronic synapse proposals resort to circuit tricks: shared electrodes across multiple junctions, encoding weights in RF response, or ganging binary devices. Here, four device configurations were fabricated on IMEC's 300 mm MRAM pilot line, 128 devices in total, and the elongated junctions, with length-to-width aspect ratios near five, produce stable multi-weight switching that the 60 nm circular control devices do not.1
The tuning knob is an in-plane magnetic field. At high fields the switching is deterministic and essentially binary. As the field is reduced, the distribution of switching amplitudes widens, domain walls nucleate and pin at varying locations, and the number of accessible resistance states grows: each elongated configuration reaches at least six states at a characteristic field, and one 400 nm device reaches eight states at zero field, a three-bit synapse. On the network side, the claim is more double-edged. Mapping the measured device statistics into a crossbar simulator yields 96.1% MNIST accuracy if all devices are assumed to achieve their configuration's maximum state count, but 48.6% when the measured device-to-device variation is included. Both numbers are in the paper; only one of them describes the wafer they built.
How it works
The devices are spin-orbit-torque MTJs with perpendicular magnetic anisotropy: a cobalt-iron-boron free layer on a tungsten heavy-metal track with a magnesium oxide barrier, fabricated with 193 nm immersion lithography. The stack's resistance-area product is 10.9 ohm square micrometres, and the tunnel magnetoresistance falls from 80.2% in the circular control to 44.8%, 39.1%, and 30.5% across the three elongated configurations, a readout-margin cost the authors attribute to a fabrication process tuned for conventional MRAM. Switching is probed with in-plane fields from minus 300 to 300 oersted in 50 oersted steps, 25 programming cycles per field.
The multilevel mechanism is domain-wall physics, and the paper's supporting simulations pin down its origin. During a switching pulse, a leakage current through the MTJ stack, which the authors denote as a compensation voltage applied to the top electrode, superimposes a spin-transfer torque on the spin-orbit torque from the track. That stack current is non-uniform across the junction, creating a lateral potential difference between the track and the edges of the top electrode, and domains nucleate on one side. At zero field the magnetization stays largely perpendicular during the pulse, domains form, and a domain wall can remain in the device and be pushed to different pinning sites by subsequent pulses, which is what generates the intermediate resistance states. At higher fields the in-plane component breaks the symmetry, switching becomes deterministic and complete, and the device collapses back to binary behaviour. The simulations, run at zero kelvin to isolate the domain-formation physics, reproduce this trend: remove the stack leakage current and no domains form at all, with the free layer simply canting in-plane. A resistance state counts as addressable if it is reproducibly reached after five or more consecutive pulses within a single programming cycle.
Programming is therefore stochastic and iterative, exactly as in filamentary RRAM. An iterative write-verify algorithm, with a cap of 25 cycles, repeatedly reads and pulses the device until it lands on the target resistance; the worst-case programming energy is 36.8 pJ for the best configuration, against 1.17 pJ for a single successful write. The device figures themselves are respectable: write speeds of 10 to 30 nanoseconds, write energies of 1.5 to 1.7 pJ per pulse, per-state write noise of 0.44 to 0.77%, and device-to-device variation of 5.1 to 5.7%, all measured across 20 devices per configuration, with a crossbar simulator mapping these statistics into a differential two-device bit cell supporting four-bit signed weights.
Where a skeptic should push
The most load-bearing assumption is that the best device's behaviour is representative of what a full crossbar gets. The paper itself supplies the refutation, and it is worth dwelling on. The eight-state result is one device in the 400 nm configuration; the network's 96.1% assumes every synapse in the array reaches its configuration's maximum measured state count, an assumption the wafer statistics contradict by construction, since the states appear stochastically and vary from device to device. The honest figure is 48.6%, nearly fifty points lower, and below that of a well-tuned binary-weight network in many published comparisons. The authors do not hide this; they foreground lowering device-to-device variation as the required next step. But a reader who stops at the abstract leaves with 96.1%, and that asymmetry is exactly the failure mode this field keeps repeating.
Push further on what was not measured. The network evaluation is a simulation fed with measured statistics, not a fabricated array running inference. Retention and endurance of the intermediate, domain-wall-pinned states are not characterized; the oft-quoted MRAM endurance of ten to the fourteen cycles is a literature citation for the technology, not a measurement of these multilevel states, and a pinned domain wall that survives 25 programming cycles is a different object from a synapse that holds a weight for ten years. Tunnel magnetoresistance has already halved relative to the control, so the read window is narrowing as state count grows. And the energy arithmetic deserves suspicion in both directions: nominal 1.5 pJ writes assume the write-verify loop usually succeeds early, while the 25-cycle worst case is sixteen to a hundred times more expensive depending on configuration.
Hold organoid synapses to the 48.6% test
The non-obvious implication for organoid intelligence is that this paper accidentally standardizes the honest way to report any substrate's synaptic capability. The silicon field's convention is to lead with the best-device, uniform-population number; the organoid field's convention is to gesture at "billions of synapses" with no population statistics at all. The 96.1%-versus-48.6% gap quantifies what that costs: nearly half the claimed capability evaporates when you insist the population match the measurement. Every claim about organoid compute capacity, plasticity-enabled learning, or synapse-count parity with silicon deserves the identical treatment: performance with the measured cell-to-cell, batch-to-batch, and developmental variation included, stated next to the ideal figure, or the number is decoration.
The mechanistic parallel is tighter than a reporting convention, because the device physics here is biological in spirit. The multilevel states are stochastic, appearing with device-dependent probability over repeated programming cycles, and they only become usable because an external write-verify loop supervises the stochasticity until the desired state sticks. Biological synaptic plasticity is the same architecture with the supervisor internalized: a stochastic molecular update rule, shaped by spike timing, that lands at a quasi-stable weight without any external read-write loop. The energy and latency overhead of the paper's 25-cycle write-verify, up to 166.8 pJ per programmed synapse in the worst configuration, is a concrete price tag for what tissue plasticity does for free at the point of storage. That comparison is the paper's quiet gift to biological computing: it prices the supervisor.
There is also a genuine opportunity in the control knob itself. The in-plane field trades state count against determinism along a smooth continuum: add symmetry breaking and the device becomes a reliable binary element; remove it and the same device becomes a stochastic multilevel one. A single physical substrate, two computational personalities, selected by a global control parameter. Neural tissue has the same structure at the level of modulatory tone, where diffuse neuromodulation shifts populations between regular and irregular firing regimes, and hybrid tissue-silicon systems could exploit both: silicon front-ends whose reliability mode is field-switchable, coupled to tissue whose excitability mode is pharmacologically switchable, with the control problem being to keep the two regimes matched. The threat, beyond inflated numbers, is specification drift in exactly those hybrids: a system architected around the 96.1% device will meet the 48.6% wafer in production, and an organoid stack architected around one healthy batch will meet developmental variability in the clinic.
The bottom line
Established by direct measurement: elongated spin-orbit-torque MTJs on a 300 mm wafer exhibit field-tunable multilevel switching, at least six states per configuration and up to eight in the best device, through domain-wall nucleation and pinning, with a plausible device-level mechanism supported by simulation. Established by simulation only: the corresponding crossbar performance, which is 48.6% under measured variation and 96.1% under an assumption the wafer data contradict. What would confirm the promise is retention and endurance data on the pinned intermediate states and a demonstrator array whose measured population statistics close the gap to the single-device result. For organoid intelligence the durable lesson costs nothing to adopt: report the population number, with variation, next to the hero number, and price the supervisor whenever the substrate cannot program itself.
Frequently asked questions
How many resistance states did the devices achieve?
Each of the three elongated configurations achieved at least six distinct resistance states at its optimal in-plane field, and one 400 nm device reached eight states at zero field, equivalent to a three-bit synapse. The 60 nm circular control devices remained binary. States were counted as addressable only if reproducibly reached after five or more consecutive switching pulses within a programming cycle.
What creates the intermediate resistance states?
Domain-wall physics. A leakage current through the MTJ stack during the switching pulse adds a spin-transfer torque that is non-uniform across the junction, nucleating magnetic domains on one side. At low in-plane fields the domain wall survives between pulses and pins at different locations, each pinning site giving a different resistance. At high fields the symmetry breaking makes switching complete and deterministic, collapsing the device back to two states. Micromagnetic simulations without the stack leakage current produce no domains, supporting this explanation.
Why do the two accuracy numbers differ so much?
The 96.1% MNIST figure assumes every device in the simulated crossbar reaches its configuration's maximum measured number of resistance states. The 48.6% figure uses the actually measured device-to-device variation, and because multilevel states appear stochastically and differ from device to device, the real population supports fewer effective weight levels than the idealized one. The gap of nearly fifty accuracy points is the cost of assuming the best device represents the wafer.
How does programming work, and what does it cost?
Programming is iterative write-verify, borrowed from RRAM practice: pulse the device, read its resistance, and repeat until the target state is reached, capped at 25 cycles. Single-pulse write energies are 1.17 to 6.67 pJ depending on configuration, but a worst-case 25-cycle sequence costs 29.3 to 166.8 pJ per synapse. Write speeds are 10 to 30 nanoseconds per pulse, and per-state write noise is below 0.8%.
How does this compare with other analog synapse technologies?
On the paper's own comparison table, the multilevel MTJs offer much lower write noise than RRAM and phase-change devices and far faster writes than electrochemical RAM, at picojoule write energies, with device-to-device variation around 5 to 6%. What they lack, relative to those rivals, is a demonstrated retention and endurance story for the intermediate states, and a read window that has already shrunk to roughly half the tunnel magnetoresistance of binary MRAM.
What does this mean for biological computing specifically?
Two things. First, a reporting standard: substrate capability claims should be quoted with measured population variation, as the paper itself does with 96.1% versus 48.6%, and organoid systems need that discipline even more. Second, a priced comparison: stochastic multilevel storage supervised by an external write-verify loop is what biological plasticity does internally, and the loop's energy and latency overhead, up to 166.8 pJ per programmed synapse here, is a concrete measure of what living tissue saves by updating its own weights.
References
- N. Zogbi, D. Giuliano, V. Kateel, V. D. Nguyen, J. Chatterjee, A. Palomino, R. Carpenter, S. Rao, and J. A. C. Incorvia. Synaptic Multi-Weight Wafer-Scale Domain Wall Magnetic Tunnel Junctions Driven by Spin Orbit Torque. arXiv:2611.08405 [cond-mat.mes-hall]. 2026. https://arxiv.org/abs/2611.08405. Accessed 2026-11-07.