Matrix-vector multiplication is the operation neural networks spend most of their time on, and it is the operation photonic accelerators are built to do fast. The hard part is the weights. In a photonic accelerator, a weight is a physical setting — a phase shift, a transmission level — and in most designs built to date, holding that setting costs power continuously. Turn off the heater and the weight disappears. A paper published on October 24, 2026, in Scientific Reports offers a compact answer to that specific problem: a matrix-vector multiplication engine whose weights are stored in a phase-change material, persist with no power applied, and occupy a footprint of 185 by 21 micrometres. It is a design, validated in simulation. No chip was fabricated. That distinction belongs up front, because everything that follows is a model-based estimate of what the architecture could achieve — and the design is still worth taking seriously on those terms.
The authors are Ali Atri and Abbas Zarifkar, of the Department of Communication and Electronics at Shiraz University. Their engine performs matrix-vector multiplication using light in two transverse magnetic modes of a single waveguide, with the weights held in twenty cells of GSST, a phase-change alloy. The reported figures are striking for a device this small: a simulated total insertion loss of 2.76 dB for the fundamental mode and 3.64 dB for the first-order mode at the telecom wavelength of 1550 nm, operation across a 40 nm band, and switching pulses measured in nanoseconds. The question, as with every simulation-only photonic design, is how much of that survives contact with a fab.
The weight problem, stated plainly
A photonic neural network computes by routing light through a structure whose transmission encodes the weight matrix, then reading out the result at photodetectors. The most established way to build that structure is a mesh of Mach–Zehnder interferometers, each tuned by a thermo-optic phase shifter. But a thermo-optic weight is volatile in the practical sense: the phase shift exists only while the heater is powered. Hold a thousand weights and you are paying a thousand heaters’ worth of static power, continuously, whether the accelerator is computing or idle. For inference — where a trained model may sit loaded for weeks running the same weights — that holding cost is the power budget.
Phase-change materials offer the obvious escape. A cell of a chalcogenide alloy has two solid states, amorphous and crystalline, with sharply different refractive indices, and it stays in whichever state it was last switched to. No holding power, no refresh. The weight becomes a material property rather than a powered setting. The field has known this for years; the difficulty has been finding a phase-change material that changes its refractive index enough to matter without also absorbing the light it is supposed to route. The standard material, Ge2Sb2Te5 (GST), couples the two: a large index change comes with a large loss change, so a weight written in GST attenuates the signal it is meant to shape.
GSST — Ge2Sb2Se4Te1 — is the alloy engineered to break that coupling. In a 2019 Nature Communications paper, Yifei Zhang and colleagues at MIT introduced the Ge–Sb–Se–Te family and showed that the optimized composition combines broadband transparency from 1 to 18.5 µm with a refractive-index contrast of about 2.0, a figure of merit improved by more than two orders of magnitude over GST. Atri and Zarifkar’s design is built on that material property: weights written in GSST modulate light by changing its phase velocity, with far less of the absorption penalty that made GST weights lossy.
The design: two modes, twenty cells
The engine has three functional blocks, and the paper reports a simulated loss figure for each, which is what makes the design checkable at the component level.
First, a multimode-interferometer power splitter divides the input light, with a simulated insertion loss of 0.98 dB. Then a directional-coupler demultiplexer separates the fundamental transverse magnetic mode (TM0) from the first-order mode (TM1), with roughly 10 dB of crosstalk suppression and 0.87 dB of insertion loss. The two modes are the heart of the design’s throughput claim: mode-division multiplexing carries two independent input channels in a single waveguide, so the engine processes two data streams in the footprint of one. Doubling the modes doubles the parallelism without doubling the waveguides — the same logic that makes wavelength-division multiplexing attractive, applied to the spatial shape of the light instead of its colour.
At the centre sits the weighting core: four independently programmable paths, each containing five GSST cells in series, for twenty cells in total. Each cell is switched electrically between its amorphous and crystalline states, changing the transmission of its path. By programming the cells individually, the device can be reconfigured into 1,296 distinct matrix settings, according to the paper. The simulated extinction ratios — the contrast between a path’s most and least transmissive settings — are 11.29 dB for TM0 and 17.3 dB for TM1. Those are usable contrasts for analogue weighting, though the gap between the two modes is itself a reminder that the modes do not behave identically: the first-order mode, with more of its field near the waveguide edges, interacts differently with the GSST cells, and the design’s performance is mode-dependent throughout.
The complete engine, all three blocks together, shows a simulated total insertion loss of 2.76 dB for TM0 and 3.64 dB for TM1 at 1550 nm. For a full matrix-vector engine — splitter, demultiplexer, and weighted paths combined — those are low numbers. A loss of under 4 dB means most of the input light survives to the detectors, which matters because every decibel of optical loss in an inference accelerator eventually has to be paid back as laser power. The design also maintains functionality across a 40 nm window, from 1540 to 1580 nm, which tolerates laser wavelength drift and temperature variation, and leaves room for coarse wavelength multiplexing on top of the mode multiplexing.
The footprint is the figure that earned the headline. At 185 by 21 micrometres, the whole engine is roughly the size of a large dust mote. Photonic matrix engines are usually measured in square millimetres. This one is not, because it uses one waveguide and two modes where a conventional design would use many waveguides, and because GSST cells are sub-micron structures. Compactness is not a vanity metric here: in photonic computing, footprint determines how many engines fit on a chip, and therefore how large a matrix the chip can hold.
Switching in nanoseconds — on a simulator
The weights are written electrically, and the paper’s electro-thermal simulations put numbers on the process. A 5-volt pulse lasting 33 nanoseconds crystallizes a GSST cell. A stronger 12-volt pulse lasting just 5 nanoseconds melts the material, which then quenches rapidly back into the amorphous state. The asymmetry is characteristic of phase-change switching: crystallisation is the slow step, requiring the material to be held at temperature long enough to order; amorphisation is the fast one, a melt followed by a quench fast enough to freeze disorder in place. Both pulses are short by any standard — nanosecond-scale reconfiguration means the weights could, in principle, be rewritten between inference batches rather than only at load time.
Two caveats belong here. First, these are simulated pulse parameters from electro-thermal modelling, not measurements from a fabricated device. Second, the paper’s abstract reports no current or resistance figures, so no switching energy can be derived from the published numbers — and no energy-per-operation figure should be inferred. A 5-volt pulse tells you nothing about energy without knowing how much current flowed. The honest statement is that switching looks fast in simulation, and the energy cost is unreported.
What is established, by the material physics rather than by this paper alone, is the zero-static-power claim. Once a GSST cell is in its crystalline or amorphous state, it remains there without power — through idle periods and power cycles. For an inference accelerator that spends much of its life waiting between queries, that is the property that matters most: the power budget becomes the cost of computing and of rewriting weights, not the cost of holding them.
What the simulated numbers actually show
Read together, the paper’s figures describe a component with a coherent story: low loss, broadband tolerance, fast reconfigurability, and weights that cost nothing to hold. Each figure is individually plausible — the component losses are in line with what multimode interferometers and directional couplers achieve, the GSST behaviour matches the material literature, and the switching physics is standard phase-change behaviour. The design’s novelty is the combination: mode multiplexing plus non-volatile phase-change weights in a single ultra-compact engine, where previous non-volatile photonic weights have generally been demonstrated in single-mode devices.
Most importantly, the paper reports no neural-network result. There is no accuracy figure on MNIST or any other benchmark, no demonstration of the engine computing even a toy inference task. That is consistent with what the paper is: a device design paper, not an accelerator paper. But it means the gap between this engine and a working photonic accelerator is the entire system stack — input encoding, weight programming, detection, calibration, and a network mapped onto the hardware — none of which exists yet for this design.
The honest ledger
The limitations deserve the same weight as the figures, because they define what this paper does not establish.
First, and most fundamentally, nothing here has been built. The design is validated through electro-thermal and optical simulation. Simulated insertion loss is not measured insertion loss: real devices carry fabrication roughness, sidewall angles, overlay errors, and GSST films whose actual optical constants differ from the literature values used in the model. Until a device is fabricated and measured, the 2.76 and 3.64 dB figures are targets, not results.
Second, there is no system-level energy, latency, or throughput figure. No energy per multiply-accumulate, no operations per second, no comparison of a complete system — laser, modulators, this engine, detectors, control electronics — against a digital baseline. The energy case rests on the zero-static-power property of GSST, which is real, and on the low simulated loss, which is unmeasured. The conversion from those component properties to a system-level efficiency advantage has not been made, and it is exactly where photonic accelerator claims have historically come unstuck.
Third, the reliability questions that govern every phase-change device are unaddressed in the published material. How many switching cycles can a GSST cell survive before it degrades? Does repeated crystallisation drift the cell’s transmission? Do neighbouring cells thermally disturb each other when one is pulsed at 12 volts? This paper’s abstract reports no endurance data, and until such data exists, the reconfiguration speed is a capability without a lifetime.
Fourth, the two-mode asymmetry is a system problem, not just a device curiosity. An accelerator that routes half its data through a channel with 3.64 dB of loss and 17.3 dB of extinction and the other half through a channel with 2.76 dB and 11.29 dB needs calibration to make the two channels agree — and calibration is itself a cost, in control circuitry and in characterisation time. Adding more modes, the obvious scaling route, makes the asymmetry worse, since higher-order modes interact more strongly with waveguide imperfections and with each other. The 1,296 configurations, meanwhile, are the count of a small engine — four paths, twenty cells — not a matrix of any size that would run a real model layer. Scaling is asserted by the authors, not demonstrated.
Finally, a note on sourcing. Every figure in this article comes from the paper’s abstract and the bioengineer.org coverage of October 24, 2026. The full text was not accessible at the time of writing, so details that live in the body — the derivation of the 1,296 configurations, the waveguide geometry, any additional simulation results — could not be checked, and none are asserted here.
What would make this story matter
This desk has recently covered photonic AI hardware that exists: fabricated chips and deployed systems. This paper belongs in a different category — a design to watch — and it should be judged by the milestones appropriate to that category.
First, a fabricated device. The single most informative next step is a chip: the engine built, with measured insertion loss and extinction ratio for both modes, reported against the simulated figures. The gap between the two numbers would tell the field more than the simulation alone.
Second, measured switching: real pulse parameters, real switching energy per cell, and endurance data across many cycles. Nanosecond switching is only useful if the cell survives being switched, and energy per switch is the number that determines whether fast reconfiguration is affordable or merely possible.
Third, a network-level demonstration — even a small one. A toy inference task running on the fabricated engine, with accuracy reported, would move this from a component design to an accelerator programme. Without it, the engine is a multiplier in search of a network.
Fourth, scaling evidence: more modes in one waveguide without crosstalk swamping the signal, or tiled engines sharing a chip without thermal crosstalk between them. The compactness argument only pays off if the engine can be replicated.
Until those appear, the paper is best read as a clean statement of a design idea whose time may be coming: non-volatile weights are the missing piece for inference accelerators that do not burn power holding their weights, and mode multiplexing is a genuine way to double throughput per waveguide. The numbers are simulated, the chip does not exist, and the distance between the two is the whole story.
Sources and further reading
Primary source: Atri, A., and Zarifkar, A., “Non-volatile, compact, and broadband multimode matrix vector multiplication for photonic neuromorphic computing,” Scientific Reports (2026), published October 24, 2026, DOI: 10.1038/s41598-026-77326-7. Received May 13, 2026; accepted October 19, 2026. Authors affiliated with the Department of Communication and Electronics, School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran. All device figures — 185 × 21 µm footprint, MMI splitter 0.98 dB insertion loss, directional-coupler demultiplexer with approximately 10 dB crosstalk suppression and 0.87 dB insertion loss, twenty GSST cells in four paths of five, 1,296 matrix configurations, extinction ratios of 11.29 dB (TM0) and 17.3 dB (TM1), 5 V / 33 ns crystallisation and 12 V / 5 ns melt pulses, total insertion loss of 2.76 dB (TM0) and 3.64 dB (TM1) at 1550 nm, 40 nm bandwidth from 1540 to 1580 nm — are from the paper’s abstract; the full text was not accessible at the time of writing, and no figure from the body is cited. Coverage via bioengineer.org, October 24, 2026, confirms the same figures and states explicitly that the design is validated through electro-thermal simulations rather than a fabricated chip. Background on GSST: Zhang, Y., et al., “Broadband transparent optical phase change materials for high-performance nonvolatile photonics,” Nature Communications 10, 4279 (2019), DOI: 10.1038/s41467-019-12196-4 — Ge2Sb2Se4Te1 with broadband transparency (1–18.5 µm) and refractive-index contrast of approximately 2.0, breaking the index–loss coupling of GST.