GHz photonic lasers emulate two spiking neuron types on silicon

What the work claims

The authors claim that two-section InP quantum-well lasers monolithically integrated on silicon can act as high-speed photonic spiking neurons. By changing the DC bias, the same device can produce either integrate-and-fire-like or resonate-and-fire-like pulsation. The evidence is an experimental parametric sweep across cavity length and gain-to-saturable-absorber ratio, plus direct tests of threshold, temporal integration, and refractoriness in the integrate-and-fire regime.¹

How it works

Each device is a two-section laser: a gain section and a saturable absorber (SA) section. The gain current and the reverse bias on the SA are the two control knobs. Cavity length L and the ratio of SA length to cavity length R_ABS are the design knobs. The PIC uses Sagnac loops as cavity facets with reflectivities of 10% and 40%.

At low gain current and low SA bias the laser enters Regime I: broad pulses with full width at half maximum (FWHM) of 14 to 38 ns and a nearly constant pulse repetition frequency (PRF) of about 16 MHz. The PRF stays roughly fixed as current increases, while peak power rises. The authors link this to resonate-and-fire dynamics driven by optothermal "Canard" spikes.

At higher gain current (above 80 mA) and higher SA bias the laser enters Regime III: stable sub-nanosecond pulses whose PRF rises strongly with current up to 1.8 GHz. Peak power saturates shortly after the onset of pulsation. This is attributed to Q-switching from gain-SA dynamics and is taken as the photonic analogue of integrate-and-fire behaviour.

To test the integrate-and-fire analogy, the authors bias a laser near the Regime III boundary and apply 0.5 ns electrical pulses to the SA. A 0.8 V pulse elicits an optical spike; a 0.3 V pulse does not. Two 0.3 V pulses 0.4 ns apart do trigger a spike, while two pulses 1.1 ns apart do not, showing temporal integration and decay. Two 0.8 V pulses 0.5 ns apart produce one spike, but at 0.8 ns separation they produce two, indicating a refractory period. The inferred minimum resting period, 1 / 1.8 GHz ~ 0.55 ns, is shorter than the 0.8 ns recovery.

A design sweep across nine lasers with L = 400, 600, and 800 um and R_ABS = 3%, 5%, and 8% (with a 10% device used for L = 600 um) shows that Regime III is suppressed when R_ABS exceeds 8%, and that longer cavities widen the Regime III operating range and enable higher PRFs.

Where a skeptic should push

The most load-bearing assumption is the "neural isomorphism" itself: that laser self-pulsation is functionally equivalent to biological neuronal spiking. The authors demonstrate excitability, integration, and refractoriness, but they do so by applying electrical pulses to the SA section of a single isolated laser. There is no optical input from a silicon waveguide synapse, no network, no weighting, and no learning.

Second, the time scales are biologically extreme. A 0.4 ns integration window and a ~0.5 ns refractory period are six orders of magnitude faster than typical cortical neuron dynamics. It is not obvious what spike-time coding scheme can exploit such short windows, or whether fan-out, synchronisation, and plasticity can scale to useful networks at these speeds.

Third, efficiency is underspecified. The reported "external quantum efficiency" is 24.6 uW/mA at zero SA bias and 13.1 uW/mA at 3 V, but this is optical output per milliamp of drive current, not wall-plug or system efficiency. Coupling losses are 6 to 7 dB, the quantum efficiency is described as "relatively low" because of the low-reflectivity Sagnac facets, and energy per spike is not reported.

Fourth, the data are not publicly available, so the regime maps and the exception at L = 800 um, R_ABS = 8% cannot be independently checked. The R_ABS > 8% generalisation is contradicted by that exception, so the design rule "should not exceed 5%" is a cautious engineering heuristic rather than a proven bound.

OI implication: speed and the missing synapse

For organoid intelligence and biological computing, this paper is best read as a competing-substrate benchmark, not a supporting result. It demonstrates that a manufactured photonic device can replicate two canonical spiking-neuron behaviours at gigahertz rates in a CMOS-compatible PIC platform. That directly challenges the argument that event-driven spiking computation is the natural advantage of living tissue.

The opportunity is narrow but real. If a stable optical read/write interface to organoid neurons can be built, photonic devices of this kind could serve as a very fast routing or encoding layer for a hybrid system. The shared physical substrate - silicon - is a plausible integration target. But the paper does not demonstrate any such interface, and GHz optical pulses are far faster than the millisecond time constants of organoid neurons, so the coupling would require deliberate down-conversion or time-division schemes.

The threat is clearer. Biological computing often justifies itself by analogy to the brain's efficiency and temporal dynamics. Here, a non-biological device achieves temporal dynamics that are orders of magnitude faster, in a form that can be lithographically replicated. The remaining defensible claims for organoid substrates then become biological ones: self-repair, three-dimensional connectivity, developmental plasticity, and the ability to learn from biochemical signals. Those claims are genuine open hypotheses, not established advantages.

The most honest conclusion is that the speed and integration niches for biological computing have both narrowed. A photonic spiking neuron can already do what organoid advocates often gesture toward: fast, event-driven, low-latency spikes on silicon. What it cannot yet do - and what organoids might still offer - is to learn, rewire, and tolerate damage while operating. The burden of proof shifts to the biological side to show that those properties matter enough to overcome the life-support, variability, and I/O overheads that photonics avoids.

The bottom line

The paper establishes a concrete experimental result: the same III-V-on-silicon laser can be biased into integrate-and-fire and resonate-and-fire pulsation regimes, with the fastest regime reaching 1.8 GHz. That is a real advance in photonic neuromorphic devices. What remains unproven is whether these lasers can be assembled into scalable networks with programmable weights and learning. For organoid intelligence, the result sharpens the competitive landscape. It does not make organoid computing impossible, but it removes one common rhetorical crutch - that spiking dynamics are a unique biological advantage - and leaves the case for living substrates resting on harder-to-demonstrate properties such as plasticity and self-repair.