In this paper
AbstractIntroductionSynchronous digital circuits as analogue spacetimesDesign of a synchronous digital circuit as an analogue black holeHardware implementationMeasurementsConclusionsAcknowledgmentsAppendixThe ruleThe holeThe deviceVerificationMeasurement protocolsError analysisConverters and noiseOffset of the instrumentCalibrationHorizon from the profileTonesEnergy flux and gainThe causal cone and the dispersion of the ruleReferencesWorksPhysics and Astronomy
Engineering a digital acoustic black hole
Abstract
We show how to design and build an analogue black hole as a synchronous digital circuit. We define the sound speed, the flow and the horizon of such a circuit in terms of its own logic, and realise one on a field-programmable gate array, where the black hole is probed electrically with a signal generator and an oscilloscope. We measure the stability of the horizon, its causal cone, the ringdown and rotational superradiance. A single flipped bit inside the horizon leaks out by an amount that falls exponentially with depth. Digital circuits thus form a reconfigurable class of gravitational analogues and realise an analogue black hole as a cellular automaton.
Introduction
In 1981, Unruh identified an analogy between black holes and flowing fluids [1]: sound cannot travel upstream in a fluid that moves faster than sound, just as light cannot escape from a black hole. Sound waves in a moving fluid obey the same equation as a field propagating in curved spacetime, with the surface where the inward flow reaches the speed of sound acting as an event horizon.
Such analogue black holes bring into the laboratory phenomena like Hawking radiation [2], the ringing of a perturbed black hole, and the amplification of waves by a rotating one (superradiance [3, 4]), turning theoretical questions into experimental ones: whether these effects survive the short-distance structure of the medium; how a horizon acts on the waves that reach it; and which features of black holes are universal, and which are particular to Einstein’s equations. Laboratory experiments have realized analogue black holes with flowing water [5, 6, 7, 8], ultracold atomic gases [9, 10], optical fibres [11, 12], superfluid helium [13] and superconducting circuits, proposed as Hawking-radiation analogues [14, 15] and realised as a qubit chain [16]. Early theoretical work includes Refs. [17, 18, 19, 20, 21, 22]; See Ref. [23] for a review.
Here we describe and realise a new class of analogue black holes, implemented as synchronous digital circuits on a field-programmable gate array chip. The chip’s registers hold a grid of cells, and a local update rule, wired as logic gates, is applied to every cell at each clock tick. Specific choices of the update rule produce a flow with a horizon. Such a device is an ordinary digital circuit and can be probed by injecting an electrical signal, which is converted into a wave that scatters off the horizon; the response is read as a voltage and recorded on an oscilloscope. It is a classical digital device: its horizon is a property of the state of a synchronous logic circuit, i.e. a cellular automaton in which every quantity is a word of bits. Each cell of the digital circuit stores a finite number of bits in its registers, so the wave, the flow and the horizon take only finitely many values, and the dynamics is an exact map on this finite set of states. What such a device offers is an analogue black hole whose short-distance structure is known exactly. Its dispersion relation follows from the update rule of the circuit, and whether horizon physics survives this structure can be tested by direct measurement.
The device is used to measure five properties of the analogue black hole, each read as a voltage: its stability, the horizon itself, the causal cone that defines it, the ringdown (the quasinormal modes of the hole in its reservoir) and superradiance.
Synchronous digital circuits as analogue spacetimes
The defining element of black hole analogues is the emergence of a horizon; this requires a disturbance, a speed at which the disturbance spreads, and a flow that carries it. In a fluid these are sound, the speed of sound and the velocity of the fluid. Each has a precise counterpart in a synchronous digital circuit.
A synchronous digital circuit consists of registers, which hold bits, combinational logic, which computes new bits from old ones, and a clock: at every tick all registers are replaced at once by the output of the logic [24]. Let the registers form a regular array of sites, each holding a word of bits, and let the logic, the same at every site, read only the words within a distance [25]:
with a fixed Boolean map, static bits that may differ from site to site, an external input, and so that the neighbourhood is a square.
The circuit is deterministic, so a disturbance has an exact meaning. Consider two copies of the circuit, run with the same input, whose states at are identical except for one flipped bit, and follow the sites at which the two copies hold different words. After one tick a word can differ only if read a word that differed, that is, only within sites of the flipped bit: at most a square, and after ticks at most a square of . Whether the difference fills that square depends on and on the state. In the plane of position and time, the sites that may differ fill a cone with apex at the flipped bit and slope : the cone of the wiring (a classical analogue of the Lieb–Robinson bound of lattice quantum systems [26, 27]). It is the same at every site and in every direction, and by itself it contains no horizon. Figure 1 draws the circuit, the rule and this cone for .

Figure 1. A synchronous digital circuit as a medium. (a) Registers hold the words ; the logic , the same at every site, computes from the words within , the static bits and the input , Eq. (1), and at every clock edge all registers are replaced at once. (b) The registers as an array of sites, each holding a word of bits; shaded, the words within of (drawn for ). (c) Two copies run with the same input and one bit flipped at : the sites at which they may differ (black) are those within of the flipped bit, a square of side , a cone of slope in every direction. This is the speed limit of the wiring.
How the disturbance propagates inside the cone of the wiring depends on and on the state of the circuit. Suppose that implements conservation laws. A physical example is a “mass” associated with each site, with the word read as a list of numbers, and a “total mass” that the conservation law requires to stay constant. The same can be done for a quantity analogous to the momentum, with a direction attached to each entry. Such definitions are those of a lattice-Boltzmann scheme [25, 28] and allow one to interpret the evolution of the digital circuit, on scales much larger than , as that of a fluid. Assuming the circuit settles to a steady state, a small disturbance of that state travels as a sound wave, at a speed set by and by the state and bounded by . The bulk of the difference between the two copies then stays inside a narrower cone, that of the sound; the rest spreads by dispersion at the speed of the wiring, decaying at every tick.
Let be the speed at which the edge of this narrower cone, the one that holds the bulk of the difference made by a bit flipped at , advances in the direction ; it may be negative. After ticks the cone extends from to along : its width grows at the rate and its centre moves at the rate . This suggests to define
as the rate at which the cone opens and the rate at which it drifts: the speed of sound of the circuit and the velocity of its flow, defined by the circuit properties alone. If the rule disperses, and are then the rates for disturbances long compared with the spacing of the sites, and the shorter ones run ahead or behind as precursors. Figure 2 shows the cone with no flow, with a subsonic and with a supersonic flow.

Figure 2. Sound speed and flow velocity read off the cone. Space–time diagrams of the sites reached by the difference between two copies of the circuit after one bit is flipped at (black mark): the cone of the wiring (grey, slope ) and, inside it, the sound cone (blue), whose edges advance at and ; dashed, its centre. (a) No flow. (b) Inflow slower than sound: the cone drifts with the flow (red arrow) but still opens on both sides. (c) Inflow faster than sound: , the cone is carried inward and no disturbance reaches the outer side; the horizon is where this first happens.
We can define a horizon as a closed curve , with outward normal , on which , with just inside. On the state carries differences inward as fast as they spread, and inside it faster. Where the cone of long disturbances does not contain the site it starts from, for some direction , that is , the circuit has an ergoregion.
The horizon belongs to the state of the circuit, not to its wiring: the wires cross in both directions and the logic is the same on both sides. On scales of a few sites a perturbation contains a dispersive part, which spreads at the speed of the wiring. This part is suppressed at long scales, but at short scales it may cross , yielding a small leak. If short disturbances travel faster than long ones, as they do in the circuit built below, the leak comes from just inside and falls off with depth: the horizon is sharp for long disturbances and leaky for short ones.
This is reminiscent of analogue horizons with superluminal dispersion [29, 30].
Design of a synchronous digital circuit as an analogue black hole
The construction of the previous section is formal: it does not guarantee that for a given map a horizon forms or that it is stable. These are properties of a running device, to be measured through an input and an output read from its state.
The design of the device corresponds to a choice of , the map , the static bits and the input . Here, with the shallow-water model of black holes in mind [5, 14], we choose them so that, on large scales, the circuit mimics a shallow layer of water flowing over a bed. As in shallow water, the wave speed depends on the depth, so the speed of sound of the circuit becomes a property of its state. We follow the construction of ref. [31]. Each word is a set of 25 floating-point numbers of 32 bits, one for each of a fixed set of directions reaching up to two sites, so that : the word reads the depth at the nearest neighbours, so that reads the words within sites. The map acts in two steps. First, each site forms the depth and the momentum , adds a push that sets the slope of the depth and of the bed at its neighbours, and relaxes its towards a fixed function of and . Then each moves to the site its direction points to. The relaxation and the moves conserve depth and momentum, up to the rounding of 32-bit arithmetic; the push only exchanges momentum between neighbouring sites, except for the slope of the bed, which acts as an external force. Water enters or leaves only at the reservoir, the drain, the outflow and the drive described below, and the absorbing layer removes momentum. On 32-bit words it is a Boolean map of the form of Eq. (1), whose long-wavelength limit is the shallow-water fluid, with and the velocity of the water.
The static bits lay out the analogue black hole: the reservoir (the sites beyond an outer radius, whose words are reset at every tick to a layer of fixed depth flowing inward with a swirl), the drain (the central sites, whose words are scaled down at every tick so that a fixed volume leaves; it forms the hole and is switched off once the sill is on), the sill (a ring of sites carrying a raised bed, over whose crest the inflow crosses the wave speed, which holds the horizon away from the centre), the outflow (the sites inside the sill, which copy at every tick the state of a ring just outside them, so that the water leaves freely), and the drive ring (the sites that receive the input). The input consists of two numbers per tick, added to words of the drive ring with angular patterns and : they set the amplitude and the sense of rotation of a wave of azimuthal number sent towards the hole. The controls use the same circuit: with the swirl of the reservoir set to zero the hole does not rotate, and reversing the sign of one input reverses the sense of the wave. The drain and the outflow are not of the form of Eq. (1): the drain uses the total depth of the central sites, and the outflow copies sites along a ray. In the held state only the outflow acts, inside the horizon, and its sites are left out of the causal-cone analysis (Appendix).
The design of the circuit as a shallow-water fluid allows one to predict where the horizon forms: where crosses zero, that is over the crest of the sill. The swirl makes the flow rotate, at an angular velocity on , with an ergoregion just outside.
Treating the circuit as a fluid is only formal. Whether a real circuit behaves as an analogue black hole is a question that only a measurement answers. In the following section we describe how this design is implemented in hardware, where it can be probed by electrical signals.
Hardware implementation
The circuit is implemented by translating the map into a hardware description, registers and logic, placed and routed on a field-programmable gate array [24]. In the configured chip the map exists as gates and wires, with no instructions in between.
A hardware implementation of Eq. (1) would require one copy of per site, far more than the chip holds. We therefore adopt the standard alternative for array (stencil) computations in hardware, to keep the words of the array in the board’s memory and apply a single copy of to the sites one after another [32, 33]. Every site is updated from the words of the previous tick, so the folded circuit computes the same bits as the unfolded one (Figure 9, the Appendix).

Figure 9. The device. (a) The bench: generator and oscilloscope, the two converter modules on the pin header, and the board, whose chip holds in its programmable logic next to a processor; the array lives in the board’s memory. Dashed: links used only to start runs and collect records. (b) One tick, to scale.
The processor on the same chip starts each tick, passes the two input numbers to the drive, forms the ring sums that give the output, and writes it. The state of the circuit evolves in the logic alone, so the processor affects neither the evolution of the circuit nor the nature of the measurement. The device is deterministic, and this allows the circuit to be checked at the three stages of its construction; we find that the chip computes as expected. Each protocol is given in the Appendix.
Measurements
The analogue black hole on the chip is probed by a two-channel signal generator, which supplies the input through analogue-to-digital converters, and by an oscilloscope, which records the observable written to a digital-to-analogue converter (Figure 9, the Appendix). The figures below show the voltages recorded on the oscilloscope together with the board’s digital record of the same quantity and the value of the 64-bit numerical twin [31]; the uncertainties of the oscilloscope readings are given in the Appendix.
Stability. Every 200 ticks the board reads back the state of the circuit, finds the radius of the horizon and writes it to the output (Figure 3, panel 1; Figure 4). Started from rest, the flow settles with its horizon at 103.2 sites; when the sill is switched on, the horizon moves to 114.29 sites and stays there to one part in for ticks, three hours on the oscilloscope. The non-rotating copy of the hole has its horizon at 114.72 sites.

Figure 3. The four measurements, schematically: a signal generator at the input of the chip, an oscilloscope at its output. In each panel the input signal is on the left, the circuit with the horizon (thick circle) in the middle and the oscilloscope trace on the right; blue marks where the measurement acts.
Figure 4. (a) Horizon radius against time: formation without the sill, the step when the sill is switched on, and the hold of ticks, for the rotating hole and its non-rotating copy; lines, oscilloscope; circles, board record. (b) The hold, zoomed: oscilloscope minus board record; dash-dotted, numerical twin. (c) Froude number, the ratio of inflow speed to wave speed ( per unit), against the probe radius set by the generator; grey dots, raw samples over twelve sweeps; black, binned; dashed, board record; thick grey, numerical twin.
The horizon. One generator channel sets a probe radius, and the board returns the ratio of the inflow speed to the wave speed at that radius (Figure 3, panel 2). A slow triangular sweep of the generator moves the probe in and out, and the oscilloscope, plotting output against input, draws the profile of the flow (Figure 4c). The board’s record of the same sweeps crosses one at 114.29, as does the horizon found from the full state.
The causal cone. The circuit is run twice from the held (steady) state with one bit flipped between the two runs, and the difference of the depth fields is recorded (Figure 5). The set of differing words spreads at the speed of the wiring; the bulk of the difference forms a much narrower cone (Figure 5a,Figure 5b). When the flipped bit sits outside the horizon, the cone of the difference drifts inward but opens on both sides, as in Figure 2b, and its outer edge advances at the speed (Figure 5d). When it sits inside, the whole cone is carried inward, as in Figure 2c, and its bulk does not reach the outside. The experiment is repeated with the bit flipped at different distances from the centre, and for each run we record the energy of the difference that reaches a circle well outside the horizon. Figure 5c, Figure 5d, Figure 5e show that the horizon is sharp for long waves and leaky for short ones, by an amount that falls exponentially with the depth of the flipped bit inside the horizon. The leak starts as short waves: in what first arrives beyond 135 sites, 83% of the power of above the rounding floor lies at , whereas inside 90% lies below it (Figure 8b). It also depends on direction, as the dispersion of the rule does (Figure 10): on the chip the decay length is sites along a lattice axis and along a diagonal. A flipped bit leaks far more than a smooth depth bump, because it excites every wavenumber of the lattice, including the short waves that climb out. The scale of the fall-off and its dependence on direction follow from the dispersion of the rule, with nothing fitted (Appendix).
Figure 5. (a) Difference of the depth fields of two runs that differ by one bit at , flipped inside and outside the horizon, on a logarithmic scale; white circle, the horizon; cyan, the acoustic front of the fluid limit. (b) The same against radius and tick, summed over angle; dashed, the characteristics of the fluid limit; dotted, the wiring limit . (c) Energy reaching beyond 135 sites against the depth of the flip below the horizon, relative to flips well outside; maximum over the run, with the rounding floor included, which sets the level of 0.04 for deep flips. (d) Outward speed of the difference against radius, with of the fluid limit. Points, medians over bins of four sites of the centroid speeds measured over 30 ticks in every run; bars, 1.25 times the standard error within the bin. (e) Energy beyond 135 sites at relative to the energy launched, with the rounding floor subtracted; triangles, points within 5% of the floor, drawn as upper limits at that level; filled, chip; open, software circuit; line, exponential fit over the three launch angles of the software circuit.

Figure 8. Post-processing of the causal-cone read-backs. (a) Fraction of the sites within of the flipped bit (, chip) whose words differ, against radius and tick (grey scale, 0 to 1); dotted, ; thin, , the reach of the push; dashed, the characteristics of the fluid limit; cyan, the contour at of the energy of the difference, the cone of Figure 5b. (b) Power spectrum of at for a flip at , inside the horizon () and beyond the detection circle (), with the rounding floor (flip of bit 13) in the outer annulus; solid, chip; dashed, software circuit. (c) Leak above rounding against depth for the three launch angles of the software circuit and for the chip at and , with the fitted decay lengths.
Figure 10. Dispersion of the rule on a still layer. (a) Phase speed over against wavenumber along an axis (filled) and the diagonal (open), at the depths of the horizon (0.68), of 120 sites (1.26) and of 150 sites (1.9); points, the rule run in 64-bit arithmetic; lines, the exact dispersion of the linearised rule (solid, axis; dashed, diagonal). (b) Decay rate per tick. Above the waves decay within a period.
The ringdown. A pulse from the generator perturbs the flow on a ring outside the horizon with a pattern of azimuthal number , and the circuit writes, at every tick, the amplitude of that pattern on a second ring (Figure 3, panel 3). For to 6 the frequencies of the tones on the oscilloscope agree with the board’s record and with the numerical twin to 1% or better (Figure 6). All tones decay, so the hole is stable. The longest-lived is the tone, at , which rings for nine periods on the oscilloscope. Table III in the Appendix lists the tones. The absorbing layer that starts at (Appendix) removes short waves but returns part of the long waves of the tones, so the hole rings in a box bounded by the layer: the tones are the quasinormal modes of hole and box together. Moving the inner edge of the layer in the twin between and 290 leaves the frequencies of to 6 unchanged within 2% while their decay rates fall steadily with the size of the box: the frequencies are those of the hole, and the decay rates measure the loss per round trip of the box, to the horizon, to the layer and to viscosity. The tone moves with the edge and is a mode of the box. It lies in the superradiant band, , but its turning radius ( sites) coincides with the outer layer, so it reaches the horizon only by tunnelling across the whole barrier; what the rotation returns is outweighed by the loss, a third of the energy per period, and the tone decays. In a twin whose boundary is five times farther the tones give way to tones that decay within one or two periods, the ringdown of the hole alone, which the array of the present board is too small to hold.
Figure 6. Ringdown on the oscilloscope. (a) Amplitude of the response after the pulse, rebuilt from the two passes (one per component), against the board record; the shaded window is the one used for the tones. (b) The tone, recorded for ticks. (c) Frequencies and damping rates of the tones, to 6, in units of : oscilloscope, board record and numerical twin.
| 1 | 0.835(2) | 0.028(1) | 0.834 | 0.029 | 0.840 | 0.027 |
| 2 | 1.958(9) | 0.229(9) | 1.955 | 0.229 | 1.962 | 0.230 |
| 3 | 1.590(5) | 0.137(9) | 1.596 | 0.128 | 1.602 | 0.132 |
| 4 | 1.439(13) | 0.095(10) | 1.426 | 0.098 | 1.429 | 0.100 |
| 5 | 1.695(1) | 0.134(4) | 1.695 | 0.134 | 1.700 | 0.135 |
| 6 | 1.550(13) | 0.128(18) | 1.561 | 0.111 | 1.561 | 0.112 |
Table III. Tones of the ringdown on ring 130, frequency and damping rate in units of : oscilloscope (both quadratures, from two passes; from the in-phase pass alone), the board’s digital record of the same runs, and the numerical twin, all with the same extraction. Errors in the last digits. The tones are those of the hole in the box bounded by the absorbing layer at 250, the frequencies set by the hole and the decay rates by the box; in twins with the boundary at 750 and 1500 they are absent (Appendix, Tones).
Superradiance. The two generator channels, in quadrature, drive a wave of azimuthal number that turns with the hole or against it, and the output is the net energy flux through four rings just outside the horizon (Figure 3, panel 4). The scan on the oscilloscope covers six frequencies of the co-rotating wave, two of the counter-rotating one, the non-rotating copy and the undriven circuit (Figure 7). The co-rotating flux is outward below and inward above it: the rotating hole gives energy to the wave inside the superradiant band and absorbs it beyond. The counter-rotating wave and the wave on the non-rotating copy are absorbed at every frequency, and the undriven output lies at the resolution of the instrument. Every point agrees with the board’s record and with the numerical twin (Figure 7b); the gain, the ratio of the outward flux to the flux absorbed by the non-rotating copy, is on the oscilloscope, the board and the twin alike. This is the gain with the absorbing layer; without it the twin gives 0.41 [31].
Figure 7. Superradiance on the oscilloscope. (a) Net energy flux through the rings two to eight sites outside the horizon against the drive frequency, for the wave turning with the hole (black), against it (orange), and on the non-rotating copy (grey); signed logarithmic scale, positive is outward. Filled symbols: oscilloscope; open circles: board record; crosses: numerical twin. The flux of the co-rotating wave changes sign at the band edge . (b) Oscilloscope against board record, point by point.
Conclusions
We have built an analogue black hole out of logic gates. Its horizon is a property not of the wiring but of the state: a closed curve in a lattice of 32-bit words along which the emergent flow carries disturbances inward exactly as fast as they spread outward. The hole holds that horizon to one part in over ticks; rings down, in the box set by its absorbing layer, with tones whose frequencies match those of a 64-bit numerical twin to 1% and whose damping rates match them within the measurement errors; gives energy to co-rotating waves below with a gain of 0.31 and absorbs them above it; and lets a single flipped bit inside the horizon escape only through the dispersive tail of the rule, by an amount that falls exponentially with the depth of the flip. That leak is the signature of the wavenumber-dependent horizon of a medium with superluminal dispersion [29, 30], here measured directly.
Three features set the digital hole apart from its fluid, atomic and optical counterparts. It is exactly deterministic: a run repeats bit for bit, so two copies that differ by a single flipped bit can be subtracted, and the causal structure of the horizon is measured rather than inferred. Its short-distance physics is fixed by the update rule: the scale and the anisotropy of the measured leak agree with the dispersion relation computed from it. And it is reconfigurable: geometry, rotation and dissipation are changed by rewriting static bits, the rule itself by a few hundred lines of logic.
Nothing in the construction is specific to two dimensions or to shallow water. A three-dimensional hole costs only memory and logic. A rule of a few bits per site would let every site carry its own logic and tick at the clock, with intrinsic fluctuations and the throughput to average them; whether a transonic, rotating flow exists in such a rule is an open question. Digital circuits thus form a reconfigurable class of gravitational analogues with the analogue black hole becoming a cellular automaton.
Acknowledgments
This work was conducted using the Miyabi Supercomputing System at the Joint Center for Advanced High Performance Computing (JCAHPC).
Data availability and AI disclosure. All data used in this paper are available in the Zenodo repository 10.5281/zenodo.23967575. AI tools were used to assist with code for the simulation sweeps, data analysis and figure plotting, and with language editing; all figures are plotted directly from the data in the repository.
Appendix
The rule
Each site of a array holds 25 single-precision numbers , one per velocity , , with weights , , , , and . One tick maps to :
; , with , the centred nearest-neighbour difference, the bed and the absorbing layer.
.
with , , . With , the forcing term as defined in ref. [31], the moments of relax at rate 0 (orders 0, 1), (order 2; viscosity ) and 0.6 (orders ).
Each moves to .
The hole
The static word lays out the hole. Beyond (reservoir) each word is reset every tick to the equilibrium of depth 2 and , , ( for the non-rotating copy). The sill is . The hole is formed with a drain: for the words are multiplied by , with rising linearly from 0 at to 1 at and over the central sites on the previous tick, so that the volume leaves every tick. Once the sill is on the drain is off, and the water leaves through an outflow: inside every word is replaced by the equilibrium of the site at on the same ray, captured on the previous tick, with radial velocity raised to at least . Neither is local in the sense of Eq. (1): the drain uses the sum , fed back by the processor with one tick of lag, and the outflow copies a site up to 100 sites away, also one tick late. In the held state only the outflow acts; the ring it copies lies 13 sites inside the horizon, in the supercritical layer. The drive acts on a ring at 170: , and set every tick from the two inputs. In the ringdown and superradiance runs an absorbing layer , beyond , damps what reaches it: short waves are absorbed, the long waves of the tones are partly returned. The hole is formed from rest in 32,000 ticks with the drain and without the sill (horizon at 103.2), the drain is then switched off and the sill and the outflow are switched on for 12,000 ticks, and the state is held for a further ticks (1.1 ticks after the sill is switched on, Figure 4); this held state starts every measurement.
The device
Figure 9 shows the signal chain and one tick.
Signals. A two-channel generator and oscilloscope (Digilent Analog Discovery 3) form the instrument. The generator channels feed a two-channel 12-bit ADC (Pmod AD1, 0–3.3 V), sampled once per tick to give ; a 16-bit DAC (Pmod DA3, 0–2.5 V), written once per tick while the observable, feeds oscilloscope channel 1, and channel 2 records one generator channel. Both converters sit on the header of the board, are driven by the processor, and share one ground with the instrument. A workstation programs the instrument over USB and starts runs over Ethernet; neither link carries a signal during a measurement. With no circuit between them, the chain copies the generator to the output within one input step (0.8 mV). The board runs at 0.12 s per tick (0.2 s in the ringdown runs; the hold runs unpaced, at about 0.10 s per tick); one turn of the horizon takes about 2,800 ticks, so every signal lies far below one hertz.
Board. On the KV260 board the Zynq UltraScale+ XCK26 combines the programmable logic, which holds , with a processor. The array ( sites of 32 words of 32 bits, 25 of them populations, 75 MB) and the static word (16 per site, 38 MB) are kept in DDR4 and streamed through the logic once per tick; at 100 MHz and 16.5 clock cycles per site one pass takes 96 ms. The device is thus the chip together with the memory and converters of its board.
One tick. The processor reads the two inputs, hands to the logic and starts the tick. The logic passes all $5.9\times10^5$ sites through in 96 ms and copies out the flagged rings. From these the processor forms the observable ( on a ring, or the energy flux through rings outside the horizon), writes it to the DAC and waits for the end of the period. The output holds one value per tick and answers the input of the previous tick. The processor returns to $\Phi only the total depth of the central sites and, in the ringdown runs, the pulse. In the hold and profile runs it also reads the array back every 200 ticks to form averages over angle; in the ringdown and superradiance runs it never does.
Verification
The rule is a C++ source synthesised with AMD Vitis and Vivado 2026.1. It was checked against a PyTorch reference at three stages: the source compiled as a program, the synthesised gates in cycle-accurate co-simulation on a array, and the configured chip read back digitally. All three agree with the reference to the last place of single precision ( to in the after one tick, in the ring captures). Chip and compiled source agree to about one unit in the last place per tick, not bit for bit; the chip repeats itself bit for bit. The circuit closes timing at 100 MHz and uses 72,372 of the chip’s 117,000 look-up tables, 86,482 registers and 252 of its 1,248 DSP multipliers.
Numerical twin. The rule also exists as software: the shallow-water lattice-Boltzmann model of ref. [31], run in 64-bit arithmetic on graphics processors. It shares with the device the rule and the layout of the hole, and nothing else. Each measured quantity is available three times: as a voltage on the oscilloscope, as the digital record of the board, and as the value of the twin (Table I).
Measurement protocols
Every run begins with 10 ticks at 0 V on the output, which marks tick 0 on the oscilloscope; the board logs what it read and wrote at every tick.
Hold. Every 200 ticks the board reads back the state, averages over angle in rings one site wide, finds by linear interpolation where the inflow speed equals , and writes .
Profile. Every 200 ticks the processor forms the profile of from a read-back; at every tick it reads the probe voltage on the second input, sets , and writes at 1 V per unit.
Ringdown. At a rising edge of the first input above 0.5 V the processor multiplies the words by . At every tick the logic captures the ring at 130 and the processor writes , ; a second pass writes .
Superradiance. The two inputs, calibrated against the generator, give and ; the drive is , ramped over 4000 ticks, with . After 8000 further ticks, 16,000 ticks are recorded. At every tick the logic captures the rings ; the processor forms on each and writes , the sign alternating from tick to tick.
Error analysis
Three effects limit the oscilloscope readings: the resolution of the converters and the noise of the instrument; the offset of the instrument, which is measured and subtracted wherever the output passes through a known level and cancels in the flux; and calibration, of the probe input and of the output chain. Table II gives the budget.
Converters and noise
The output step is 38 μV (16 bits over 2.5 V), the input step 0.8 mV (12 bits over 3.3 V). The oscilloscope carries 0.5 mV rms of noise. On the scales used, 1 mV is 0.01 sites in the hold, 0.001 in (0.006 sites at the crossing) in the profile, and in the ringdown and flux outputs, to being the gain chosen for a swing of about 0.5 V ( for the undriven output).
Offset of the instrument
The offset wanders by about 1 mV within an hour and changes from run to run; after a re-zero against levels set by the board the oscilloscope read 1.0003 times the output plus mV. Where a run contains a stretch of known output (the ticks before the pulse in the ringdown runs, before the drive in the flux runs) the offset is measured there and subtracted. In the flux runs the sign alternates on alternate ticks, and demodulating with , the output and its sign, removes any offset that varies slowly compared with two ticks.
Calibration
Probe input. At the turning points of the profile sweep, where the one-tick lag does not matter, the radii read by the board and by the oscilloscope agree at the inner end (90.59 on both) and differ by 0.22 sites at the outer end: a 0.3% difference in gain, consistent with the calibration of the input converter (0.995). At the horizon the correction is −0.075 sites; choosing the turning points from the 0.2, 0.5 or 1% extremes of the sweep changes it by about 0.01.
Output chain. Point by point, the flux on the oscilloscope differs from the board’s record by −0.16 to +0.07% (mean −0.07%); 0.16% is taken as the calibration uncertainty. It cancels in , a ratio of two fluxes.
Horizon from the profile
The probe sweeps 90 to 160 sites and back twelve times, and the crossing of is found in each sweep by a straight-line fit. The one-tick lag shifts the crossing outward on outward sweeps and inward on inward ones and cancels in their mean; with the probe-gain correction and the uncertainty of the offset the result is 114.35 ± 0.03 sites, against 114.29 in the board’s record.
Tones
Box. The same hole was struck in the same way in twins of and words, with the free disc extended to and 1500 and the damping layer moved to its edge. The horizon radius and are unchanged to 0.3%, and none of the tones of Table III appears; the pencil finds instead poles at to decaying at to , within one to two periods, with amplitudes ten times smaller. In the twin these poles are stable at every once the undriven response of the lattice, recorded without a pulse and at as large as the pulse’s, is subtracted, and they are unchanged when the pulse is doubled; the twin agrees where its extraction is stable (). Radius of the layer. In the twin the absorbing layer, which starts at in the table, was moved to 150, 200, 225 and 290. From 225 to 290 the frequencies of to 6 stay within 2% of the table while the decay rates fall monotonically with the radius, by a factor of 1.3 to 2.5; at 150 the tones are replaced by the damped family of the large boxes. The frequency follows the radius (0.99, 0.92, 0.84, 0.72 at 200, 225, 250, 290): it is a mode of the box. The tones of the table are therefore those of the hole in the box bounded by the layer, the frequencies set by the hole and the decay rates by the box. Each tone is extracted from ticks 1200 to 3900 after the pulse by a matrix pencil at model orders 4 to 10, keeping poles that recur at five or more orders; the same extraction is applied to the oscilloscope, the board’s record and the numerical twin, all on ring 130. The error combines the scatter of the pole over the orders with the difference between oscilloscope and board record, which measures how instrument noise moves the extraction. The tone (30,600 ticks) is extracted from the in-phase pass alone by the cluster method, the scatter over windows being the error. The offset is taken from the ticks before the pulse, and poles with per tick, which describe its drift, are excluded. On the wire the frequencies agree with the numerical twin within 0.012 and the damping rates within 0.016 (Table III).
Energy flux and gain
Each run lasts 28,000 ticks with chopped output. The flux is the mean of the demodulated ticks after the transient; the statistical error is the scatter of the means of four equal windows divided by two, 0.005 to 0.17%. Without drive the output is on the oscilloscope and in the board’s record, the resolution at the gain used (). The gain , the co-rotating flux over that of the non-rotating copy, at is 0.3107 ± 0.0002 on the oscilloscope and 0.3110 in the board’s record. The error is statistical. The flux is outward in the band with every treatment of the outer boundary, and the size of the gain, a ratio to the flux absorbed by the non-rotating copy, varies with that treatment between 0.31 and 0.41 over the protocols of ref. [31].
The causal cone and the dispersion of the rule
Protocol. From the held state the chip runs 400 ticks; every two ticks the board reads back, for a 401 × 401 window around the centre, the depth and a 32-bit hash of the 25 words of each site. The run is repeated with bit 20 of flipped at at the site nearest a chosen radius (8% of that number), keeping per radial bin of one site the sum of , the largest and the number of sites whose hashes differ, and the full map of at 13 ticks. A third run without the flip reproduced the first at every site and tick. On the chip the flips were placed on one ray at 11 radii from 104.5 to 124.5; on the software circuit at 15 radii and three angles (, , ). The level of rounding is set by a flip of bit 13, whose effect is lost after about 60 ticks. The outward speed is that of the centroid of the energy of the difference beyond the launch radius, followed over 30 ticks.
Wiring cone, wave-number content and anisotropy. Figure 8 shows three further quantities formed from the same read-backs. (a) The number of sites per radius at which any of the 25 words differs, from the hashes, fills the square reached by the wiring: within of the flipped bit over 90% of the words differ after 20 ticks, while the energy of the difference stays within the sound cone. The words differ out to for the first ticks, because the push reads the depth at the nearest neighbours before the populations move by up to two sites, so that the reach of is ; the differences carried by the push alone fall below rounding within about 40 ticks, after which the set grows at two sites per tick. The sites are left out: inside the words are copies of the ring at 101.5 (the outflow), so a difference that reaches that ring appears along the whole ray inside it on the next tick. (b) Inside the horizon 90% of the power of lies at ; beyond 135 sites, at , the rounding floor accounts for 46% of the power and 83% of the excess lies at , where the waves of Figure 10 outrun the inflow. (c) The decay length of the leak, fitted as in Figure 5e for each launch angle, is sites at , at and at on the software circuit ( pooled, the fit drawn in Figure 5e), and on the chip at and at (seven flips at radii 104 to 120, run after a clean restart of the board; at those angles the chip and software circuit give the same escaped energy against depth to 1–3%). The leak does not depend on the kick: on the software circuit, flipping bit 18, 20 or 22 of , or bit 22 of an outward or inward population (, ) instead, gives the same escaped energy against depth and decay lengths of , and sites. Nor does it depend on the measure of energy: with the momentum read back as well, the second-order wave energy of the difference is about a tenth of at launch, the rest being vortical, and the outward flux of this energy through the circle at 135 sites, accumulated in time, saturates within 100–170 ticks and stays flat to 400: its final value falls off with depth over sites against for the snapshot of Figure 5e on the same runs. The cone runs have no absorbing layer; the reservoir at 300 sites is the outer boundary, acts too far for anything to return within 400 ticks, and the circle at 135 sites lies outside the ergosurface at 121.
Dispersion and the leak. Linearising one tick about a uniform layer of depth and velocity gives, for each wavevector , a 25 25 matrix whose eigenvalues hold the frequencies and decay rates of the modes; two of them are sound waves (lines in Figure 10, for a still layer). Runs of the rule in 64-bit arithmetic on a still layer in a periodic box (points) follow the linearised rule to of the perturbation. Long waves travel at , shorter ones faster, to leading order as : the hydrostatic force replaces by only for disturbances long compared with the two-site difference that computes it. Waves shorter than about ten sites () travel at 0.3–0.5 sites per tick, up to 2.5 times at the depth of the horizon, and decay within 4–30 ticks.
The fall-off of the leak with depth follows from this dispersion, with nothing fitted. On the held flow the linearised rule gives the frequency and the decay rate of the outgoing sound wave along a ray. Inside the long waves have and are carried inward; a component launched at depth climbs out only if its phase speed exceeds the inflow there, that is above a threshold , which rises from 0.4 at to 0.7 at sites along an axis, and from 0.5 to 0.9 along a diagonal. As it climbs its frequency is conserved, its wavenumber falls and it decays at the rate , so the escaping fraction of the flat spectrum of a flipped bit is , with and the inner integral taken along the ray to 135 sites. The short waves found outside at tick 40 (Figure 8b) are the faster and more strongly damped part, which arrives first. The computed leak falls off over 1.9 to 2.2 sites along an axis and 1.4 to 1.6 along a diagonal, for spectra flat in one and in two dimensions. The measured lengths are 2.5 ± 0.3 and 1.6 ± 0.3 on the software circuit and 3.1 ± 0.6 and 1.9 ± 0.2 on the chip: the computation gives the scale of the leak and its dependence on direction, with lengths shorter than the measured ones by up to about two standard errors.
From a depth of eight sites a flipped bit delivers about of what it delivers from outside, against at most for smooth depth bumps 1.5 to 3 sites wide [31], which excite much less of the short waves. The leak responds only weakly to the relaxation rate of the moments of order : varying it from 0.3 to 1.4 (the published value is 0.6) moves the horizon by 0.04 sites and the decay length from 2.7 ± 0.5 to 1.8 ± 0.3 sites, because the damping at is set by the viscosity.
| quantity | oscilloscope | board record | numerical twin |
| horizon radius, rotating hole (sites) | |||
| horizon radius, non-rotating copy (sites) | |||
| horizon from the flow profile (sites) | |||
| tones, to (frequency) | agree to 1% or better (Table III) | ||
| energy flux, co-rotating, | |||
| energy flux, non-rotating copy, | |||
| gain at | |||
| sign change of the co-rotating flux | between and |
Table 1. The same quantities on the oscilloscope, in the digital record of the board, and in the numerical twin. Oscilloscope values carry their statistical error and, for the held radii, the systematic of the instrument offset (Table II); fluxes in the last digits carry the statistical error only, with a calibration uncertainty of 0.16%.
| source | size | horizon hold | horizon profile | tones | energy flux |
| output converter, 16 bit | – | ||||
| oscilloscope noise | rms | in extraction | |||
| input converter, 12 bit | – | – | – | ||
| offset of the instrument | to | <br> | removed | cancels | |
| probe gain | – | (corrected) | – | – | |
| output gain | – | – | – | ||
| lag of one tick | – | (cancels) | – | – | |
| extraction | – | – | – | – | – |
Table 2. Uncertainty budget of the oscilloscope readings. Horizon radii in sites; tones in units of ; energy flux as a relative error. “Corrected”: a measured shift that was removed; the number is what remains.
References
References
- [1]W. G. Unruh, Phys. Rev. Lett. 46, 1351 (1981).
- [2]S. W. Hawking, Nature 248, 30 (1974).
- [3]Ya. B. Zel’dovich, JETP Lett. 14, 180 (1971).
- [4]R. Brito, V. Cardoso, and P. Pani, Superradiance, Lect. Notes Phys. 906 (Springer, 2015).
- [5]S. Weinfurtner, E. W. Tedford, M. C. J. Penrice, W. G. Unruh, and G. A. Lawrence, Phys. Rev. Lett. 106, 021302 (2011).
- [6]L.-P. Euvé, F. Michel, R. Parentani, T. G. Philbin, and G. Rousseaux, Phys. Rev. Lett. 117, 121301 (2016).
- [7]T. Torres, S. Patrick, A. Coutant, M. Richartz, E. W. Tedford, and S. Weinfurtner, Nat. Phys. 13, 833 (2017).
- [8]T. Torres, S. Patrick, M. Richartz, and S. Weinfurtner, Phys. Rev. Lett. 125, 011301 (2020).
- [9]J. Steinhauer, Nat. Phys. 12, 959 (2016).
- [10]J. R. Muñoz de Nova, K. Golubkov, V. I. Kolobov, and J. Steinhauer, Nature 569, 688 (2019).
- [11]T. G. Philbin, C. Kuklewicz, S. Robertson, S. Hill, F. König, and U. Leonhardt, Science 319, 1367 (2008).
- [12]J. Drori, Y. Rosenberg, D. Bermudez, Y. Silberberg, and U. Leonhardt, Phys. Rev. Lett. 122, 010404 (2019).
- [13]P. Švančara et al., Nature 628, 66 (2024).
- [14]R. Schützhold and W. G. Unruh, Phys. Rev. Lett. 95, 031301 (2005).
- [15]P. D. Nation, M. P. Blencowe, A. J. Rimberg, and E. Buks, Phys. Rev. Lett. 103, 087004 (2009).
- [16]Y.-H. Shi et al., Nat. Commun. 14, 3263 (2023).
- [17]T. Jacobson, Phys. Rev. D 44, 1731 (1991).
- [18]W. G. Unruh, Phys. Rev. D 51, 2827 (1995).
- [19]M. Visser, Class. Quantum Grav. 15, 1767 (1998).
- [20]L. J. Garay, J. R. Anglin, J. I. Cirac, and P. Zoller, Phys. Rev. Lett. 85, 4643 (2000).
- [21]U. R. Fischer and M. Visser, Phys. Rev. Lett. 88, 110201 (2002).
- [22]R. Schützhold and W. G. Unruh, Phys. Rev. D 66, 044019 (2002).
- [23]C. Barceló, S. Liberati, and M. Visser, Living Rev. Relativity 14, 3 (2011).
- [24]D. M. Harris and S. L. Harris, Digital Design and Computer Architecture, 2nd ed. (Morgan Kaufmann, Waltham, 2013), Chap. 3.
- [25]T. Toffoli and N. Margolus, Cellular Automata Machines (MIT Press, 1987).
- [26]E. H. Lieb and D. W. Robinson, Commun. Math. Phys. 28, 251 (1972).
- [27]M. B. Hastings and T. Koma, Commun. Math. Phys. 265, 781 (2006).
- [28]S. Succi, The Lattice Boltzmann Equation for Fluid Dynamics and Beyond (Oxford University Press, 2001).
- [29]S. Corley and T. Jacobson, Phys. Rev. D 54, 1568 (1996).
- [30]J. Macher and R. Parentani, Phys. Rev. D 79, 124008 (2009).
- [31]A. Flachi, Shallow-water black holes and lattice superradiance, arXiv preprint, submitted concurrently (2026).arxiv.org/abs/2611.10539
- [32]H. T. Kung, Why systolic architectures?, Computer 15, 37 (1982).
- [33]K. Sano, O. Pell, W. Luk, and S. Yamamoto, in Proc. Int. Conf. Field-Programmable Technology (IEEE, 2007), pp. 233–236.




