Time-Multiplexed Ising Machine

• Time-multiplexed Ising machine is a hardware architecture that encodes multiple spin states as distinct time slots within a single physical loop, enabling efficient Ising Hamiltonian minimization.
• Diverse implementations—from optical to acoustic and microwave systems—utilize delay-based feedback and measurement techniques to realize programmable couplings and effective spin dynamics.
• Practical insights include careful synchronization, delay management, and controlled bifurcation, with demonstrated applications in MAX-CUT, Sudoku, number partitioning, and supply-chain optimization.

A time-multiplexed Ising machine is a hardware optimizer that represents many effective spins as distinct time slots circulating in a single physical loop or delay element, rather than as spatially separate oscillators or probabilistic bits. Across optical, microwave, acoustic, spintronic, and mixed-signal realizations, the common principle is that binary variables $s_i \in \{\pm 1\}$ are encoded in a phase, amplitude, or stochastic state associated with each time bin, while couplings $J_{ij}$ and fields $h_i$ are applied through delayed reinjection, measurement-feedback, digital threshold control, or weighted accumulation so that the hardware relaxes toward low-energy configurations of an Ising or QUBO objective (Takata et al., 2016, Ovcharov et al., 2024, Love et al., 15 Jun 2026).

1. Conceptual basis and Ising formulation

The unifying objective is minimization of an Ising Hamiltonian. Several realizations in the literature use the form

$$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$

with $s_i \in \{+1,-1\}$, although some papers note an equivalent sign convention differing by a global sign (Takata et al., 2016, Litvinenko et al., 2023, Vadde et al., 2 Jul 2026). In time-multiplexed machines, these spins are not instantiated as separate hardware nodes; instead, one loop carries a pulse train or sequence of logical updates, and each temporal slot acts as one spin (Marandi et al., 2014, Litvinenko et al., 2022).

This architecture is used to solve combinatorial optimization problems that admit Ising or QUBO encodings. MAX-CUT is the canonical benchmark. In optical, surface-acoustic-wave, spin-wave, and bulk-acoustic-wave implementations, antiferromagnetic couplings encode graph edges so that minimizing $H$ is equivalent to maximizing the cut (Takata et al., 2016, Litvinenko et al., 2023, Litvinenko et al., 2022, Vadde et al., 2 Jul 2026). Other demonstrated mappings include number partitioning with $J_{ij}=a_i a_j$ and $h_i=0$, Sudoku with one-hot penalty Hamiltonians on 729 spins, Boolean satisfiability via quadratized 2-SAT and 3-SAT encodings, and multi-period supply-chain allocation via a block-diagonal QUBO with slack-bit capacity constraints (Vadde et al., 2 Jul 2026, Love et al., 15 Jun 2026, Ubale et al., 24 Oct 2025).

A central distinction from spatially multiplexed Ising hardware is that time multiplexing replaces $N$ physical spin elements with one shared physical loop and $N$ temporal degrees of freedom. This allows dense logical connectivity without duplicating the core nonlinear element for every spin, although the implementation burden shifts to delay management, memory, feedback bandwidth, and slot-to-slot synchronization (Takata et al., 2016, Litvinenko et al., 2023, Love et al., 15 Jun 2026).

2. Time-multiplexing principle and spin encodings

The basic timing relation is that the number of supported spins is the ratio of total loop delay to slot duration. In optical and wave-based systems, the loop stores a train of pulses separated so that adjacent slots do not overlap. In delay-line oscillator models this is written as

$J_{ij}$0

with $J_{ij}$1 the round-trip delay and $J_{ij}$2 the slot width (Ovcharov et al., 2024). In early optical coherent Ising machines, a single ring cavity hosted several femtosecond pulses, with $J_{ij}$3 (Marandi et al., 2014). In the 16-bit coherent Ising machine, 16 pulses spaced by about $J_{ij}$4 ns circulated in a cavity with repetition frequency $J_{ij}$5 GHz (Takata et al., 2016). In the 50-spin surface acoustic wave machine, a 12 $J_{ij}$6s circulation held 58 RF pulses, of which 50 served as spins and 8 provided measurement/feedback latency (Litvinenko et al., 2023). In the 2,048-spin bulk acoustic wave machine, two serially connected 707 $J_{ij}$7s delay lines supported $J_{ij}$8 time slots, with $J_{ij}$9 used as spins (Vadde et al., 2 Jul 2026). In the 64-spin optoelectronic machine, a 64 ns optical frame at 1 GHz encoded 64 spins in 64 time slots (Love et al., 15 Jun 2026).

The spin state itself is encoded differently across platforms, but typically by a bistable phase variable. In degenerate optical parametric oscillator systems, the two stable phases are $h_i$0 and $h_i$1, mapped to $h_i$2 and $h_i$3 (Takata et al., 2016, Marandi et al., 2014). In spin-wave and surface-acoustic-wave machines, a phase-sensitive amplifier binarizes RF pulse phases to $h_i$4 or $h_i$5, again realizing $h_i$6 (Litvinenko et al., 2022, González et al., 2023, Litvinenko et al., 2023). In delay-line oscillator models under subharmonic injection locking, a convenient mapping is

$h_i$7

where $h_i$8 is the phase of the $h_i$9-th pulse relative to the reference at $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$0 (Ovcharov et al., 2024). The 64-spin CMOS-assisted machine uses amplitude-coded optical pulses and a saturating nonlinearity that bifurcates spins to $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$1, so it is amplitude-based rather than phase-coherent (Love et al., 15 Jun 2026).

A different time-multiplexed interpretation appears in the single-STNO Ising computer. There, one tunable true random number generator based on a spin torque nano-oscillator is reused across time slots, so one physical stochastic source acts as multiple logical p-bits. Each 100 ns window produces one bit, and logical spins are realized serially rather than as co-propagating wave packets (Zhang et al., 2022). This suggests that “time-multiplexed Ising machine” is a broader architectural category than pulse-in-loop photonics alone.

3. Coupling mechanisms and dynamical evolution

Time-multiplexed Ising machines differ most strongly in how they implement $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$2 and $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$3. In all-optical coherent Ising machines, couplings are realized by reinjecting delayed replicas of pulses into the main cavity. Constructive mutual injection implements ferromagnetic couplings, while destructive mutual injection implements antiferromagnetic couplings (Takata et al., 2016). The earlier four-spin optical network implemented couplings by tapping about $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$4 of intracavity power and reinjecting after integer-multiple pulse delays; the injection phase set the sign of $$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$5 (Marandi et al., 2014). Spin-wave and surface-acoustic-wave machines also use delay-based reinjection, but the feedback may be computed digitally and then reinjected with programmable phase and amplitude (Litvinenko et al., 2022, Litvinenko et al., 2023).

Measurement-feedback architectures replace direct optical or RF interference by explicit readout and digitally synthesized feedback. In the SAW machine, the FPGA computes

$$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$6

then applies sign through a two-state phase shifter and magnitude through an attenuator before injecting the feedback pulse into the correct time slot (Litvinenko et al., 2023). The 2,048-spin bulk acoustic wave machine similarly measures each pulse, digitizes it, computes

$$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$7

and reinjects an analog correction, with the effective discrete-time update captured by

$$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$8

(Vadde et al., 2 Jul 2026). The 64-spin optoelectronic solver implements an explicitly mixed-signal gradient step,

$$H(s) = -\sum_{i<j} J_{ij}\, s_i s_j - \sum_i h_i\, s_i,$$9

where $s_i \in \{+1,-1\}$0 is a saturating nonlinearity and $s_i \in \{+1,-1\}$1 acts as a learning rate (Love et al., 15 Jun 2026).

In oscillator-based analytical models, the physical evolution is described by delay-differential equations. One formulation uses

$s_i \in \{+1,-1\}$2

with

$s_i \in \{+1,-1\}$3

(Ovcharov et al., 2024). In that model, the gain-compression coefficient $s_i \in \{+1,-1\}$4 and amplitude-dependent frequency nonlinearity $s_i \in \{+1,-1\}$5 determine a synchronization region, and the highest global minimum probability occurs near the edge of synchronization rather than deep inside it (Ovcharov et al., 2024).

The single-STNO system replaces field coupling by a threshold update rule. Its incremental coupling rule is

$s_i \in \{+1,-1\}$6

with thresholds clamped to $s_i \in \{+1,-1\}$7. This introduces memory in threshold space and avoids the “oscillatory wrong state” that arises under a simpler reset-to-$s_i \in \{+1,-1\}$8 rule (Zhang et al., 2022).

A recurrent dynamical theme is that good performance is obtained not by maximal forcing, but by controlled bifurcation. Optical coherent Ising machines report that gradual pumping improves success relative to abrupt pumping (Takata et al., 2016). Surface-acoustic-wave and bulk-acoustic-wave systems use attenuator-controlled start-up schedules in which effective coupling is strongest when pulse amplitudes are small and decreases as amplitudes saturate (Litvinenko et al., 2023, Vadde et al., 2 Jul 2026). The numerical delay-line oscillator study likewise reports best global minimum probability near a synchronization boundary, not far inside the locked regime (Ovcharov et al., 2024). This suggests a common operational principle across otherwise different hardware.

4. Physical implementations across platforms

The literature now spans several distinct physical realizations of time-multiplexed Ising hardware.

Platform	Spin encoding and coupling	Reported scale or metric
Optical coherent Ising machine	DOPO pulse phases $s_i \in \{+1,-1\}$9; optical delay-line mutual injection	4 spins with no computational error detected in 1000 runs (Marandi et al., 2014); 16 spins with 99.8% success on a cubic graph (Takata et al., 2016)
Spin-wave and surface-acoustic-wave machines	RF pulse phases binarized by PSA; delay-line or FPGA-mediated feedback	8-spin SWIM solved in less than 4 $H$0s consuming 7 $H$1J (Litvinenko et al., 2022); 50-spin SAWIM solved MAX-CUT in less than 340 $H$2s consuming 0.62 mJ (Litvinenko et al., 2023)
Acoustic and optoelectronic large-scale systems	Measurement-feedback with FPGA or on-chip MAC fabric	2,048-spin BAWIM for MAX-CUT, number partitioning, and Sudoku (Vadde et al., 2 Jul 2026); 64 all-to-all connected spins at 1 GHz with nanosecond annealing (Love et al., 15 Jun 2026)

The optical coherent Ising machine established the archetype. Marandi et al. implemented four spins as four femtosecond pulses in one ring cavity and used mutual injections to program a frustrated four-node MAX-CUT instance, reporting that in 1000 runs of the machine no computational error was detected (Marandi et al., 2014). The 16-bit extension used time-division-multiplexed femtosecond degenerate optical parametric oscillators and programmable delay lines to realize one-dimensional rings and a 16-vertex cubic graph; under well-locked conditions, only two failures were observed in 1000 cubic-graph runs, for an empirical success rate of 99.8% (Takata et al., 2016).

Spin-wave Ising machines transferred the same architectural logic to magnonics. In the 8-spin SWIM, 34–68 ns long 3.125 GHz spin-wave pulses in a 7 mm YIG waveguide encoded spins, and delayed weighted reinjection implemented $H$3. The system solved 4- and 8-spin MAX-CUT instances in less than 4 $H$4s while consuming only 7 $H$5J (Litvinenko et al., 2022). A later SWIM study added a hardware global Zeeman term by injecting a phase-locked RF tone, allowing ferromagnetic ordering to be induced against antiferromagnetic pairwise coupling in a 4-spin ring (González et al., 2023).

Surface-acoustic-wave and bulk-acoustic-wave systems exploit the same time-slot logic but with solid-state acoustic delay lines. The 50-spin SAWIM used a 50 mm long LiNbO$H$6 delay line, all-to-all FPGA reprogrammability, and a parametric phase-sensitive amplifier. It solved arbitrary 50-spin MAX-CUT problems in less than 340 $H$7s consuming only 0.62 mJ, with close to 3000 solutions per second and a figure of merit of 1610 solutions/W/s (Litvinenko et al., 2023). The 2,048-spin BAWIM used two serially connected 20.5 MHz quartz delay lines supporting all-to-all coupling with 15-bit resolution, found approximate MAX-CUT solutions in 341 ms, and also demonstrated number partitioning and Sudoku (Vadde et al., 2 Jul 2026).

The most integrated optoelectronic realization in the dataset is the 64-spin pulsed TDM solver. A 65 nm CMOS chip stores the full spin vector, applies 4-bit signed $H$8 and $H$9 weights from on-chip memories, accumulates the weighted sums in current mode, and closes the loop with a compact optical fiber delay. It operates at 1 GHz and benchmarks 2-SAT, 3-SAT, and dense MaxCut (Love et al., 15 Jun 2026).

Finally, the STNO implementation shows that time multiplexing is not restricted to wave-packet storage. One tunable true random number generator, validated by all 11 NIST SP 800-22 tests, can be reused acting as a p-bit array by time division multiplexing, enabling a whole Ising computer to be implemented by one single STNO (Zhang et al., 2022).

5. Performance characteristics and demonstrated applications

Reported performance varies strongly by platform, problem class, and success metric. The 16-spin optical coherent Ising machine experimentally gave more than 99.6% of success rates for one-dimensional Ising ring and NP-hard instances, and specifically 99.8% on a 16-vertex cubic graph under well-locked conditions (Takata et al., 2016). The four-spin predecessor reported no computational error detected in 1000 runs on the smallest NP-hard Ising instance (Marandi et al., 2014).

Wave-based microwave and acoustic systems emphasize time-to-solution and energy. The 8-spin SWIM solved MAX-CUT in less than 4 $J_{ij}=a_i a_j$0s while consuming only 7 $J_{ij}=a_i a_j$1J (Litvinenko et al., 2022). The 50-spin SAWIM solved arbitrary 50-spin MAX-CUT problems in less than 340 $J_{ij}=a_i a_j$2s, using 1.82 W total active power and about 0.62 mJ per solution; for one representative graph, the exact-solution probability reached 100% at a global coupling attenuation of $J_{ij}=a_i a_j$3 dB (Litvinenko et al., 2023). The 64-spin optoelectronic machine reported a mean annealing time of about 700 ns and, for 3-SAT clause-to-variable ratios $J_{ij}=a_i a_j$4, $J_{ij}=a_i a_j$5, and $J_{ij}=a_i a_j$6, time-to-solution values of 7.4 $J_{ij}=a_i a_j$7s, 13.8 $J_{ij}=a_i a_j$8s, and 159.2 $J_{ij}=a_i a_j$9s for 99% success at 100% clause accuracy, with corresponding energies to solution of 2.9 $h_i=0$0J, 5.5 $h_i=0$1J, and 63.7 $h_i=0$2J (Love et al., 15 Jun 2026).

Large-scale acoustic hardware broadens the application range beyond graph cuts. The 2,048-spin BAW system demonstrated approximate MAX-CUT, number partitioning, and Sudoku. It reached 90% of the heated-ballistic simulated bifurcation best energy on a 2,048-node, 10%-density MAX-CUT instance in about 341 ms, achieved 99% or better of HbSB’s best MAX-CUT score across runs, outperformed HbSB on number partitioning at the 99.9%-approximate level for most sizes from 32 to 2,048, and converged to fully correct Sudoku solutions on 729 spins (Vadde et al., 2 Jul 2026).

The supply-chain QUBO study demonstrates a time-multiplexed coherent Ising machine as an application accelerator rather than a physics benchmark. On a problem with approximately 4,100 variables and over one million quadratic terms, Quanfluence’s time-multiplexed CIM achieved an energy of $h_i=0$3, selected 288 distinct SKUs, allocated 226,813 units, produced 12.75 million dollars profit, and had zero capacity violations (Ubale et al., 24 Oct 2025). The formulation used profit, similarity, risk, cardinality, and exact slack-bit capacity penalties in one block-diagonal QUBO.

The STNO-based machine reported lower-fidelity logical primitives than the wave-based systems but demonstrated a different resource trade-off. With the incremental coupling rule, NOT-gate accuracy reached 74%, XOR-gate accuracy 64%, and factorization of 35 achieved 87% accuracy using a single physical oscillator time-shared across logical p-bits (Zhang et al., 2022).

These figures are not directly comparable because they use different tasks, graph families, success definitions, and runtime accounting. A recurring misconception is therefore that “time-multiplexed Ising machine” denotes one standardized benchmark regime. The literature instead shows a family of architectures with different nonlinearities, update laws, and evaluation protocols.

6. Scaling behavior, limitations, and open issues

Scalability is the principal motivation for time multiplexing, but each platform inherits its own limiting mechanisms. Optical coherent Ising machines require longer cavities or lower repetition rates to host more time slots, careful dispersion management to preserve femtosecond pulse integrity, and stable multi-delay phase control as graph size and connectivity grow (Takata et al., 2016). The early optical work argued that time multiplexing gives intrinsically identical oscillators and linear hardware scaling in the number of delay bands, but arbitrary dense graphs still require synchronized modulation of amplitude and phase on the delay lines (Marandi et al., 2014).

Wave-based microwave and acoustic machines exchange optical phase stability for timing and bandwidth constraints. In SWIM, usable spin count is limited by spin-wave dispersion, attenuation, cross-talk, gating jitter, and amplifier nonlinearity (Litvinenko et al., 2022, González et al., 2023). SAWIM improves on spin-wave dispersion because SAWs exhibit linear dispersion over wide amplitude and frequency ranges, but scaling remains constrained by delay time, pulse width, detector response, and feedback latency (Litvinenko et al., 2023). BAWIM emphasizes that higher-frequency delay lines can reduce runtime dramatically; the paper projects about 0.462 ms to reach the same target at 16.455 GHz, about 750 times faster than the 20.5 MHz device (Vadde et al., 2 Jul 2026).

Parameter tuning is another recurring issue. The numerical delay-line oscillator study finds a sharp transition line in the $h_i=0$4 plane above which global minimum probability collapses to zero, and the optimal operating region remains near the lower edge of synchronization across all tested coupling topologies (Ovcharov et al., 2024). SAWIM reports an optimum overall coupling strength: too-weak coupling fails to switch spins reliably, while too-strong coupling leads to local-minimum trapping and broad non-Gaussian energy distributions (Litvinenko et al., 2023). The 16-spin optical machine shows that abrupt pumping traps the single-mode model in metastable domain-wall configurations, whereas gradual pumping and multimode dynamics yield 100% simulated success on the tested instances (Takata et al., 2016).

Noise, nonideal analog behavior, and hidden degrees of freedom remain important. The optical 16-spin machine observed repeated formation and decay of ground-state patterns over about 40 $h_i=0$5s due primarily to pump noise (Takata et al., 2016). The global-bias SWIM study observed mixed 3+1 states in a 4-spin ring that are not minima of the ideal Hamiltonian, attributing them to amplitude mismatches from low-noise-amplifier gain compression (González et al., 2023). The STNO system notes temporal correlations from using a single entropy source, although operation in non-overlapping windows and threshold tuning near $h_i=0$6 mitigate synchronization artifacts (Zhang et al., 2022).

There is also an architectural debate over direct mutual injection versus measurement-feedback. In time-multiplexed coherent Ising machines, all-optical delay-line coupling is low-loss and parallel, but measurement-feedback architectures offer far more flexible dense programmability (Takata et al., 2016). A distinct controversy concerns nonclassical resources. A Gaussian analysis of time-multiplexed CIMs finds that conventional delay-line coupling has a fundamental steady-state entanglement limit $h_i=0$7, while a measurement-feedback scheme based on nonlocal homodyne detection can reach much smaller steady-state $h_i=0$8, including $h_i=0$9 for $N$0 and $N$1 (Yanagimoto et al., 2019). This does not establish a computational advantage, but it does show that different feedback architectures change the accessible correlation structure.

A plausible implication is that “time-multiplexed Ising machine” should be understood less as a single device class than as a systems-design pattern: one shared nonlinear substrate, one temporal register of spins, and one coupling mechanism synchronized to circulation. Within that pattern, optical DOPO networks, spin-wave and acoustic delay lines, serial stochastic p-bit systems, and CMOS-assisted pulsed loops realize different trade-offs among programmability, spin count, thermal stability, coupling precision, and time-to-solution (Takata et al., 2016, Zhang et al., 2022, Litvinenko et al., 2023, Vadde et al., 2 Jul 2026, Love et al., 15 Jun 2026).