Ising machines: Hardware solvers for combinatorial

optimization problems
Naeimeh Mohseni1,2,3 , Peter L. McMahon4,* , and Tim Byrnes5,1,6,7,8,†
1 State Key Laboratory of Precision Spectroscopy, School of Physical and Material Sciences, East China Normal
University, Shanghai 200062, China
2 Max-Planck-Institut für die Physik des Lichts, Staudtstrasse 2, 91058 Erlangen, Germany
3 Department of Physics, University of Erlangen-Nuremberg, Staudtstr. 5, 91058 Erlangen, Germany
arXiv:2204.00276v1 [quant-ph] 1 Apr 2022

4 School of Applied and Engineering Physics, Cornell University, Ithaca, NY 14853, USA
5 New York University Shanghai, 1555 Century Ave, Pudong, Shanghai 200122, China
6
NYU-ECNU Institute of Physics at NYU Shanghai, Shanghai 200062, China
7 National Institute of Information and Communications Technology, Tokyo 184-8795, Japan
8 Department of Physics, New York University, New York, NY 10003, USA
* pmcmahon@[Link]
† [Link]@[Link]

ABSTRACT

Ising machines are hardware solvers which aim to find the absolute or approximate ground states of the Ising model. The
Ising model is of fundamental computational interest because it is possible to formulate any problem in the complexity class
NP as an Ising problem with only polynomial overhead. A scalable Ising machine that outperforms existing standard digital
computers could have a huge impact for practical applications for a wide variety of optimization problems. In this review,
we survey the current status of various approaches to constructing Ising machines and explain their underlying operational
principles. The types of Ising machines considered here include classical thermal annealers based on technologies such as
spintronics, optics, memristors, and digital hardware accelerators; dynamical-systems solvers implemented with optics and
electronics; and superconducting-circuit quantum annealers. We compare and contrast their performance using standard
metrics such as the ground-state success probability and time-to-solution, give their scaling relations with problem size, and
discuss their strengths and weaknesses.

Key points
• Dedicated hardware solvers for the Ising model are of great interest due to the many potential practical applications and
the end of Moore’s law which motivate alternative computational approaches.
• Three main computing methods that Ising machines employ are classical annealing, quantum annealing, and dynamical
system evolution. A single machine can operate based on multiple computing approaches.
• Today, Ising hardware based on classical digital technologies are the best performing for common benchmark problems.
However, the performance is problem dependent and alternative methods can perform well for particular classes of
problems.
• For particular crafted problem instances, quantum approaches have been observed to have a superior performance over
classical algorithms, motivating quantum hardware approaches and quantum-inspired classical algorithms.
• Hybrid quantum-classical and digital-analog algorithms are promising for future development; they may harness the
complementary advantages of both.

Introduction
Computers have had an enormous impact on society, infiltrating almost every aspect of our daily lives. One type of problem
that conventional computers have particular difficulty in solving are hard combinatorial optimization problems. Such problems
typically involve finding an optimal configuration, defined by a cost function, among a very large number of potential candidate
configurations. Examples of such problems include the travelling salesman problem, Boolean satisfiability problems (i.e. SAT
problems), MaxCut, to name a few. In a practical setting, such combinatorial optimization problems are of relevance to a wide
a b d Energy
thermal

Jij=1
tunneling
B
Minimum gain
A

Attractors

Figure 1. The Ising model, combinatorial problems, and its energy landscape. (a) An example of an 8 spin Ising model.
On each node is a two valued spin (arrows). Edges correspond to assigning an energy Ji j = 1. (b) Examples of combinatorial
optimization problems: the travelling salesman problem (top) and MaxCut (bottom). In the travelling salesman problem, the
aim is to find the shortest possible route that visits each city exactly once and returns to the origin city. For MaxCut, the
graph is equivalent to the Ising model in (a) and the the optimal division is indicated by the dashed line. The travelling
salesman problem can be mapped onto the Ising model by encoding the information of the city and its route ordering as a
spin variable. The total number of spins required is the square of the number of cities. (c) Schematic energy landscape of the
Ising model and various mechanisms used in Ising machines to overcome local minima. Shown are thermal excitations using
in classical thermal annealing, quantum tunneling in quantum annealing, the minimum gain principle in coherent Ising
machines, and attractors in dynamical system evolution.

variety of applications such as planning, logistics, manufacturing, financial portfolio management, computer vision, artificial
intelligence, machine learning, bioinformatics, drug design, and a variety of chemical and physical materials problems1–4 . In
many cases, such combinatorial optimization problems are instances of NP-complete problems, which represent the hardest
problems within the NP class. A well-known result states that it is possible to map any problem in NP to a NP-complete
problem in polynomial time5, 6 . In this sense, if there were a way of solving any combinatorial optimization problem in the
NP-complete class with an improvement over conventional computing methods, then this would have an enormous impact to
a large number of practical applications.
To give an example of such an NP-complete problem, consider MaxCut (Fig. 1b). Here, one starts with a graph, where
some of the vertices are connected via edges. The aim then is to group the vertices into two types such that the number of
edges (i.e. links between the two groups) is as large as possible. MaxCut is of direct relevance to problems such as circuit
design7, 8 , machine learning9, and computer vision10, 11 , and therefore even without any mapping is an important problem in
its own right. In a brute force solution of this problem, one requires checking every one of the exponential ways the vertices
can be grouped, and finding the largest number of edges. The MaxCut problem can be recast in a physics language as a spin
glass problem (Fig. 1a). To do this, put a binary valued spin on each vertex σi ∈ ±1, and assign a positive interaction constant
Ji j = 1 between the connected vertices and 0 otherwise. The value of spin then encodes which group a vertex is in, and lowers
the overall energy for connected spins if they are in different groups σi σ j = −1. This can be written as an Ising Hamiltonian
N N
HP = ∑ Ji j σi σ j + ∑ hi σi , (1)
i, j=1 i=1

where N is the number of vertices or spins. We have also included a linear hi term for generality, although for MaxCut it is
not required. Finding the minimum energy of (1) is then equivalent to solving MaxCut. We note that the Ising Hamiltonian
(1) can be related to a Quadratic Unconstrained Binary Optimization (QUBO) problem under a simple change of variables
σi = 1 − 2xi, xi ∈ {0, 1}, and hence they can be regarded as equivalent problems.
The most common way that such optimization problems are solved are on large-scale high performance classical comput-
ers, using variants of Monte Carlo methods. With the demise of Moore’s law it is of interest whether alternative methods —
perhaps based on unconventional methods of computing — could be used to solve such optimization problems. Alternatives
to Turing’s concept of a deterministic digital computing machine12 have a long history of investigation, in particular, the
analog computers used for physical simulators to investigate complex problems13. In an analog computer, the computation is
performed using coupled physical systems which evolve continuously according to their physical dynamics, implemented by
analog electronics or mechanical systems, for example. These were used predominantly in the first half of 20th century when

2/23
digital computing speeds were insufficient, and continued to be used for several decades after for specialized applications such
as flight simulation, although even these applications have been now rendered obsolete.
The recent interest in the field of quantum simulation14–16 in many ways mirrors this development of classical analog
computers, and led to a resurgence of interest in realizing analog Ising model simulators. Quantum simulation was in fact one
of the early motivations for realizing a quantum computer, based on Feynman’s conjecture that a quantum computer should
be able to simulate quantum systems more efficiently than classical computers14 — proven true later by Lloyd17 . However,
a large-scale, fault-tolerant quantum computer is still a challenging goal, technologically. On the other hand, advances in
the manipulations of many-body quantum systems using cold atoms, ions, or artificial qubits potentially allow for a way
of simulating complex quantum systems without requiring the full controllability of a quantum computer15, 16, 18–22 . This
led to the idea that Ising models might be realizable using a quantum simulation approach23, 24 , where alternative models of
computation could be utilized in order to find the ground state more efficiently. The first large-scale physical implementation
of the quantum approach was produced by D-wave Systems, where a 128 qubit quantum annealer was realized, followed by
larger scale systems25, 26 . Today, there are numerous approaches, incorporating a variety of different techniques (both classical
and quantum), which will be described and compared in this review.
In the classical realm, one of the main drawbacks of analog computers in comparison to digital computers is that they
are more susceptible to error, due to the analog storage of the information. Nevertheless, analog computers can have several
advantages over digital computers. First, the operation of the analog computer is typically highly parallelized. For a system
consisting of many coupled systems which encode the information, each of the systems evolves in parallel, in contrast to digital
computers where parallelization is performed across multiple processors. Second, there is no additional overhead due to the
implementation of digital logic. In many cases the time evolution of a physical system is continuous, but is discretized and
evolved in a step-wise sequence, which requires additional resources not required in analog simulation. Analog computing
is in many ways analogous to the way that the brain operates: there is no predefined algorithm, and its operation is inherently
massively parallel and asynchronous. This “natural computing” approach has intrigued researchers for decades, from both of
the point of view of improvements over current computing, as well as understanding the way that biological systems compute.
In this review, we survey various hardware devices that have been developed with the aim of solving the Ising model; here
we call such devices “Ising machines”. An important caveat is what exactly we mean by solve the Ising model. In many
applications suboptimal but still good solutions are acceptable in practice, hence we shall consider primarily heuristic and
approximate solvers. We focus on discussing their underlying operating principles27, and introduce the types of technologies
that have been used to implement them. This includes variations of classical thermal annealers, quantum annealers, as well
as dynamical system-based solvers including the coherent Ising machine, which have attracted interest recently. Other types
of novel computing devices such as those based on hybrid quantum-classical systems are also described. We discuss the
performance of the investigated devices, focusing on the scaling with regard to the size of the Ising problem.

Operating Principles of Ising Machines

Classical thermal annealing
One of the fundamental concepts that is encountered in connection to solving the Ising model — and optimization problems
in general — is annealing. Inspired by concepts in statistical mechanics, the configurations corresponding to the lowest values
of a cost function (i.e. energy) are found by gradually lowering the effective temperature of a system. The basic observation
is that at thermal equilibrium, a classical physical system follows statistics according to a Boltzmann (or Gibbs) distribution

exp(− kEBnT )
pn = , (2)
Z
where Z = ∑n exp(− kEBnT ), kB is the Boltzmann constant, T is thermodynamic temperature, and En is the energy of each spin
configuration of the Ising model (1) that we have labeled by n = {σ1 , . . . , σN }. There are 2N different energy configurations
labeled by n. The lowest energy states appear with higher probability, and the probability of obtaining the ground state, i.e. the
desired solution of the Ising model, increases as the temperature is lowered. To produce such a state at thermal equilibrium,
the system evolves according to a master equation typically of the form
d pn
= −wnm pn + wmn pm , (3)
dt
where rates wnm are the rates for transition from the nth to the mth energy state. The rates are taken such that in the limit
of t → ∞ the probability distribution follows (2). If (3) is evolved long enough, one is guaranteed to obtain the low energy
solutions for a sufficiently low temperature. The main problem is that particular energy landscapes require extremely long
times before thermal equilibrium is reached, due to the possibility of getting trapped in a local minima (Fig. 1c). The solution

3/23
 a b c d

Stochastic magnetic tunnel junctions Memsistor Coupled electrical oscillators CMOS

e f g

Boltzmann Machine Photonic annealer Coherent Ising machine

Figure 2. Example technologies used to realize various types of Ising machines. (a) Stochastic magnetic tunnel junction
showing difference between conventional MRAM and the probabilistic bit from Ref.29 . P is the parallel and AP is the
antiparallel orientations of the magnetic layers. (b) Memristor crossbar array to perform matrix-vector multiplication from
Ref.30 . (c) Metal insulator VO2 system to realize coupled electrical oscillators from Ref.31 . (d) 28 nm CMOS chip to realize
a 1 million spin Boltzmann machine from Ref.32 . (e) Co atoms on surface of black phosphorus interacting with a scanning
tunneling microscope to realize a Boltzmann machine from Ref.33 . (f) Spatial light modulator based photonic annealer from
Ref.34 . (g) Coherent Ising machine measurement-feedback loop from Ref.35 .

to this problem is to gradually lower the temperature, or anneal the system such that at each temperature the system has a
chance to equilibrate. Geman and Geman showed that by reducing the temperature with an inverse logarithmic dependence
with time, one is guaranteed to obtain the ground state28 .
As a classical computer algorithm, simulated annealing (SA) remains one of the most popular algorithms that can be
applied to optimization problems. Run on a classical digital computer, it is preferable to perform an equivalent stochastic
sampling approach, rather than run (3) directly, due to the exponential resources required. As such, typically one uses various
Monte Carlo algorithms employing the Metropolis-Hastings algorithm36, 37 , such that the desired Boltzmann distribution is
obtained. More sophisticated classical algorithms are known to provide substantial improvement over SA such as parallel
tempering38, 39 , population annealing40, and isoenergetic cluster moves41 to mention a few. For both parallel tempering and
population annealing, multiple copies of the system are prepared in random initial states. For parallel tempering, each copy
has a different temperature parameter. The temperature is increased for the copies that perform poorly and is decreased for the
ones perform successfully. In population annealing, poorly performing copies are probabilistically removed and those which
perform successfully are replicated, while reducing the temperature42, 43 .
Simulated annealing is most commonly implemented algorithmically in conventional programmable digital classical com-
puters, but has also been implemented in dedicated hardware using digital hardware accelerators and analog natural computing
approaches. When implemented on dedicated hardware, this provides the chance to exploit the parallelization of digital hard-
ware accelerators and analog computing. For the analog computation approach, numerous physical implementations of Ising
and related models have been realized or proposed, including magnetic devices29, 44–51 , optics34, 52, 53 , memristors30, 54 , spin-
switches55 , quantum dots56 , single atoms33 , microdroplets57, and Bose-Einstein condensates24, 58 (see Fig. 2). For example,
Camsari, Datta and co-workers have demonstrated that stochastic magnet tunnel junctions can act as probabilistic bits, fluc-
tuating due to thermal energy between either parallel or antiparallel mutual orientations of magnetic domains44–46 (Fig.2a).
These are mutually coupled to realize an arbitrary Ising interaction by measuring the orientation of the bits and adjusting
the barrier energy between the two orientations. This was used to factor integers via an adiabatic procedure29. In another
approach, Strachan and co-workers used memristors to perform an analog matrix multiplication of the Ising matrix to evalu-
ate the energy; utilizing the intrinsic hardware noise, they performed a highly parallelized implementation of a Ising model
annealer30 (Fig.2b). For optical systems, the Ising model was realized by encoding the spins using the phase of the light, and
a recurrent feedback network was used to produce the Ising couplings34, 52, 53 (Fig.2f). This was shown to converge towards
the Boltzmann distribution (2), with the primary advantage being the fast parallelized spin updates.
For digital-electronic approaches, hardware accelerators using CMOS application-specific integrated circuits (ASICs)59–63
and field-programmable gate arrays (FPGAs)64–68 have been investigated to solve the Ising model as a type of domain-specific
computing. For example, Yamamoka and co-workers at Hitachi used CMOS circuits to implement 2 × 104 Ising spins, where
each spin interacts with up to 5 local spins59 . Here, random thermal effects were introduced by either introducing random
spin-flips during calculation of spin values, or applying a low supply voltage to the memory cells, which also introduces ran-

4/23
domness at the level of the hardware. Matsubara and colleagues at Fujitsu developed a 8192-spin Ising machine with full con-
nectivity that is based on a digital-CMOS-chip implementation of SA, where spin updates are performed in parallel60, 61, 64, 65 .
The parallelization allows for a large speedup in comparison to a serial implementation of SA. We note that in the context
of machine-learning accelerators69, hardware implementations of Boltzmann machines have been investigated with CMOS
ASICs32, 70, 71 (Fig. 2d), FPGAs72–77 , and GPUs78, 79 . The similarity of the underlying energy model of Boltzmann-machine
hardware accelerators suggests that such technologies could be adapted to act as Ising solvers. Recently, Patel, Salahuddin and
co-workers demonstrated an FPGA implementation of the restricted Boltzmann machine’s stochastic sampling algorithm to
solve the Ising problem67. Here the problem is mapped to a bipartite version and each group of spins is updated applying par-
allel SA63, 80 . The inherent parallelism of this architecture allows parallel sampling which provides substantial improvement
over SA.

Dynamical system solvers

In a thermal annealer, at any given point of time during the evolution, the system is ideally in a state that is at thermal
equilibrium, following the Boltzmann distribution. Likewise, as we shall see in the next section, in a quantum annealer the
system ideally remains in the ground state of the instantaneous Hamiltonian. To ensure these conditions, annealing must
proceed sufficiently slowly such as to maximize the probability that the minimum energy state of the Ising model is obtained.
In contrast to such annealing-based approaches, alternative strategies where the system evolution is much faster than thermal
equilibriation and adiabatic timescales also exist. In such dynamical system approaches, the state of the system is driven
towards the lowest energy state of the Ising model. An early example of such a dynamical solver was proposed by Hopfield
and collaborators81, 82 where electronic circuits were used to realize (1). In this section, we explain three types of dynamical
system solvers: coupled oscillators, coherent Ising machines, and chaotic systems.
Oscillator-based computing
In the 1950s, von Neumann and Goto pioneered a type of analog computer based on coupled oscillators called the “parametron”
computer83–85 . Here the state information, such as the configuration of an Ising spin, is represented by the phase of an oscillator.
In the presence of a non-linearity, an oscillator with resonant frequency ω0 can be phase locked with a pump frequency 2ω0
with two possible stable phases, 0 or π that represent the digital information. In its original conception, information processing
in the parametron computer occurs as a sequence of logical gates. However, it was also shown that a computation can be
performed in a more parallel, natural computing approach (see Refs.86, 87 for a review). Such coupled oscillators can be used
for solving combinatorial optimization problems, such as the Ising model. The basic idea of oscillator based computing can
be captured by the Kuramoto model. This describes a system of oscillators mutually coupled by an interaction88–90 . Consider
N oscillators, labelled by the index i that oscillate with frequency ω0 . Denote the phase of the ith oscillator by φi . Mutually
coupling the oscillators, the dynamical system can be described by
d φi
= ω0 + K ∑ Ji j sin(φi − φ j ) + Khi sin(φi − ω0t), (4)
dt j

where K is a coupling parameter which controls the overall contribution of the Ising dynamics. In the rotating frame, the
sin(φi − φ j ) factor has two steady-state solutions ddtφi = 0, where the phases are either in or out of phase. The sin(φi − ω0t)
term is also stable when φi − ω0t = 0, π . This means that the dynamical set of equations will converge to a particular
configuration of phases, from which a spin readout can be performed. For the constant Ji j and hi = 0 case the Kuramoto
model can be analytically solved to show a dynamical phase transition between unsynchronized oscillators to an synchronized
one for particular interaction strengths. This type of dynamics has been applied to numerous types of applications in artificial
intelligence problems, such as image processing, pattern recognition and generation86, 87 .
Wu, Chua and co-workers proposed using such oscillators to solve the graph coloring problem using a system of coupled
LC circuits91, 92 . Here the aim is to color the vertices of a graph with k colors such that no adjacent vertices have the same
color. For k ≥ 3 this is a NP-complete problem; in Ref.91 it was found that for k = 2 the scheme is able to correctly find
solutions, but for the more difficult k = 3 case the scheme only succeeded for a subset of problem instances. This approach
was theoretically further developed, including explicitly extending to the case of solving Ising problems and analyzing various
possible physical implementations93–98 . Oscillator networks have been experimentally demonstrated with various systems
such as bulk analog electronic95, 99–101 , VO2 insulator-to-metal transition31, 102, 103 (Fig. 2c), spin104, 105 and integrated CMOS
electronic106–108 oscillators to solve problems such as graph coloring, maximum independent set, and the Ising model. In
several of these studies, by adding noise and turning on the interactions smoothly, it was found that this enabled the network
to find low-energy solutions of the Ising model.
A related approach called memcomputing uses networks of Boolean logic gates as a dynamical system solver. These
have conceptual similarities with oscillator-based Ising machines even if they are not explicitly constructed from networks of
oscillators109, 110 . Such approaches have been applied successfully to frustrated-loop Ising model instances111, 112 .

5/23
Coherent Ising Machine
Yamamoto and colleagues have investigated a particular class of oscillator-based Ising machines, dubbed “coherent Ising
machines” (CIMs), that are naturally suited to being implemented with optical oscillators35, 113–121 (Fig. 2g). Each Ising
spin σi in a CIM is encoded in the phase φi of light in an optical mode. To enforce binary spin values φi = 0, π , CIMs use
degenerate optical parametric oscillators (DOPOs), which are a form of parametric oscillator in which phase-sensitive gain
yields oscillations either in-phase or out-of-phase with respect to the oscillator’s pump light114, 122 . Each DOPO represents
a single spin and is part of a network of DOPOs that are coupled together such that the coupling between a pair of DOPOs
is proportional to the Ising spin-spin coupling Ji j . Several ways to realize couplings have been proposed24, 35, 114–116 , but for
the experimental demonstrations performed thus far, the details of the coupling scheme are not crucial for understanding the
Ising-solving capability of each CIM implementation. If one models DOPOs as classical oscillators, then the time evolution
of a CIM can be modeled by the following system of coupled differential equations in the rotating frame:114, 117

dai
= −γ ai + ra∗i − κ |ai|2 ai − g ∑ Ji j a j − ghi + ni , (5)
dt j

where each ai is a complex number representing the optical field in the ith mode, γ is the decay rate of the photons from each
mode, r is the amplification provided by OPO gain, κ is the coefficient of nonlinear loss due to OPO gain saturation, g is a
coupling constant determining the strength of the Ising interactions, and ni are Langevin noise operators associated with the
photon decay and nonlinear gain. The effect of the Ising terms −g ∑ j Ji j a j − ghi can be thought of as additional loss terms
that act on the ith mode.
The key new elements in these equations of motion relative to those for Kuramoto oscillators (4) are that the oscillator
amplitudes are also explicitly considered, in addition to their phases, and there are now loss and gain terms; it is these
terms that are responsible for DOPOs having an oscillation threshold. If one considers a single DOPO, the loss and gain
terms result in the DOPO being bistable: above threshold, a DOPO will oscillate either exactly in-phase or exactly out-of-
phase. Since the Ising terms can be interpreted as spin-configuration-dependent loss, one can interpret the DOPO network
as having a collective-oscillation threshold that is lowest when the Ising terms are smallest, and hence when the represented
spin configuration has minimum energy. If the CIM is operated in a way where the gain r is slowly increased from 0 (when
the DOPO network is below threshold) to ever-higher values (i.e., r is not a constant, but rather a monotonically increasing
function of time t), then, in the absence of noise ni , the DOPO configuration with the lowest loss should oscillate first (see the
minimum gain principle illustrated in Fig. 1c) and the solution to the Ising problem can be read out by measuring the phases
of the light from each DOPO. An important point to note is how slowly r can be increased and have the CIM still oscillate
in the ground state for a length of time sufficient to allow measurement: even in the complete absence of noise (which is
not experimentally realistic, but can be programmed in a computer simulation), the CIM does not find the exact solution to
arbitrary Ising problems in polynomial time. An important technicality that arises in the CIM model (5), as well as in other
oscillator-based Ising machines where the oscillators have both amplitude and phase degrees of freedom (as opposed to just
phase), is that if the amplitudes |ai | of the oscillators are not equal, then the system will tend to minimize the energy of an
Ising instance with a different Ji j matrix than the desired one. This phenomenon is sometimes referred to as a broken mapping
due to amplitude heterogeneity. An intuitive fix is to add a feedback mechanism that forces the amplitudes |ai | to be equal;
this has been studied for XY machines123 and Ising machines118.
The classical description of a CIM (5) is sufficient to explain the results obtained in the experimental demonstrations35, 115, 116, 119, 124–126
to date, since these experiments have used DOPOs with fairly large roundtrip (photon) loss. However, with sufficiently low
loss, each DOPO can generate an appreciable amount of quadrature squeezing, and in this regime the CIM’s dynamics are
more faithfully modeled quantum mechanically117. An interpretation for CIM operation that arises in the quantum-mechanical
formulation is that each DOPO begins in a squeezed state that is approximable by a coherent superposition of in-phase and
out-of-phase coherent states |α i + | − α i, so the below-threshold state of the CIM is one in which every spin configuration
is represented in superposition, and when the CIM goes through threshold, one of the configurations is selected. It is an
open question to what extent quantumness of the DOPO network may improve (or impair) the computational performance
of a CIM120 . A quantum model for a machine conceptually quite similar to a quantum-regime CIM, in which superpositions
|α i + | − α i are also formed, has been studied by Goto127; Goto found that the machine acts as an adiabatic quantum computer
when the pump rate (the equivalent of r in the CIM model above) is increased from 0 sufficiently slowly. This theoretical con-
nection suggests that insights into the solution mechanisms of quantum annealers might be helpful for understanding CIMs,
especially CIMs in which the coupling between OPOs is conservative rather than dissipative, and vice versa.
Besides the CIM, there have been proposals and demonstrations of several different types of optical and optoelectronic
Ising and Ising-like machines in addition to the ones already cited in the subsection on thermal annealers: systems based on
coupled lasers128–130 , optoelectronics131, exciton-polaritons94, 132–135 , and electromechanical systems136, 137 .

6/23
Chaos in dynamical-system solvers
In an ergodic system, the dynamics are such that the system visits all parts of configuration space. This is an attractive
idea in the context of solving the Ising model, since in many approaches getting trapping in local minima is the cause of
the exponential slowdown. Numerical studies studying thermal relaxation have showed that the process is strongly non-
ergodic, and do not visit all parts of configurational space138 . Several studies have suggested by modifying the dynamics
to include chaos, this can lead to an improvement in performance118, 139–141 . Ercsey-Ravasz and Toroczkai proposed and
studied139 a limit-cycle-free dynamical system whose fixed-point attractors are the solutions of a given optimization problem.
The particular optimization problem they designed their system for was k-SAT, which is—like the Ising problem—an NP-
complete decision problem with an NP-hard optimization version. The formulation of the dynamical system involved both
state variables si corresponding to the variables in the k-SAT problem (analogous to spin variables for an Ising problem) and
auxiliary variables ai . The theoretical property of the dynamical system that it avoids becoming stuck in local minima of the
k-SAT cost function is very appealing. However, this comes at a price: the auxiliary variables ai grow exponentially in time.
Two interpretations or implications of this are as follows: an analog hardware implementation of the dynamical system will
require an exponentially growing amount of energy to operate (a prototype CMOS demonstration142 for problems with up to
50 variables artificially capped the signals representing the auxiliary variables at 1 V), and a digital hardware implementation
that integrates the differential equations will need to use exponentially small timesteps because the differential equations
become stiff (observed in Refs.139, 140 ).
It was found empirically in Ref.139, through numerical simulations, that the dynamical system undergoes a transient period
of chaos while solving difficult instances of the k-SAT problem, but not when solving easy instances. It was suggested in this
work that chaos might be unavoidable in approaches to solving hard optimization problems. A discrete-map optimization
algorithm143 applied to solving both k-SAT and Ising problems was also found to exhibit chaotic dynamics. The general
approach in Ref.139 for designing a limit-cycle-free dynamical system that avoids being trapped in k-SAT local minima
through the use of auxiliary variables has been adopted for Ising solving118, and has been implemented and benchmarked with
an FPGA141 .

Quantum approaches
Quantum annealing
Quantum annealing (QA)4, 144, 145 is a heuristic algorithm based on the quantum adiabatic theorem and was first proposed by
Apolloni et al.146 and studied in the context of the Ising model by Kadowaki and Nishimori147. In this algorithm, the system
is initially prepared in the ground state of a Hamiltonian H0 where its ground state is known. A common choice for this initial
Hamiltonian is
N
H0 = − ∑ σix , (6)
i=1

⊗N
√ of qubits, and the ground state is the uniform superposition of all possible configurations |+i ,
where N denotes the number
where |+i = (|0i + |1i)/ 2. The Hamiltonian is gradually reweighted to the desired problem Hamiltonian HP according to

H = (1 − λ (t))H0 + λ (t)HP, (7)

where λ (t) ∈ [0, 1] is the annealing schedule. The annealing process can be viewed as H0 introducing quantum fluctuations
originating from the non-commutability of HP and H0 . These fluctuations are gradually reduced to reach the low-energy
configuration of the classical energy function HP . Based on the quantum adiabatic theorem, if one performs the sweep
sufficiently slowly, the system remains in its instantaneous ground state throughout the evolution144, 148 . The sweep time for
which the adiabaticity can be achieved is proportional to a negative power of the minimum energy gap between two lowest
energy levels during the sweep149–152 . The use of quantum fluctuations in QA has been hypothesized as a potential resource
for a speedup over classical methods. Quantum tunneling allows the system to pass through energy barriers which have a
higher energy than available in the state (see Fig. 1c). However, despite several decades of investigation, the computational
role of coherent tunneling in providing speedup still is not completely understood144, 153, 154 . Part of the reason for this is
the difficulty of simulating QA on classical computers due to the large computational overhead. The only quantum hardware
that has been able to directly test QA with a large number of qubits to date is that developed by D-Wave Systems. While
this technology still suffers from limitations such as the presence of decoherence, control errors, and limited connectivity,
several studies have shown that quantum effects do play a role in the D-Wave machine155–157 . For problem instances that
possess tunneling barriers, QA and quantum-inspired classical algorithms that mimic tunneling155 have been shown to have
an advantage over SA.

7/23
Hybrid Quantum-Classical Algorithms
The aim of variational quantum algorithms158 is to solve classical and quantum optimization problems by combining a
parametrized quantum circuit with a classical optimizer to obtain the variational parameters. The parametrized quantum
circuit can be thought of as preparing a variational quantum state, which is optimized to give the lowest energy state of a given
Hamiltonian. These algorithms are believed to be strong candidates to achieve a practical quantum advantage on noisy inter-
mediate scale quantum (NISQ) devices159 . In the context of combinatorial optimization problems, the quantum approximate
optimization algorithm (QAOA)160 has particularly attracted a lot of interest, partially as a result of the existence of theoretical
guarantees on the approximation ratio that it can achieve for certain classes of optimization problems160, 161 .
The QAOA algorithm can be viewed as a Trotterized version of QA with a parametrized annealing pathway162. The system
is initially prepared in |+i⊗N , the ground state of the Hamiltonian (6). The parameterized quantum circuit transfers the initial
state to the ground state of the target problem Hamiltonian (in the ideal case) by alternately applying the unitary operator
corresponding to the problem Hamiltonian e−iγ j HP and the unitary operator e−iβ j H0 . This sequence generates the following
quantum variational state

|ψ (β , γ )i = e−iβ p H0 e−iγ p HP · · · e−iβ1 H0 e−iγ1 HP |+i⊗N , (8)

where γ = (γ1 , γ2 , · · · , γ p ) ∈ [0, 2π ] p and β = (β1 , β2 , · · · , β p ) ∈ [0, π ] p are 2p variational parameters and p determines the
circuit depth. Next, a classical optimizer is applied to find the optimal β , γ that optimizes the energy expectation E(β , γ ) =
hψ (β , γ ) |HP | ψ (β , γ )i by updating the variational parameters iteratively. Various approaches have been applied for this
classical optimization step such as brute force grid search160, gradient descent methods163, and machine learning models164.
A key feature of QAOA is that the computational power increases with p162, 165 in contrast with QA where the performance
does not always improve with annealing time160 . It is has been found that under reasonable complexity-theoretic assumptions,
QAOA with p = 1 can not be efficiently simulated with classical computers166, or it implies that P=NP. This has led to
speculations that QAOA may be able to demonstrate a quantum computational advantage in the context of an optimization
problems on near term quantum computers166. However, the class of problems that can be solved efficiently with shallow
circuits may not be representative for problems of practical interest. For example, for all-to-all connected Ising models and
MaxCut, it has been shown that deep circuits may be required167. Therefore, to benchmark computational advantages of
QAOA against classical algorithms one needs to go far beyond problems that can be solved with a shallow circuit and explore
power of QAOA at intermediate depths. However, at current technological levels such circuits are prone to decoherence and
gate errors167.
QAOA has been demonstrated at the small-scale on platforms such as superconducting qubits167, photonics168 to trapped
ions165 . To date, no large-scale (i.e., N > 50) demonstrations of QAOA have been experimentally performed. We note that
classical simulations showing expectation results for single-layer (p = 1) QAOA on problems with up to N = 105 spins have
been performed169, but as is the case with quantum annealers, it is expected that large-scale quantum hardware will be needed
to properly evaluate the performance of QAOA in general.

Other quantum algorithms

Several other quantum algorithms have been proposed to solve combinatorial optimization problems (see Ref.170 for a re-
view). These include using amplitude amplification based approaches, and quantum simulated annealing (not to be confused
by simulated quantum annealing below), where the aim is to prepare the quantum Gibbs state, a superposition state with
Boltzmann probabilities (2) as the amplitudes. Attaining the quantum Gibbs state is done by performing a quantum walk,
such that after many iterations the desired coherent Gibbs state is obtained171–173 . Then in a similar way to thermal annealing,
the temperature is gradually lowered in order to obtain a low-energy state.

Other classical algorithms

Quantum-inspired classical algorithms
Novel types of classical algorithms have been proposed which are inspired by quantum algorithms. Such quantum-inspired
algorithms are run on conventional computing hardware or on digital hardware accelerators, hence are classical approaches,
but use concepts that originate from quantum mechanics in the algorithm. We briefly summarize several approaches in this
direction as below.
In simulated quantum annealing (SQA), quantum Monte Carlo is applied to estimate the low-energy states of the QA
Hamiltonian174–176 . To perform the quantum Monte Carlo, a stoquastic QA Hamiltonian is mapped to a classical Hamiltonian
by introducing an extra spatial dimension, corresponding to imaginary time. A stoquastic Hamiltonian is characterized by
having only nonpositive off-diagonal elements in the computational basis. The new Hamiltonian has the equivalent equilibrium
properties with the original QA Hamiltonian177. The mapping can be implemented both in discrete time by applying the
Trotter-Suzuki decomposition, or in continuous time by applying a path integral177. Quantum Monte Carlo in the continuous

8/23
time limit samples the equilibrium thermal state of a quantum system (as opposed to directly simulating its unitary time
evolution) and can generate Boltzmann distributed states (2). At sufficiently low temperatures, SQA can mimic tunneling
effects. SQA can also generate entangled ground states that occur during the adiabatic evolution. It can thus faithfully predict
the performance of QA for stoquastic Hamiltonians.
Several other quantum-inspired classical algorithms based on dynamical system evolution have been proposed which we
briefly describe. The simulated coherent Ising machine simulates the CIM on a classical computer and has shown a speedup in
comparison to a physical implementation of a CIM applying FPGA178 and GPU179 . The key observation here is that simulating
such behavior is described by a set of coupled equations of the form (4) or (5) only scale with the number of spin variables N,
rather than the configurational space 2N . This makes a simulation of the coupled oscillator system efficient.
Another approach is simulated bifurcation (SB), which is based on simulating adiabatic evolutions of classical nonlinear
Hamiltonian dynamical systems. This algorithm is the classical counterpart of bifurcation-based adiabatic quantum computa-
tion127 . Two branches of the bifurcation in each non-linear oscillator represents two states of each Ising spin. In 2019, Toshiba
developed an FPGA and GPU-based SB machine showing excellent performance due in part to its high parallelizability180, 181 .
The operational mechanism of SB algorithm is based on an adiabatic and ergodic search. Later, two other variants of SB
were introduced, called the ballistic simulated bifurcation algorithm (bSB) and the discrete simulated bifurcation algorithm
(dSB)182 that far outperform the original SB in terms of both speed and solution accuracy. These new algorithms apply new
approaches, such as a quasi-quantum tunneling effect. Recently, a multi-chip architecture using a partitioned version of the
SB algorithm was implemented with FPGAs, showing that large scale Ising problems can be handled using the method183.
Both CIM simulations and the SB algorithm are highly amenable to parallelization, by simultaneously updating at each time
step N coupled-oscillator variables. This is in contrast to SA, which canonically involves sequential updates of spins, with
simultaneous updates allowed only for isolated spins.
Yet another quantum-inspired algorithm involves tensor networks which are a powerful framework that provides represen-
tations of complex quantum states based on their entanglement structure184. Tensor networks have been applied as an ansatz
to solve optimization problems184, 185 . Such an approach was used in the context of dynamic portfolio optimization, which
can be encoded as an Ising problem186.

Machine learning approaches

Machine learning algorithms can be applied to either boost the performance of traditional classical solvers and quantum algo-
rithms187 or can work as a stand-alone solver. For example, some machine learning algorithms have been applied to accelerate
Monte Carlo simulations188, 189 . It also has been applied as a stand-alone solver and found to have excellent performance in the
context of portfolio optimization185. The emergence of the methods that are more sample-efficient make them more scalable
to large-scale problems. Deep learning based methods applying reinforcement learning190, 191 , graph neural networks192–194 ,
and neural attention mechanism195 also have been investigated as solvers for combinatorial optimization problems.

Computational complexity
One may ask how the various computing approaches discussed in this review relate to computational complexity, and what
the prospects are to devise an Ising machine that can solve Ising problems efficiently (i.e., in polynomial time). While the
?
P = NP question remains an open problem, it is widely conjectured that P 6= NP, i.e., that certain problems in NP, including
the Ising problem, are fundamentally more difficult to solve than those in P. Indeed, decades of computer science, physics,
mathematics, and operations research has failed to find a polynomial-time algorithm that solves any NP-complete problem.
The explosion of interest in quantum computing since the 1990s was kicked off by the discovery that integer factorization
could be performed in polynomial time on a quantum computer196. It is thus conjectured that the BQP complexity class—
decision problems a quantum computer can solve in polynomial time with an error probability of at most 1/3—is a larger class
than P, i.e., P ⊆ BQP. The associated class for a probabilistic Turing machine, the BPP class, is meanwhile conjectured to be
equivalent to P, i.e., P = BPP, which in general remains unproven but is true if a suitable pseudorandom number generator is
available197 . Although there is no proof that quantum or probabilistic computers cannot solve NP-complete problems such as
the Ising problem in polynomial time, it is considered unlikely198.
The complexity-class arguments above concern solving the Ising problem in the sense of being able to find the exact
ground state for all possible problem instances (all possible Ji j , hi in (1)). However, as mentioned in the introduction, an
approximate solution with an energy close to the true ground state is often acceptable for practical applications. We may
distinguish three approaches for solving combinatorial optimization problems such as the Ising problem: exact algorithms,
approximation algorithms, and heuristic algorithms. Exact algorithms are ones designed to find exact solutions in a way that
guarantees that the returned solutions are exactly optimal. Approximation algorithms return solutions that are not necessarily
optimal but provide a guarantee on how far a returned solution is from being optimal. Heuristic algorithms return solutions
without any guarantee on the quality of them; because of this lack of theoretical guarantee, the primary basis for trusting an

9/23
heuristic algorithm is from prior empirical (benchmarking) results. Both approximation and heuristic algorithms tend to be
practical to run on large problems.
For approximation algorithms, we may ask what solution quality (formally, approximation ratio) one could hope to guar-
antee with any algorithm. The MaxCut problem, and hence the Ising model, is APX-hard (i.e. approximable-hard). The main
consequence of this is that, assuming P 6= NP, there exists no polynomial-time approximation algorithm for the Ising problem
that will guarantee a solution arbitrarily close to the exact solution199. However, there does exist a polynomimal-time approx-
imation algorithm for MaxCut that will find solutions a fixed distance from the optimal solution: the Goemans-Williamson
algorithm is guaranteed to find solutions within ≈ 12% of the optimal value200. It is also known that it is NP-hard to approxi-
mate MaxCut with solutions guaranteed to be closer than ≈ 6% to the optimal201, so it is expected (assuming P 6= NP) that no
polynomial-time approximation algorithm achieving this approximation closeness will be possible. In many practical settings
it is desirable to find solutions to MaxCut problems having distance from the optimal solutions that is better than ≈ 12% or
even ≈ 6%, and this motivates the use of heuristic algorithms to solve MaxCut in practice.
Most Ising machines are heuristic solvers—that is, they can be thought of as physical machines realizing heuristic opti-
mization algorithms. As such, they typically do not provide any approximation-ratio guarantees. The potential advantages of
Ising machines largely lie outside the realm of complexity theory: there is the possibility that Ising machines have a polynomi-
ally improved scaling or constant-prefactor advantage in comparison to existing heuristic algorithms running on conventional
processors. i.e., it is generally expected that Ising machines, regardless of their underlying algorithm or practical hardware
implementation, will still have runtimes scaling exponentially in order to achieve near-optimal solutions, but that the exponent
may be smaller than for a conventional solver, or that the constant factor in front of the exponent may be smaller. A small
difference in the exponent can make a large difference in runtimes for large problem sizes, and the fast clock speeds of various
physical implementations, which give rise to constant-factor improvements, could lead to significant practical speedups versus
conventional state-of-the-art solvers.

Computation performance comparisons

Since the utility of any Ising machine is in its ability to solve a given Ising problem both quickly and accurately, an important
task is to benchmark the performance and compare between competing methods. Here we will direct our attention particularly
to various Ising solvers that have been experimentally tested for relatively large systems N ≥ 50. We consider only large
systems as it is rather difficult to extract any scaling relation for smaller systems, and by choosing experimentally realized
systems this will tend to focus on technologies that are relatively near to maturity. For the figures of merit, we focus upon two
of the most commonly employed quantifiers: the success probability and the time-to-solution. We first define each of these.
For the success probability, this is defined as the probability that the exact ground state of the Ising model is obtained for
a single run of the Ising machine. The success probability is a quantity that depends inherently on algorithmic parameters.
For example, for annealing methods under ideal conditions, longer annealing times generally result in higher success proba-
bilities. In this sense, the success probability can also be made arbitrarily close to 1 in the ideal case. However, other factors
typically prohibit approaching unit success probability, due to practical considerations. For example, in quantum annealers
the annealing time must be within the coherence time in order to maintain a quantum superposition. In this sense, the success
probability still has a meaning, since it often involves a trade-off with practical considerations. For our comparison of success
probabilities, we generally quote the best performing value available in the literature.
One of the limitations of the success probability as a figure of merit is that it does not take into account of how much time
it takes for a single run of the Ising machine. An Ising machine will typically perform multiple runs when attempting to solve
a problem and Ising machines are often optimally operated for a choice of run parameters where the success probability for
a single run is not maximized, but where each run takes only a short amount of time. The time-to-solution is another figure
of merit that takes into account of the time taken to perform a single run on a given Ising machine, as well as the success
probability. Suppose that r runs of a particular scheme are performed, which each have a success probability psuc of obtaining
the ground state . Then the collective probability of getting at least one successful run, in which the ground state is found, is
1 − (1 − psuc )r . Now let us demand this collective probability to be a desired value, say 99%, and each run takes a time τ . The
time-to-solution is then related to the success probability as
ln 0.01
Tsol = τ . (9)
ln(1 − psuc)
This measure takes into account of differences in clock speed of various approaches, and also allows for various approaches
to choose their optimal parameters such that the best performance of the machine can be extracted.
In Fig. 3, we show the performance of various Ising machines, quantified by the success probability for random instances
of Sherrington-Kirkpatrick (SK) problems and dense MaxCut problem instances. We note that while the same types of models
are used for the comparison in Fig. 3 and 4, the same problem instances were not necessarily used, as we have compiled results

10/23
 SK MaxCut
1 1 MRT*
0.50 0.50
success probability

RBM

success probability
RBM
CIM2
CIM2
0.10 CIM1 0.10
0.05 0.05 CIM1
DWAV CIM3*
a DWAV CIM3*
b
0.01 0.01
0 20 40 60 80 100 120 140 0 50 100 150 200
N N

Figure 3. Success probability comparison of various Ising machines. The probability of obtaining the ground state is
shown for (a) Sherrington-Kirkpatrick (SK) and (b) dense MaxCut problems. For the SK problem, Ji j is chosen from ±1
with equal probability. The MaxCut problem is mapped onto the Ising model by setting Ji j to 0 and 1 with equal probability.
In both cases hi = 0. The labels for each line and their references are given in Table 1. Error bars on original data where
present have been omitted for clarity. Here, CIM1, CIM2, CIM3, DWAV and RBM are benchmarked on the same problem
instances. Data reported for theoretical predictions, rather than being directly measured from an hardware implementation,
are labeled with *.

from different studies. While this does not make our comparisons perfect, we hope that this nevertheless gives a sense of state
of the art of various approaches. In numerous works, the general scaling behavior is observed to follow the relation

psuc ∝ e−bN , (10)

where b is a fitting parameter. Keeping in mind the interpretational caveat mentioned above about how the success probability
can for some Ising-machine approaches been made high at the expense of very long runtimes in a way that is ultimately
not useful, Fig. 3 gives evidence suggestive that at current technological levels, SA-based approaches, such as the restricted
Boltzmann machines (RBM), implemented on digital hardware give the best scaling with N. One of the reasons for the
high success probability of RBM is the inherent parallelism of this architecture which allows parallel SA updating. We do
however note that this figure does not include results from state-of-the-art dynamical-systems algorithms such as the CIM
with amplitude-heterogeneity correction118, 141 and SB182 , and based on these algorithms’ excellent performance on the G-set
MaxCut instances, one may anticipate that they would be competitive with RBM-based solvers. It is notable that the quantum
annealer has a particularly poor performance in comparison to other methods. This can be understood119 as a consequence
of the benchmarked D-Wave machine having qubit connectivity given by a low-degree (Chimera) graph that cannot natively
implement either the dense MaxCut or SK models (see Table 1). An embedding procedure that requires ∝ N 2 physical qubits
is used to realize the equivalent graph, and this puts the D-Wave annealer at a disadvantage in comparison to the other listed
approaches featuring all-to-all spin connections. It is for this reason that the success probability has a relation which more
2
resembles psuc ∝ e−bN .
Figure 4a and 4b shows the time-to-solution metric for the MaxCut and SK models. The best performing methods for the
SK model employ classical digital hardware, where RBM and TBM show the lowest time to solution. For MaxCut problem,
the lowest time-to-solution achieved for a physically implemented machine is RBM. We note that the MRT, PRIS, and CIM3
curves involve theoretical prediction of the time-to-solution, rather than a direct measurement of the time. The scaling relation
for most of the curves follows the phenomenological relation
√
Tsol ∝ ec N
. (11)

For the cases with a relatively small range of available data the square root behavior may not yet be visible. The D-wave
results are better approximated by an exponential relation Tsol ∝ ecN which requires ∝ N 2 physical qubits due to the limited
chimera connectivities of the qubits. One should note that this is a hardware implementation limitation that gives a different
scaling and not the computational mechanism itself, and may be improved in the future204–207 . For problem instances with

11/23
 SK MaxCut
PT3 BLS
CIM2 SA2 0.100 DWAV1 CIM2
SA1

time to solution (s)

10

time to solution (s)

DWAV1 CAC PT2 CIM1
FDA1
0.001
PT1
0.01 CIM1 TBM1
PT1 RBM
RBM 10 5
PRIS*
5
10 CIM3*
7 MRT*
CIM3* a 10
b
0 200 400 600 800 1000 0 50 100 150 200
N N

3R3X Logical-Planted/deceptive
4 1 SA3
10 HFS
10 PT4
time to solution (s)

time to solution (s)

2 MEM
10 TBM2 -1 SQA1
SAT 10 PT+ICM

1 DWAV3
FDA2
-3
-2 10
10 DWAV4
-4 -5
10 DWAV2 c 10 d
0 300 600 0 1000 2000
N N

Figure 4. Time-to-solution (TTS) comparison of various Ising machines. The time to obtain a 99% success probability
of obtaining the ground state is shown for the (a) SK; (b) dense MaxCut problems; (c) 3R3X problems (reproduced from
Ref.202 ); (d) logical-planted (dashed lines, from Ref.155 ) and deceptive cluster loops (solid lines, from Ref.203 ) instance
classes. For the SK and MaxCut cases, the Ising model definitions are as in Fig. 3. The 3R3X, logical planted, and crafted
problem definitions can be found in Refs.202 ,155 ,203 respectively. The labels for each line and their references are given in
Table 1. Error bars on original data where present have been omitted for clarity. For both SK and MaxCut, CIM1, CIM2,
CIM3, DWAV and RBM are benchmarked on the same problem instances. We note that for each reference, the best time to
solution quoted are taken for each N. For results showing multiple annealing times, we have taken results optimized over
annealing times. Data reported for theoretical predictions, rather than being directly measured from an hardware
implementation, are labeled with *.

sparse connectivity D-wave was observed to have a more favorable scaling119 . These results show that the connectivity is an
important factor that determines the performance of an Ising machine — in Table. 1, the hardware connectivity of the different
Ising machines are given.
In Fig. 4c, we show results from Ref.202 , which compare various Ising machines for 3R3X problems, which have a golf-
course energy landscape structure with known exact solutions. The best performing approach in this case is the SATonGPU
approach, which is a highly parallelized version of a SAT algorithm implemented on a GPU. The Fujitsu Digital annealer and
Toshiba bifurcation machine achieves almost similar scaling although has a larger prefactor than the SATonGPU approach. The
memcomputing results are based on classical simulation of a proposed system, hence with dedicated hardware to implement
it, some performance improvement might result202. Although there are fewer studies performed for this problem class, these
results again suggest that the best performing solvers today are based on digital computing hardware.
To show the potential of quantum approaches, we also discuss additional problem classes where it is expected that QA
has advantages over a class of classical methods155–157 despite the mentioned limitations of D-wave. In Fig. 4d, the optimum
time-to-solution for the class of logical-planted (LP) problems that are constructed such that they promote the presence of
tunneling barriers is compared (dashed lines). For these problems it is expected that barriers can be traversed more effectively
by quantum, rather than thermal fluctuations. Here D-Wave and SQA shows a scaling advantage over SA155 . The superior per-
formance of SQA implies that tunneling through barriers may not be considered the exclusive advantage of quantum hardware.
However, one should note that that SQA can not be applied for non-stoquastic Hamiltonians where there is a sign problem,
and as such the power of QA for non-stoquastic Hamiltonians require further exploration. Non-stoquastic Hamiltonians are

12/23
 Hardware Benchmark
Ising machine/Algorithm Acronym Operating principle Hardware Parallelization Reference
connectivity problem
Breakout local search BLS Local search & simulated annealing algorithm CPU All-to-all No SK Fig. 3a141
Chaotic amplitude control CAC Dynamical chaotic algorithm FPGA All-to-all Yes SK Fig. 3a141
Coherent Ising machine (NTT) CIM1 Dynamical oscillator Hybrid (Optical/FPGA) All-to-all Yes MaxCut, SK Fig. S6119
Coherent Ising machine (Stanford) CIM2 Dynamical oscillator Hybrid (Optical/FPGA) All-to-all Yes MaxCut, SK Fig. S6119
Coherent Ising machine CIM3 Dynamical oscillator algorithm *Predicted All-to-all Yes MaxCut, SK Fig. S10119
D-Wave quantum annealer 2Q DWAV1 Quantum annealer Superconducting qubits Chimera Yes MaxCut, SK Fig. 3b, 4c119
D-Wave quantum annealer Advantage1.1 DWAV2 Quantum annealer Superconducting qubits Chimera Yes 3R3X Fig. 2202
D-Wave quantum annealer 2KQ DWAV3 Quantum annealer Superconducting qubits Chimera Yes LP Fig. 2155
D-Wave quantum annealer 2KQ DWAV4 Quantum annealer Superconducting qubits Chimera Yes Deceptive Fig. 1203
Fujitsu digital annealer FDA1 Simulated annealing algorithm ASIC All-to-all Yes SK Fig. 7a61
Fujitsu digital annealer FDA2 Simulated annealing algorithm ASIC All-to-all Yes 3R3X Fig. 2202
Hamze–de-Freitas–Selb HFS Tree sampling CPU All-to-all No Deceptive Fig. 1203
Memcomputing MEM Dynamical logic gate algorithm CPU All-to-all Yes 3R3X Fig. 2202
Memristor annealing MRT Simulated annealing algorithm *Predicted All-to-all Yes MaxCut Fig. 6a, 6b30
Photonic recurrent Ising sampler PRIS Oscillator based annealer *Predicted All-to-all Yes MaxCut Fig. 2b53
Parallel tempering PT1 Simulated annealing algorithm CPU All-to-all No MaxCut, SK Fig. S12119
Parallel tempering PT2 Simulated annealing algorithm CPU All-to-all No SK Fig. 7a61
Parallel tempering PT3 Simulated annealing algorithm CPU All-to-all No SK Fig. 3a141
Parallel tempering PT4 Simulated annealing algorithm CPU All-to-all No 3R3X Fig. 2202
Isoenergetic cluster moves+ parallel tempering PT+ICM Monte Carlo algorithm Digital- CPU All-to-all Yes Deceptive Fig. 1203
Restricted Boltzmann machine RBM Simulated annealing algorithm FPGA All-to-all Yes MaxCut,SK Fig. 3, 467
Simulated annealing SA1 Simulated annealing algorithm CPU All-to-all Yes SK Fig. 3a141
Simulated annealing SA2 Simulated annealing algorithm CPU All-to-all No SK Fig. 7a61
Simulated annealing SA3 Simulated annealing algorithm GPU All-to-all Yes LP Fig. 2155
SATonGPU SAT SAT algorithm GPU All-to-all Yes 3R3X Fig. 2202
Simulated quantum annealing SQA1 Quantum Monte Carlo algorithm GPU All-to-all Yes LP Fig. 2155
Toshiba bifurcation machine TBM1 Discrete simulated bifurcation algorithm FPGA All-to-all Yes SK Fig. 3c182
Toshiba bifurcation machine TBM2 Discrete simulated bifurcation algorithm GPU All-to-all Yes 3R3X Fig. 2202

Table 1. Types of Ising machine examined in Figs. 3 and 4. For each type of Ising machine, the operating principle,
hardware type, hardware connectivity, and the parallelization are shown. We note that if the results of Fig. 3 and 4 are for
theoretical predictions, rather than being directly measured from an hardware implementation, the hardware type is quoted as
being “*Predicted”. In the Parallelization column we indicate approaches where simultaneous updates of Ising spins are
performed. The Reference for publication source of the plotted data is given.

important from a computational complexity perspective as adiabatic quantum computation with non-stoquastic Hamiltonians
is equivalent to the circuit model of quantum computing208. Therefore, they can simulate other universal models with at most
a polynomial resource overhead. The fact that D-Wave outperforms SA confirms the presence and advantage of quantumness
but the superior performance of SQA suggests that current QA hardware is still dominated by classical dynamics and needs
to be improved. We note that for the LP problem class, there are classical algorithms that outperform or have comparable
performance with D-Wave155 . In Fig. 4d (solid lines), we show results from a separate study203 that compares D-wave with
classical heuristic algorithms for another class of specially crafted problem, called the deceptive cluster loop problem. For this
problem class, D-wave outperforms the best known heuristics algorithms such as parallel tempering Monte Carlo with isoener-
getic cluster moves (PT+ICM) and Hamze–de Freitas–Selby (HFS) with approximately two orders of magnitude shorter TTS.
However, no scaling improvement is evident.
Numerous other benchmarking studies of Ising machines have been performed in the literature (see for e.g. Refs. 30, 61,
67, 119, 141, 167, 178, 180, 182, 209). For example, in Ref.209 , the performance of the D-Wave hybrid solver, TBM, FDA,
and SA was benchmarked for three different classes of problem instances including SK. The results highlight the fact that the
performance of machines is problem dependent. In particular for SK model, TBM showed the best performance. In Ref.167 ,
the performance of QAOA was benchmarked on SK and MaxCut problems for problems up to 23 qubits.

Discussion and Outlook

We have surveyed various approaches to Ising machines and their computational performance. We identify three dominant
operating-principle strategies that Ising machines employ: classical annealing, quantum annealing, and dynamical system
evolution. All three approaches are all in principle compatible with the natural-computing approach, where information is
encoded in an analog and parallelized way such that the underlying physics drives the system towards the ground state.
In the dynamical-system approach, the evolution drives the state of the system far from equilibrium, in contrast to annealing
approaches where the aim is to avoid occupying high-energy states. In classical and quantum annealing, adding thermal or
quantum noise is an important component of the procedure to escape local minima in the energy landscape. On the other hand,
in dynamical approaches, the system is designed to be attracted to particular configurations that correspond to low-energy
configurations of a given Ising instance. Comparing the performance of Ising machines, interestingly, most approaches tend
to have similar scalings in terms of the error probability and the time-to-solution metrics as a function of the number of
spins, despite extremely different approaches and technologies used to realize them. The complexity of all approaches scale

13/23
exponentially with the system size, with the difference being the power within the exponent and the prefactors. This is
expected given the NP-complete complexity of the Ising problem—the battle between competing approaches is with respect
to the exponents that are achievable, where a small difference in the exponent makes a large difference in time-to-solution for
large system sizes.
While Fig. 3 and 4 suggest that classical digital methods are still the best-performing approaches at the time of writing;
analog and quantum computing technologies are rapidly developing and the landscape may completely change in a short
amount of time. Some of the best-performing approaches are based on classical digital technology which have had the
benefit of decades of development, and in many cases can be highly parallelized. In comparison, QA approaches have only
been recently developed to a scale where it can be tested, either theoretically or experimentally, and often have hardware
limitations such as limited connectivity and the presence of decoherence. The specific form of the Ising instance being solved
(e.g., the structure of the J matrix) can affect the performance dramatically. It may be that in the future, much like various
numerical algorithms are chosen based on the compatibility of a particular problem, that different Ising machines will be
utilized according to their suitability for the given problem. For example, while problem instances with small spectral gaps
are known to be hard for QA, they may not limit the performance of QAOA162 . One step further would be to explore hybrid
quantum-classical and digital-analog algorithms to gain the complementary advantage of each210 .
One point of active debate has been the role that quantum mechanics plays in coherent Ising machines and in D-Wave’s
quantum annealers. In the context of coherent Ising machines, there exist models of their operation that treat CIMs quantum
mechanically114, 117 , including, for example, a description of the initial state as being in a coherent superposition of all logical
states117 . However, experimental realizations of CIMs thus far35, 115, 116, 121, 125 have been in regimes of high photon loss,
where purely classical models can accurately describe the pertinent dynamics of the systems. The clearest indication of
this is that similar performance to demonstrated coherent Ising machines may be achieved by simulating the mean-field
dynamics119, 178, 179, 211 . With sufficiently high nonlinearity to loss in their constituent OPOs, CIMs can be firmly in the
quantum regime212 and have a strong connection with quantum annealers127. Exploring how to construct experimental CIMs
where quantum effects play a crucial role, and designing them so that quantum effects improve the performance of the machine,
are two topics of active investigation120.
For QA, the computational advantage of incoherent tunneling over certain classical methods (typically SA42 and the
Hamze–de Freitas–Selby algorithm213, 214 ) for certain classes of problems has been shown144, 154–157, 203 . However, for any
real-world problem of interest, no evidence of an unqualified quantum speedup (as defined in Ref.215 ) has been found. Perhaps
the most compelling results with the D-Wave QA so far are for a specially crafted problem class, deceptive cluster loops, for
which the QA was found to outperform in terms of time-to-solution for all classical heuristics that were tested, including
parallel tempering41. The speedup was of an approximately constant-factor nature, with no strong evidence of a scaling
advantage203. Another disadvantage of quantum annealing is that they cannot sample uniformly all low-lying states in contrast
to other heuristic algorithms such as SA-based algorithms216, 217 . In SA, after many repetitions and starting from different
initial states one can record all the configurations that minimize the problem Hamiltonian. For an optimization machine, such
an ability to sample fairly is beneficial as having different solutions for a problem is often useful. Furthermore, it is not yet
well-understood what the role of entanglement in QA is and whether it contributes to a quantum speedup218. While various
aspects of quantumness, including entanglement, might or might not aid the performance of CIMs or QAs, this uncertainty
has inspired the proposal of interesting quantum-inspired classical algorithms related to CIMs and QAs, which is a fruitful
development in its own right.
Looking to the future, there is much room for development for Ising machines. In the same way that the scaling of the
time-to-solution and other metrics for classical algorithms are known to several decimal places, the scaling of Ising machines
require more precise quantification such that competing methods can be compared. Figs. 3 and 4 are extremely preliminary
in this regard. With improved quantification and investigation for different classes of problems, a better understanding of
the suitability of various approaches for a particular problem can be known in advance. Another interesting direction is to
compare the performance of the different Ising machines for finding an approximate solution with different levels of accuracy,
since for many applications finding a high-quality solution, rather than the exact solution, is sufficient. Several forms of Ising
machine whose operation either relies on or can be enhanced by quantum-mechanical mechanisms have been proposed and
demonstrated. However, constructing large-scale quantum machines with high connectivity and low decoherence remains
an outstanding challenge for the field of quantum information processing in general, and further progress in this direction is
needed for experimental exploration of the benefits quantum-mechanical methods may bring to solving Ising problems. This
is in contrast with classical approaches, especially digital ones, which often have little difficulty in supporting full connectivity.
Given the demand for faster methods of solving optimization problems in society, and the maturity of conventional algorithms
and processors, it seems likely that the development of specialized Ising machines will continue well into the future, featuring
an exciting interplay between hardware engineering, computer science, statistical physics, and quantum mechanics.

14/23
References
1. Lucas, A. Ising formulations of many NP problems. Front. Phys. 2, 5 (2014).
2. Tanahashi, K., Takayanagi, S., Motohashi, T. & Tanaka, S. Application of Ising machines and a software development
for Ising machines. J. Phys. Soc. Jpn. 88, 061010 (2019).
3. Smelyanskiy, V. N. et al. A near-term quantum computing approach for hard computational problems in space explo-
ration. arXiv preprint arXiv:1204.2821 (2012).
4. Hauke, P., Katzgraber, H. G., Lechner, W., Nishimori, H. & Oliver, W. D. Perspectives of quantum annealing: Methods
and implementations. Reports on Prog. Phys. 83, 054401 (2020).
5. Karp, R. M. Reducibility among combinatorial problems. In Complexity of computer computations, 85–103 (Springer,
1972).
6. Mézard, M., Parisi, G. & Virasoro, M. A. Spin glass theory and beyond: An Introduction to the Replica Method and Its
Applications, vol. 9 (World Scientific Publishing Company, 1987).
7. Barahona, F., Grötschel, M., Jünger, M. & Reinelt, G. An application of combinatorial optimization to statistical physics
and circuit layout design. Oper. Res. 36, 493–513 (1988).
8. Chang, K. & Du, D.-C. Efficient algorithms for layer assignment problem. IEEE Transactions on Comput. Des. Integr.
Circuits Syst. 6, 67–78 (1987).
9. Wang, J., Jebara, T. & Chang, S.-F. Semi-supervised learning using greedy Max-cut. The J. Mach. Learn. Res. 14,
771–800 (2013).
10. Collins, T. Graph cut matching in computer vision. Univ. Edinb. (2004).
11. Arora, C., Banerjee, S., Kalra, P. & Maheshwari, S. An efficient graph cut algorithm for computer vision problems. In
European conference on computer vision, 552–565 (Springer, 2010).
12. Turing, A. M. On computable numbers, with an application to the entscheidungsproblem. Proc. Lond. mathematical
society 2, 230–265 (1937).
13. Bournez, O. & Pouly, A. A survey on analog models of computation. In Brattka, V. & Hertling, P. (eds.) Handbook of
Computability and Complexity in Analysis, 2018 (Springer in cooperation with the Association Computability in Europe,
2018).
14. Feynman, R. P. Simulating physics with computers. Int. J. Theor. Phys 21 (1982).
15. Buluta, I. & Nori, F. Quantum simulators. Science 326, 108–111 (2009).
16. Georgescu, I. M., Ashhab, S. & Nori, F. Quantum simulation. Rev. Mod. Phys. 86, 153 (2014).
17. Lloyd, S. Universal quantum simulators. Science 1073–1078 (1996).
18. Greiner, M., Mandel, O., Esslinger, T., Hänsch, T. W. & Bloch, I. Quantum phase transition from a superfluid to a Mott
insulator in a gas of ultracold atoms. nature 415, 39–44 (2002).
19. Labuhn, H. et al. Tunable two-dimensional arrays of single Rydberg atoms for realizing quantum Ising models. Nature
534, 667–670 (2016).
20. Bernien, H. et al. Probing many-body dynamics on a 51-atom quantum simulator. Nature 551, 579–584 (2017).
21. Keesling, A. et al. Quantum Kibble-Zurek mechanism and critical dynamics on a programmable Rydberg simulator.
Nature 568, 207–211 (2019).
22. Scholl, P. et al. Quantum simulation of 2D antiferromagnets with hundreds of Rydberg atoms. Nature 595, 233–238
(2021).
23. Farhi, E. et al. A quantum adiabatic evolution algorithm applied to random instances of an NP-complete problem.
Science 292, 472–475 (2001).
24. Byrnes, T., Yan, K. & Yamamoto, Y. Accelerated optimization problem search using Bose–Einstein condensation. New
J. Phys. 13, 113025 (2011).
25. King, A. D. et al. Observation of topological phenomena in a programmable lattice of 1,800 qubits. Nature 560, 456–460
(2018).
26. Harris, R. et al. Phase transitions in a programmable quantum spin glass simulator. Science 361, 162–165 (2018).

15/23
27. Vadlamani, S. K., Xiao, T. P. & Yablonovitch, E. Physics successfully implements Lagrange multiplier optimization.
Proc. Natl. Acad. Sci. 117, 26639–26650 (2020).
28. Geman, S. & Geman, D. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images. IEEE
Transactions on pattern analysis machine intelligence 721–741 (1984).
29. Borders, W. A. et al. Integer factorization using stochastic magnetic tunnel junctions. Nature 573, 390–393 (2019).
30. Cai, F. et al. Power-efficient combinatorial optimization using intrinsic noise in memristor hopfield neural networks.
Nat. Electron. 3, 409–418 (2020).
31. Shukla, N. et al. Synchronized charge oscillations in correlated electron systems. Sci. reports 4, 1–6 (2014).
32. Merolla, P. A. et al. A million spiking-neuron integrated circuit with a scalable communication network and interface.
Science 345, 668–673 (2014).
33. Kiraly, B., Knol, E. J., van Weerdenburg, W. M., Kappen, H. J. & Khajetoorians, A. A. An atomic Boltzmann machine
capable of self-adaption. Nat. Nanotechnol. 1–7 (2021).
34. Pierangeli, D., Marcucci, G. & Conti, C. Large-scale photonic Ising machine by spatial light modulation. Phys. Rev.
Lett. 122, 213902 (2019).
35. McMahon, P. L. et al. A fully programmable 100-spin coherent Ising machine with all-to-all connections. Science 354,
614–617 (2016).
36. Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H. & Teller, E. Equation of state calculations by fast
computing machines. The J. Chem. Phys. 21, 1087–1092 (1953).
37. Hastings, W. K. Monte Carlo sampling methods using markov chains and their applications. Biometrika 57, 97–109
(1970).
38. Swendsen, R. H. & Wang, J.-S. Replica Monte Carlo simulation of spin-glasses. Phys. Rev. Lett. 57, 2607–2609 (1986).
39. Earl, D. J. & Deem, M. W. Parallel tempering: Theory, applications, and new perspectives. Phys. Chem. Chem. Phys. 7,
3910–3916 (2005).
40. Wang, W., Machta, J. & Katzgraber, H. G. Population annealing: Theory and application in spin glasses. Phys. Rev. E
92, 063307 (2015).
41. Zhu, Z., Ochoa, A. J. & Katzgraber, H. G. Efficient cluster algorithm for spin glasses in any space dimension. Phys. Rev.
Lett. 115, 077201 (2015).
42. Kirkpatrick, S., Gelatt, C. D. & Vecchi, M. P. Optimization by simulated annealing. Science 220, 671–680 (1983).
43. Černỳ, V. Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm. J. optimiza-
tion theory applications 45, 41–51 (1985).
44. Camsari, K. Y., Faria, R., Sutton, B. M. & Datta, S. Stochastic p-bits for invertible logic. Phys. Rev. X 7, 031014 (2017).
45. Camsari, K. Y., Sutton, B. M. & Datta, S. p-bits for probabilistic spin logic. Appl. Phys. Rev. 6, 011305 (2019).
46. Sutton, B., Camsari, K. Y., Behin-Aein, B. & Datta, S. Intrinsic optimization using stochastic nanomagnets. Sci. reports
7, 1–9 (2017).
47. Shim, Y., Jaiswal, A. & Roy, K. Ising computation based combinatorial optimization using spin-Hall effect induced
stochastic magnetization reversal. J. Appl. Phys. 121, 193902 (2017).
48. Arnalds, U. B. et al. A new look on the two-dimensional Ising model: thermal artificial spins. New J. Phys. 18, 023008
(2016).
49. Bhanja, S., Karunaratne, D., Panchumarthy, R., Rajaram, S. & Sarkar, S. Non-Boolean computing with nanomagnets
for computer vision applications. Nat. nanotechnology 11, 177–183 (2016).
50. Lee, A. et al. A thermodynamic core using voltage-controlled spin-orbit-torque magnetic tunnel junctions. Nanotech-
nology (2021).
51. Mizushima, K., Goto, H. & Sato, R. Large-scale Ising-machines composed of magnetic neurons. Appl. Phys. Lett. 111,
172406 (2017).
52. Pierangeli, D., Marcucci, G. & Conti, C. Adiabatic evolution on a spatial-photonic Ising machine. Optica 7, 1535–1543
(2020).

16/23
53. Roques-Carmes, C. et al. Heuristic recurrent algorithms for photonic Ising machines. Nat. communications 11, 1–8
(2020).
54. Bojnordi, M. N. & Ipek, E. Memristive boltzmann machine: A hardware accelerator for combinatorial optimization
and deep learning. In 2016 IEEE International Symposium on High Performance Computer Architecture (HPCA), 1–13
(IEEE, 2016).
55. Behin-Aein, B., Diep, V. & Datta, S. A building block for hardware belief networks. Sci. reports 6, 1–10 (2016).
56. Sarkar, S. & Bhanja, S. Synthesizing energy minimizing quantum-dot cellular automata circuits for vision computing.
In 5th IEEE Conference on Nanotechnology, 2005., 541–544 (IEEE, 2005).
57. Guo, S. Y. et al. A molecular computing approach to solving optimization problems via programmable microdroplet
arrays. Matter 4, 1107–1124 (2021).
58. Byrnes, T., Koyama, S., Yan, K. & Yamamoto, Y. Neural networks using two-component Bose-Einstein condensates.
Sci. reports 3, 1–7 (2013).
59. Yamaoka, M. et al. A 20k-spin Ising chip to solve combinatorial optimization problems with CMOS annealing. IEEE J.
Solid-State Circuits 51, 303–309 (2015).
60. Matsubara, S. et al. Digital annealer for high-speed solving of combinatorial optimization problems and its applications.
In 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC), 667–672 (IEEE, 2020).
61. Aramon, M. et al. Physics-inspired optimization for quadratic unconstrained problems using a digital annealer. Front.
Phys. 7, 48 (2019).
62. Su, Y., Kim, H. & Kim, B. Cim-spin: A 0.5-to-1.2 V scalable annealing processor using digital compute-in-memory
spin operators and register-based spins for combinatorial optimization problems. In 2020 IEEE International Solid-State
Circuits Conference (ISSCC), 480–482 (IEEE, 2020).
63. Yamamoto, K. et al. Statica: A 512-spin 0.25 m-weight full-digital annealing processor with a near-memory all-spin-
updates-at-once architecture for combinatorial optimization with complete spin-spin interactions. In 2020 IEEE Interna-
tional Solid-State Circuits Conference (ISSCC), 138–140 (IEEE, 2020).
64. Tsukamoto, S., Takatsu, M., Matsubara, S. & Tamura, H. An accelerator architecture for combinatorial optimization
problems. Fujitsu Sci. Tech. J 53, 8–13 (2017).
65. Matsubara, S. et al. Ising-model optimizer with parallel-trial bit-sieve engine. In Conference on Complex, Intelligent,
and Software Intensive Systems, 432–438 (Springer, 2017).
66. Yamamoto, K. et al. A time-division multiplexing Ising machine on FPGAs. In Proceedings of the 8th International
Symposium on Highly Efficient Accelerators and Reconfigurable Technologies, 1–6 (2017).
67. Patel, S., Chen, L., Canoza, P. & Salahuddin, S. Ising model optimization problems on a FPGA accelerated restricted
Boltzmann machine. arXiv preprint arXiv:2008.04436 (2020).
68. Aadit, N. A. et al. Massively parallel probabilistic computing with sparse Ising machines. arXiv:2110.02481 (2021).
69. Reuther, A. et al. Survey and benchmarking of machine learning accelerators. In 2019 IEEE high performance extreme
computing conference (HPEC), 1–9 (IEEE, 2019).
70. Arima, Y. et al. A 336-neuron, 28 K-synapse, self-learning neural network chip with branch-neuron-unit architecture.
IEEE journal solid-state circuits 26, 1637–1644 (1991).
71. Alspector, J., Allen, R. B., Jayakumar, A., Zeppenfeld, T. & Meir, R. Relaxation networks for large supervised learning
problems. In Advances in Neural Information Processing Systems, 1015–1021 (Citeseer, 1991).
72. Skubiszewski, M. An exact hardware implementation of the Boltzmann machine. In SPDP, 107–111 (Citeseer, 1992).
73. Zhu, J. & Sutton, P. FPGA implementations of neural networks–a survey of a decade of progress. In International
Conference on Field Programmable Logic and Applications, 1062–1066 (Springer, 2003).
74. Kim, S. K., McAfee, L. C., McMahon, P. L. & Olukotun, K. A highly scalable restricted Boltzmann machine FPGA
implementation. In 2009 International Conference on Field Programmable Logic and Applications, 367–372 (IEEE,
2009).
75. Kim, S. K., McMahon, P. L. & Olukotun, K. A large-scale architecture for restricted Boltzmann machines. In 2010 18th
IEEE Annual International Symposium on Field-Programmable Custom Computing Machines, 201–208 (IEEE, 2010).

17/23
 76. Le Ly, D. & Chow, P. High-performance reconfigurable hardware architecture for restricted Boltzmann machines. IEEE
Transactions on Neural Networks 21, 1780–1792 (2010).
77. Kim, L.-W., Asaad, S. & Linsker, R. A fully pipelined FPGA architecture of a factored restricted boltzmann machine
artificial neural network. ACM Transactions on Reconfigurable Technol. Syst. (TRETS) 7, 1–23 (2014).
78. Ly, D. L., Paprotski, V. & Yen, D. Neural networks on GPUs: Restricted boltzmann machines. see
[Link] (2008).
79. Zhu, Y., Zhang, Y. & Pan, Y. Large-scale restricted Boltzmann machines on single GPU. In 2013 IEEE International
Conference on Big Data, 169–174 (IEEE, 2013).
80. Okuyama, T., Sonobe, T., Kawarabayashi, K.-i. & Yamaoka, M. Binary optimization by momentum annealing. Phys.
Rev. E 100, 012111 (2019).
81. Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proc. national
academy sciences 79, 2554–2558 (1982).
82. Hopfield, J. J. & Tank, D. W. “neural” computation of decisions in optimization problems. Biol. cybernetics 52, 141–152
(1985).
83. Von Neumann, J. Non-linear capacitance or inductance switching, amplifying, and memory organs (1957). US Patent
2,815,488.
84. Wigington, R. A new concept in computing. Proc. IRE 47, 516–523 (1959).
85. Goto, E. The parametron, a digital computing element which utilizes parametric oscillation. Proc. IRE 47, 1304–1316
(1959).
86. Csaba, G. & Porod, W. Coupled oscillators for computing: A review and perspective. Appl. Phys. Rev. 7, 011302 (2020).
87. Raychowdhury, A. et al. Computing with networks of oscillatory dynamical systems. Proc. IEEE 107, 73–89 (2018).
88. Kuramoto, Y. International symposium on mathematical problems in theoretical physics. Lect. notes Phys. 30, 420
(1975).
89. Acebrón, J. A., Bonilla, L. L., Vicente, C. J. P., Ritort, F. & Spigler, R. The Kuramoto model: A simple paradigm for
synchronization phenomena. Rev. Mod. Phys. 77, 137 (2005).
90. Breakspear, M., Heitmann, S. & Daffertshofer, A. Generative models of cortical oscillations: neurobiological implica-
tions of the Kuramoto model. Front. human neuroscience 4, 190 (2010).
91. Wu, C. W. & Chua, L. O. Application of graph theory to the synchronization in an array of coupled nonlinear oscillators.
IEEE Transactions on Circuits Syst. I: Fundamental theory applications 42, 494–497 (1995).
92. Wu, C. W. Graph coloring via synchronization of coupled oscillators. IEEE Transactions on Circuits Syst. I: Fundamen-
tal Theory Appl. 45, 974–978 (1998).
93. Wu, J., Jiao, L., Li, R. & Chen, W. Clustering dynamics of nonlinear oscillator network: Application to graph coloring
problem. Phys. D: Nonlinear Phenom. 240, 1972–1978 (2011).
94. Kalinin, K. P. & Berloff, N. G. Global optimization of spin Hamiltonians with gain-dissipative systems. Sci. Reports 8,
1–9 (2018).
95. Wang, T. & Roychowdhury, J. Oim: Oscillator-based Ising machines for solving combinatorial optimisation problems.
In International Conference on Unconventional Computation and Natural Computation, 232–256 (Springer, 2019).
96. Afoakwa, R., Zhang, Y., Vengalam, U. K. R., Ignjatovic, Z. & Huang, M. Brim: Bistable resistively-coupled Ising ma-
chine. In 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 749–760 (IEEE,
2021).
97. McGoldrick, B. C., Sun, J. Z. & Liu, L. Ising machine based on electrically coupled spin Hall nano-oscillators. arXiv
preprint arXiv:2110.08885 (2021).
98. Albertsson, D. I. et al. Ultrafast Ising machines using spin torque nano-oscillators. Appl. Phys. Lett. 118, 112404 (2021).
99. Chou, J., Bramhavar, S., Ghosh, S. & Herzog, W. Analog coupled oscillator based weighted Ising machine. Sci. Reports
9, 1–10 (2019).
100. Xiao, T. P. Optoelectronics for refrigeration and analog circuits for combinatorial optimization. Ph.D. thesis, UC
Berkeley (2019).

18/23
101. Saito, K., Aono, M. & Kasai, S. Amoeba-inspired analog electronic computing system integrating resistance crossbar
for solving the travelling salesman problem. Sci. reports 10, 1–9 (2020).
102. Parihar, A., Shukla, N., Jerry, M., Datta, S. & Raychowdhury, A. Vertex coloring of graphs via phase dynamics of
coupled oscillatory networks. Sci. reports 7, 1–11 (2017).
103. Dutta, S. et al. An Ising Hamiltonian solver using stochastic phase-transition nano-oscillators. arXiv:2007.12331 (2020).
104. Zahedinejad, M. et al. Two-dimensional mutually synchronized spin Hall nano-oscillator arrays for neuromorphic
computing. Nat. nanotechnology 15, 47–52 (2020).
105. Houshang, A. et al. A spin Hall Ising machine. arXiv:2006.02236 (2020).
106. Mallick, A. et al. Using synchronized oscillators to compute the maximum independent set. Nat. Commun. 11, 1–7
(2020).
107. Bashar, M. K. et al. Experimental demonstration of a reconfigurable coupled oscillator platform to solve the Max-Cut
problem. IEEE J. on Explor. Solid-State Comput. Devices Circuits 6, 116–121 (2020).
108. Ahmed, I., Chiu, P.-W., Moy, W. & Kim, C. H. A probabilistic compute fabric based on coupled ring oscillators for
solving combinatorial optimization problems. IEEE J. Solid-State Circuits (2021).
109. Traversa, F. L. & Di Ventra, M. Universal memcomputing machines. IEEE transactions on neural networks learning
systems 26, 2702–2715 (2015).
110. Di Ventra, M. & Traversa, F. L. Perspective: Memcomputing: Leveraging memory and physics to compute efficiently.
J. Appl. Phys. 123, 180901 (2018).
111. Sheldon, F., Traversa, F. L. & Di Ventra, M. Taming a nonconvex landscape with dynamical long-range order: Mem-
computing Ising benchmarks. Phys. Rev. E 100, 053311 (2019).
112. Aiken, J. & Traversa, F. L. Memcomputing for accelerated optimization. arXiv preprint arXiv:2003.10644 (2020).
113. Utsunomiya, S., Takata, K. & Yamamoto, Y. Mapping of Ising models onto injection-locked laser systems. Opt. express
19, 18091–18108 (2011).
114. Wang, Z., Marandi, A., Wen, K., Byer, R. L. & Yamamoto, Y. Coherent Ising machine based on degenerate optical
parametric oscillators. Phys. Rev. A 88, 063853 (2013).
115. Marandi, A., Wang, Z., Takata, K., Byer, R. L. & Yamamoto, Y. Network of time-multiplexed optical parametric
oscillators as a coherent Ising machine. Nat. Photonics 8, 937–942 (2014).
116. Inagaki, T. et al. A coherent Ising machine for 2000-node optimization problems. Science 354, 603–606 (2016).
117. Yamamoto, Y. et al. Coherent Ising machines—optical neural networks operating at the quantum limit. npj Quantum
Inf. 3, 1–15 (2017).
118. Leleu, T., Yamamoto, Y., McMahon, P. L. & Aihara, K. Destabilization of local minima in analog spin systems by
correction of amplitude heterogeneity. Phys. Rev. Lett. 122, 040607 (2019).
119. Hamerly, R. et al. Experimental investigation of performance differences between coherent Ising machines and a quan-
tum annealer. Sci. advances 5, eaau0823 (2019).
120. Yamamoto, Y., Leleu, T., Ganguli, S. & Mabuchi, H. Coherent Ising machines—quantum optics and neural network
perspectives. Appl. Phys. Lett. 117, 160501 (2020).
121. Honjo, T. et al. 100,000-spin coherent Ising machine. Sci. Adv. 7, eabh0952 (2021).
122. Marandi, A., Leindecker, N. C., Vodopyanov, K. L. & Byer, R. L. All-optical quantum random bit generation from
intrinsically binary phase of parametric oscillators. Opt. express 20, 19322–19330 (2012).
123. Kalinin, K. P. & Berloff, N. G. Networks of non-equilibrium condensates for global optimization. New J. Phys. 20,
113023 (2018).
124. Takata, K. et al. A 16-bit coherent Ising machine for one-dimensional ring and cubic graph problems. Sci. Reports 6,
1–7 (2016).
125. Inagaki, T. et al. Large-scale Ising spin network based on degenerate optical parametric oscillators. Nat. Photonics 10,
415–419 (2016).
126. Okawachi, Y. et al. Demonstration of chip-based coupled degenerate optical parametric oscillators for realizing a
nanophotonic spin-glass. Nat. communications 11, 1–7 (2020).

19/23
127. Goto, H. Bifurcation-based adiabatic quantum computation with a nonlinear oscillator network. Sci. Reports 6, 1–8
(2016).
128. Tamate, S., Yamamoto, Y., Marandi, A., McMahon, P. & Utsunomiya, S. Simulating the classical XY model with a laser
network. arXiv:1608.00358 (2016).
129. Babaeian, M. et al. A single shot coherent Ising machine based on a network of injection-locked multicore fiber lasers.
Nat. communications 10, 1–11 (2019).
130. Parto, M., Hayenga, W., Marandi, A., Christodoulides, D. N. & Khajavikhan, M. Realizing spin Hamiltonians in
nanoscale active photonic lattices. Nat. materials 19, 725–731 (2020).
131. Böhm, F., Verschaffelt, G. & Van der Sande, G. A poor man’s coherent Ising machine based on opto-electronic feedback
systems for solving optimization problems. Nat. communications 10, 1–9 (2019).
132. Lagoudakis, P. G. & Berloff, N. G. A polariton graph simulator. New J. Phys. 19, 125008 (2017).
133. Berloff, N. G. et al. Realizing the classical XY Hamiltonian in polariton simulators. Nat. materials 16, 1120–1126
(2017).
134. Kalinin, K. P. & Berloff, N. G. Simulating Ising and n-state planar Potts models and external fields with nonequilibrium
condensates. Phys. Rev. Lett. 121, 235302 (2018).
135. Kyriienko, O., Sigurdsson, H. & Liew, T. C. H. Probabilistic solving of NP-hard problems with bistable nonlinear optical
networks. Phys. Rev. B 99, 195301 (2019).
136. Mahboob, I., Okamoto, H. & Yamaguchi, H. An electromechanical Ising Hamiltonian. Sci. advances 2, e1600236
(2016).
137. Tezak, N. et al. Integrated coherent Ising machines based on self-phase modulation in microring resonators. IEEE J. Sel.
Top. Quantum Electron. 26, 1–15 (2019).
138. Bernaschi, M., Billoire, A., Maiorano, A., Parisi, G. & Ricci-Tersenghi, F. Strong ergodicity breaking in aging of
mean-field spin glasses. Proc. Natl. Acad. Sci. 117, 17522–17527 (2020).
139. Ercsey-Ravasz, M. & Toroczkai, Z. Optimization hardness as transient chaos in an analog approach to constraint satis-
faction. Nat. physics 7, 966–970 (2011).
140. Molnár, B., Molnár, F., Varga, M., Toroczkai, Z. & Ercsey-Ravasz, M. A continuous-time MaxSAT solver with high
analog performance. Nat. communications 9, 1–12 (2018).
141. Leleu, T. et al. Chaotic amplitude control for neuromorphic Ising machine in silico. arXiv: 2009.04084 (2020).
142. Yin, X. et al. Efficient analog circuits for Boolean satisfiability. IEEE Transactions on Very Large Scale Integration
(VLSI) Syst. 26, 155–167 (2017).
143. Elser, V., Rankenburg, I. & Thibault, P. Searching with iterated maps. Proc. Natl. Acad. Sci. 104, 418–423 (2007).
144. Albash, T. & Lidar, D. A. Adiabatic quantum computation. Rev. Mod. Phys. 90, 015002 (2018).
145. Das, A. & Chakrabarti, B. K. Colloquium: Quantum annealing and analog quantum computation. Rev. Mod. Phys. 80,
1061–1081 (2008).
146. Apolloni, B., Carvalho, C. & De Falco, D. Quantum stochastic optimization. Stoch. Process. their Appl. 33, 233–244
(1989).
147. Kadowaki, T. & Nishimori, H. Quantum annealing in the transverse Ising model. Phys. Rev. E 58, 5355–5363 (1998).
148. Farhi, E., Goldstone, J., Gutmann, S. & Sipser, M. Quantum computation by adiabatic evolution. arXiv preprint quant-
ph/0001106 (2000).
149. Roland, J. & Cerf, N. J. Quantum search by local adiabatic evolution. Phys. Rev. A 65, 042308 (2002).
150. Amin, M. Effect of local minima on adiabatic quantum optimization. Phys. Rev. Lett. 100, 130503 (2008).
151. Schaller, G., Mostame, S. & Schützhold, R. General error estimate for adiabatic quantum computing. Phys. Rev. A 73,
062307 (2006).
152. Lidar, D. A., Rezakhani, A. T. & Hamma, A. Adiabatic approximation with exponential accuracy for many-body systems
and quantum computation. J. Math. Phys. 50, 102106 (2009).
153. Katzgraber, H. G., Hamze, F., Zhu, Z., Ochoa, A. J. & Munoz-Bauza, H. Seeking quantum speedup through spin glasses:
The good, the bad, and the ugly. Phys. Rev. X 5, 031026 (2015).

20/23
154. Muthukrishnan, S., Albash, T. & Lidar, D. A. Tunneling and speedup in quantum optimization for permutation-
symmetric problems. Phys. Rev. X 6, 031010 (2016).
155. Albash, T. & Lidar, D. A. Demonstration of a scaling advantage for a quantum annealer over simulated annealing. Phys.
Rev. X 8, 031016 (2018).
156. Denchev, V. S. et al. What is the computational value of finite-range tunneling? Phys. Rev. X 6, 031015 (2016).
157. Boixo, S. et al. Computational multiqubit tunnelling in programmable quantum annealers. Nat. communications 7, 1–7
(2016).
158. Cerezo, M. et al. Variational quantum algorithms. Nat. Rev. Phys. 1–20 (2021).
159. Preskill, J. Quantum computing in the NISQ era and beyond. Quantum 2, 79 (2018).
160. Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization algorithm. arXiv preprint arXiv:1411.4028
(2014).
161. Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization algorithm applied to a bounded occurrence
constraint problem. arXiv:1412.6062 (2014).
162. Zhou, L., Wang, S.-T., Choi, S., Pichler, H. & Lukin, M. D. Quantum approximate optimization algorithm: Performance,
mechanism, and implementation on near-term devices. Phys. Rev. X 10, 021067 (2020).
163. Guerreschi, G. G. & Smelyanskiy, M. Practical optimization for hybrid quantum-classical algorithms. arXiv preprint
arXiv:1701.01450 (2017).
164. Khairy, S., Shaydulin, R., Cincio, L., Alexeev, Y. & Balaprakash, P. Learning to optimize variational quantum circuits
to solve combinatorial problems. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, 2367–2375
(2020).
165. Pagano, G. et al. Quantum approximate optimization of the long-range Ising model with a trapped-ion quantum simulator.
Proc. Natl. Acad. Sci. 117, 25396–25401 (2020).
166. Farhi, E. & Harrow, A. W. Quantum supremacy through the quantum approximate optimization algorithm. arXiv
preprint arXiv:1602.07674 (2016).
167. Harrigan, M. P. et al. Quantum approximate optimization of non-planar graph problems on a planar superconducting
processor. Nat. physics 1–5 (2021).
168. Qiang, X. et al. Large-scale silicon quantum photonics implementing arbitrary two-qubit processing. Nat. photonics 12,
534–539 (2018).
169. Ozaeta, A., van Dam, W. & McMahon, P. L. Expectation values from the single-layer quantum approximate optimization
algorithm on Ising problems. arXiv:2012.03421 (2020).
170. Sanders, Y. R. et al. Compilation of fault-tolerant quantum heuristics for combinatorial optimization. PRX Quantum 1,
020312 (2020).
171. Somma, R. D., Boixo, S., Barnum, H. & Knill, E. Quantum simulations of classical annealing processes. Phys. Rev.
Lett. 101, 130504 (2008).
172. Boixo, S., Ortiz, G. & Somma, R. Fast quantum methods for optimization. The Eur. Phys. J. Special Top. 224, 35–49
(2015).
173. Lemieux, J., Heim, B., Poulin, D., Svore, K. & Troyer, M. Efficient quantum walk circuits for Metropolis-Hastings
algorithm. Quantum 4, 287 (2020).
174. Bapst, V. & Semerjian, G. Thermal, quantum and simulated quantum annealing: Analytical comparisons for simple
models. In Journal of Physics: Conference Series, vol. 473, 012011 (IOP Publishing, 2013).
175. Das, A. & Chakrabarti, B. K. Quantum annealing and related optimization methods, vol. 679 (Springer Science &
Business Media, 2005).
176. Crosson, E. & Harrow, A. W. Simulated quantum annealing can be exponentially faster than classical simulated anneal-
ing. In 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 714–723 (IEEE, 2016).
177. Andriyash, E. & Amin, M. H. Can quantum Monte Carlo simulate quantum annealing? arXiv preprint arXiv:1703.09277
(2017).
178. King, A. D., Bernoudy, W., King, J., Berkley, A. J. & Lanting, T. Emulating the coherent Ising machine with a mean-field
algorithm. arXiv preprint arXiv:1806.08422 (2018).

21/23
179. Tiunov, E. S., Ulanov, A. E. & Lvovsky, A. Annealing by simulating the coherent Ising machine. Opt. express 27,
10288–10295 (2019).
180. Goto, H., Tatsumura, K. & Dixon, A. R. Combinatorial optimization by simulating adiabatic bifurcations in nonlinear
Hamiltonian systems. Sci. advances 5, eaav2372 (2019).
181. Tatsumura, K., Dixon, A. R. & Goto, H. FPGA-based simulated bifurcation machine. In 2019 29th International
Conference on Field Programmable Logic and Applications (FPL), 59–66 (IEEE, 2019).
182. Goto, H. et al. High-performance combinatorial optimization based on classical mechanics. Sci. Adv. 7, eabe7953
(2021).
183. Tatsumura, K., Yamasaki, M. & Goto, H. Scaling out Ising machines using a multi-chip architecture for simulated
bifurcation. Nat. Electron. 4, 208–217 (2021).
184. Orús, R. Tensor networks for complex quantum systems. Nat. Rev. Phys. 1, 538–550 (2019).
185. Alcazar, J. & Perdomo-Ortiz, A. Enhancing combinatorial optimization with quantum generative models. arXiv preprint
arXiv:2101.06250 (2021).
186. Mugel, S. et al. Dynamic portfolio optimization with real datasets using quantum processors and quantum-inspired
tensor networks. arXiv preprint arXiv:2007.00017 (2020).
187. Mohseni, N., Navarrete-Benlloch, C., Byrnes, T. & Marquardt, F. Deep recurrent networks predicting the gap evolution
in adiabatic quantum computing. arXiv preprint arXiv:2109.08492 (2021).
188. Bojesen, T. A. Policy-guided Monte Carlo: Reinforcement-learning Markov chain dynamics. Phys. Rev. E 98, 063303
(2018).
189. Huang, L. & Wang, L. Accelerated Monte Carlo simulations with restricted Boltzmann machines. Phys. Rev. B 95,
035105 (2017).
190. Bello, I., Pham, H., Le, Q. V., Norouzi, M. & Bengio, S. Neural combinatorial optimization with reinforcement learning.
arXiv preprint arXiv:1611.09940 (2016).
191. Dai, H., Khalil, E. B., Zhang, Y., Dilkina, B. & Song, L. Learning combinatorial optimization algorithms over graphs.
arXiv preprint arXiv:1704.01665 (2017).
192. Zhou, J. et al. Graph neural networks: A review of methods and applications. AI Open 1, 57–81 (2020).
193. Dwivedi, V. P., Joshi, C. K., Laurent, T., Bengio, Y. & Bresson, X. Benchmarking graph neural networks. arXiv preprint
arXiv:2003.00982 (2020).
194. Schuetz, M. J., Brubaker, J. K. & Katzgraber, H. G. Combinatorial optimization with physics-inspired graph neural
networks. arXiv preprint arXiv:2107.01188 (2021).
195. Vinyals, O., Fortunato, M. & Jaitly, N. Pointer networks. arXiv preprint arXiv:1506.03134 (2015).
196. Shor, P. W. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer. SIAM
review 41, 303–332 (1999).
197. Arora, S. & Barak, B. Computational complexity: a modern approach (Cambridge University Press, 2009).
198. Aaronson, S. BQP and the polynomial hierarchy. In Proceedings of the forty-second ACM symposium on Theory of
computing, 141–150 (2010).
199. Papadimitriou, C. H. & Yannakakis, M. Optimization, approximation, and complexity classes. J. computer system
sciences 43, 425–440 (1991).
200. Goemans, M. X. & Williamson, D. P. Improved approximation algorithms for maximum cut and satisfiability problems
using semidefinite programming. J. ACM (JACM) 42, 1115–1145 (1995).
201. Håstad, J. Some optimal inapproximability results. J. ACM (JACM) 48, 798–859 (2001).
202. Kowalsky, M., Albash, T., Hen, I. & Lidar, D. A. 3-regular 3-XORSAT planted solutions benchmark of classical and
quantum heuristic optimizers. arXiv preprint arXiv:2103.08464 (2021).
203. Mandra, S. & Katzgraber, H. G. A deceptive step towards quantum speedup detection. Quantum Sci. Technol. 3, 04LT01
(2018).
204. Mukai, H., Tomonaga, A. & Tsai, J.-S. Superconducting quantum annealing architecture with LC resonators. J. Phys.
Soc. Jpn. 88, 061011 (2019).

22/23
205. Onodera, T., Ng, E. & McMahon, P. L. A quantum annealer with fully programmable all-to-all coupling via Floquet
engineering. npj Quantum Inf. 6, 1–10 (2020).
206. Lechner, W., Hauke, P. & Zoller, P. A quantum annealing architecture with all-to-all connectivity from local interactions.
Sci. advances 1, e1500838 (2015).
207. Puri, S., Andersen, C. K., Grimsmo, A. L. & Blais, A. Quantum annealing with all-to-all connected nonlinear oscillators.
Nat. communications 8, 1–9 (2017).
208. Aharonov, D. et al. Adiabatic quantum computation is equivalent to standard quantum computation. SIAM review 50,
755–787 (2008).
209. Oshiyama, H. & Ohzeki, M. Benchmark of quantum-inspired heuristic solvers for quadratic unconstrained binary
optimization. arXiv preprint arXiv:2104.14096 (2021).
210. Chancellor, N. Modernizing quantum annealing using local searches. New J. Phys. 19, 023024 (2017).
211. Bilbro, G. et al. Optimization by mean field annealing. In Advances in neural information processing systems, 91–98
(1989).
212. Onodera, T. et al. Nonlinear quantum behavior of ultrashort-pulse optical parametric oscillators. arXiv:1811.10583
(2018).
213. Hamze, F. & de Freitas, N. From fields to trees. arXiv preprint arXiv:1207.4149 (2012).
214. Selby, A. Efficient subgraph-based sampling of Ising-type models with frustration. arXiv preprint arXiv:1409.3934
(2014).
215. Job, J. & Lidar, D. Test-driving 1000 qubits. Quantum Sci. Technol. 3, 030501 (2018).
216. Mandra, S., Zhu, Z. & Katzgraber, H. G. Exponentially biased ground-state sampling of quantum annealing machines
with transverse-field driving Hamiltonians. Phys. Rev. Lett. 118, 070502 (2017).
217. Zhu, Z., Ochoa, A. J. & Katzgraber, H. G. Fair sampling of ground-state configurations of binary optimization problems.
Phys. Rev. E 99, 063314 (2019).
218. Albash, T. & Lidar, D. A. Adiabatic quantum computation. Rev. Mod. Phys. 90, 015002 (2018).

Acknowledgements
We thank Scott Aaronson, Tameem Albash, Helmut Katzgraber, Timothée Leleu, Sam King, Marek Narozniak, Srikrishna
Vadlamani and Thomas van Vaerenbergh for helpful discussions and comments on the manuscript. T.B. is supported by the Na-
tional Natural Science Foundation of China (62071301); NYU-ECNU Institute of Physics at NYU Shanghai; the Joint Physics
Research Institute Challenge Grant; the Science and Technology Commission of Shanghai Municipality (19XD1423000,22ZR1444600);
the NYU Shanghai Boost Fund; the China Foreign Experts Program (G2021013002L); the NYU Shanghai Major-Grants Seed
Fund. P.L.M. thanks all his collaborators on the topic of Ising machines—especially Surya Ganguli, Ryan Hamerly, Tim-
othée Leleu, Hideo Mabuchi, Alireza Marandi, Edwin Ng, Tatsuhiro Onodera, and Yoshihisa Yamamoto—for enlightening
discussions that have shaped his understanding over the years. P.L.M. acknowledges funding from NSF award CCF-1918549,
and NTT Research for their financial and technical support. P.L.M. also acknowledges membership in the CIFAR Quantum
Information Science Program as an Azrieli Global Scholar.

Author contributions
All authors contributed in compiling the results and preparing the manuscript.

Competing interests
P.L.M. declares an interest in QC Ware Corp., a company producing software for quantum computers, to which he is an
advisor. T.B. and N.M. declare no competing interests.

Author’s note
This is a preprint of a review paper that is scheduled to appear in Nature Reviews Physics. This version does not include
changes made during the editing and production process at the journal.

23/23