Comparison

Quantum annealing vs simulated annealing vs simulated bifurcation: what the benchmarks show

Simulated annealing is a classical Monte Carlo heuristic, quantum annealing does a similar search on quantum hardware with tunnelling in place of heat, and simulated bifurcation is a classical solver that integrates oscillator dynamics in parallel. All three take the same Ising or QUBO problem and all three are heuristics. Independent benchmarks rank them differently on different problem classes, none shows one solver winning everywhere, and a well-tuned classical method is usually the one to beat. Which solver leads depends on the problem class, so the test that settles it is the one on your own instances, with the same time budget for each.

Updated

Six solvers that take the same QUBO

Each of these accepts a QUBO, or its Ising form, and returns low-energy bitstrings. They differ in how they search and in what they run on.

SolverHow it searchesRuns on
Simulated annealing (SA)Random bit flips, accepted with a probability that falls as a temperature is lowered, so the search can climb out of local minima early and settles late[1]Any CPU or GPU; free open-source codes
Quantum annealing (QA)A transverse field that is turned down while the problem is turned up; quantum fluctuations take the place of heat[2]Superconducting annealers from D-Wave, on a sparse qubit graph: degree 15[8] on Pegasus, so dense problems need embedding
Simulated bifurcation (SB)Classical oscillators whose potentials split into two wells, integrated with every variable updated at once[3]FPGAs and GPUs (Toshiba SQBM+); all-to-all, up to 10610^{6} variables[13] in the version benchmarked in 2022
Digital annealer (DA)Simulated annealing with parallel trial flips and an escape mechanism, later with parallel tempering[5]Fujitsu's custom CMOS chip; fully connected, up to 8,192 variables[13] in the 2022 benchmark
Coherent Ising machine (CIM)Pulses of light in a network of optical parametric oscillators, each settling into one of two phases, coupled by measurement and feedback[6]Laboratory optical machines; 2,000 nodes[6] fully connected in 2016
QAOAAlternating layers of gates for the cost and a transverse-field mixer, with angles tuned by a classical optimiser[7]Gate-based quantum computers and simulators; the one eigenkit implements

Quantum annealing vs simulated annealing: the difference

Both methods start in a state that explores freely and end in one that is stuck in a low-energy configuration. Simulated annealing does it with heat: at high temperature almost every flip is accepted, and as the temperature falls only flips that lower the energy survive[1]. Quantum annealing does it with a transverse field: at the start the field dominates and the state is an equal superposition of all bitstrings, and as it is turned down the problem's energy takes over. Kadowaki and Nishimori proposed it as a variant of simulated annealing and found, in small simulations with the same schedule, that it reached the ground state with a higher probability in almost all cases[2].

The two escape local minima differently. Heat lets the search climb over an energy barrier, with a cost that grows with the barrier's height. Tunnelling lets a quantum state pass through a barrier, with a cost that grows with its width and height together, so it helps most where barriers are tall and narrow. Which kind of barrier a real problem has is rarely known in advance.

The comparison is also between kinds of things. Simulated annealing is an algorithm that runs on any computer. Quantum annealing usually means a physical device, with noise, limited precision and a fixed qubit graph; its classical emulation, simulated quantum annealing by quantum Monte Carlo, is a third thing again. Simulated bifurcation is neither: it is a deterministic simulation of classical oscillators, whose only randomness is the starting point.

Quantum annealing against classical methods, study by study

StudySet-upFinding
Rønnow et al. 2014D-Wave Two with up to 503 qubits[9]; random spin glasses; against simulated classical and quantum annealingNo evidence of quantum speed-up over the whole data set; inconclusive on subsets of instances
Denchev et al. 2016D-Wave 2X; problems crafted to have tall, narrow barriers; 945 variables[10]About 10810^{8} times[10] faster than simulated annealing on one core to 99% success. The authors note that other classical heuristics solve most such instances in comparable time
Hamerly et al. 2019D-Wave 2000Q against coherent Ising machines; SK spin glasses and MaxCutThe annealer wins on degree-3 MaxCut; on dense problems it falls behind by several orders of magnitude[11] in time to solution above 50 vertices
Kowalsky et al. 2022D-Wave Advantage1.1 and five classical solvers; 3-regular 3-XORSATThe largest scaling exponent of all, 0.08[12] (see the next table)
Munoz-Bauza and Lidar 2025D-Wave Advantage with an error-suppression code, over 1,300 logical qubits[14]; 2D spin glassesA scaling advantage over parallel tempering with cluster moves for approximate solutions, at an optimality gap of at least 1.0%[14]

Each line is one problem class on one generation of hardware. The 2025 result is the first scaling advantage reported against what its authors call the top classical heuristic for this setting, and it is for approximate, not exact, optimisation.

Kowalsky et al. 2022: 3-regular 3-XORSAT, an independent benchmark

The authors fit the median time to solution of each solver as 10 to the power αn+β\alpha n + \beta, where nn is the number of variables, on instances with a planted solution that is known in advance[12]. A smaller α\alpha means the time grows more slowly as problems get bigger; β\beta sets the time on small problems.

SolverHardwareScaling exponent αPrefactor β
SATonGPU (Bernaschi et al.)GPUs, on the original cubic form of the problem0.0171[12]−5.9[12]
Fujitsu Digital AnnealerCustom CMOS chip0.0185[12]−3.56[12]
Toshiba Simulated Bifurcation MachineGPUs, through a cloud marketplace0.0217[12]−2.6[12]
Parallel tempering (the authors' code)One CPU core0.0248[12]−0.97[12]
Virtual MemComputing MachineClassical simulation of a proposed machine0.025[12]−0.6[12]
D-Wave Advantage1.1Quantum annealer0.08[12]−6[12]

The solvers fall into three groups: SATonGPU and the Digital Annealer; SB, parallel tempering and MemComputing; and the quantum annealer alone. The authors warn that one problem class does not generalise, and that the GPU code's lead comes from running 327,680 replicas in parallel on a problem simple enough to suit it.

Oshiyama and Ohzeki 2022: four QUBO solvers, three problem sets

This benchmark compares the best cost each solver reaches in a given run time, rather than the time to an exact optimum[13]. The simulated annealing is the open-source D-Wave neal on one core of a desktop CPU.

Problem setD-Wave hybrid solverFujitsu Digital AnnealerToshiba SBMSimulated annealing
45 MQLib instances, 1,000 to 10,000 variables, 5-minute runs: instances won22[13]20[13]16[13]7[13]
Random NAE 3-SAT at the phase transition, 8,192 variablesWorst; reached SBM and SA only after about 1,000 sBest from 100 to 600 s; behind SBM and SA below 100 sAbout the same as SAAbout the same as SBM
Sherrington-Kirkpatrick spin glass, 8,192 variablesClose to DA, slightly betterClose to the hybrid solverClearly best, by the 100 s[13] markAs good as the hybrid solver and DA only in runs over 600 s, with about 500 s of preprocessing

On the MQLib set the gap between the solvers' best costs was often below 0.01%, which the authors call probably negligible for some applications. Ties count as a win for each solver, and network time is excluded.

Goto et al. 2019 and 2021: SB machines against other Ising machines

These are the inventors' own benchmarks on fully connected MaxCut problems. The 2021 paper notes that some of the machines it compares against were not their latest versions.

BenchmarkSB machineCompared with
K2000, 2,000 nodes: best average cutAdiabatic SB on one FPGA, 0.5 ms[3]Coherent Ising machine, 5 ms; tuned simulated annealing, 50 ms[3]
K2000: time to reach the Goemans-Williamson SDP valueWorst of 100 trials, 58 µs[3]Coherent Ising machine, best case about 70 µs[3]
100,000 nodes, all-to-all8-GPU cluster, about 10 times[3] faster than the authors' best SAAbout 1,000 times faster than a published SA code[3]
K2000: time to 99% of the best known cutBallistic SB on one FPGA, 0.26 ms[4]STATICA chip, 1.50 ms (a simulator's prediction); coherent Ising machine, 1.1 s[4]
G22: time to the best known cutDiscrete SB on one FPGA, 2.7 s[4]Simulated CIM with chaotic amplitude control on an FPGA, estimated 12 s[4]

Why no solver wins overall, and how to choose

Put the tables side by side and the rankings disagree. The Digital Annealer leads on 3-XORSAT scaling and on random satisfiability, simulated bifurcation leads on the SK spin glass, D-Wave's hybrid solver wins the most MQLib instances, and a GPU code written for one problem beats them all on that problem. The reasons are structural, not noise: the problem classes have different energy landscapes; the metrics differ (time to an exact solution, time to 99% of the best cut, best cost in five minutes); hardware generations change between papers; some benchmarks are run by the vendor; and each paper draws its own line around what is timed, such as data loading, network transfer or preprocessing.

QAOA belongs in the comparison as the gate-based way to attack the same QUBO, but it is not yet a competitor at these sizes. None of the benchmarks above includes it, exact simulation stops at a few dozen qubits, and on hardware its results fall as problems grow beyond the chip's native connectivity[15]. It is worth learning and testing for where gate-based hardware goes, not for solving a 10,000-variable QUBO today.

What would move quantum annealing up these tables is visible in the studies themselves. Its largest win came on problems crafted to have tall, narrow barriers, where tunnelling helps most[10], and the 2025 scaling advantage needed an error-suppression code on the annealer and was for approximate answers[14]. Both point to the same two conditions: a problem whose landscape suits tunnelling, and hardware whose noise is held in check. Whether a real problem meets the first is rarely known in advance, which is why it has to be measured on your own instances.

To choose, write the problem once as a QUBO and test it on your own instances. The QUBO and Ising guide covers penalties and checking a formulation. Set a baseline with simulated annealing or parallel tempering, and with a mixed-integer solver on the original constrained model; then give each specialised solver, such as simulated bifurcation, the same time budget and compare the time to reach a target cost. Kowalsky and colleagues end on the same advice: try a suite of solvers on the problem class you care about[12].

Run it on a real QPU with EigenOrbit RouterThe QAOA side of the comparison: the same QUBO as a gate-model circuit on a real QPU.In eigenkit: pip install "eigenkit[router]", then provider=RouterConfig(device=...) (docs).

Questions

What is the difference between quantum annealing and simulated annealing?

Simulated annealing is a classical algorithm: it flips bits at random and accepts uphill moves with a probability set by a temperature that is lowered over time. Quantum annealing is done on quantum hardware: a transverse field creates quantum fluctuations that are turned down while the problem's energy is turned up, so the search can tunnel through energy barriers instead of climbing over them. Both minimise the same Ising or QUBO cost.

Is quantum annealing faster than simulated annealing?

Sometimes, on problems built to favour it, and not in general. A 2016 study measured a D-Wave 2X as many orders of magnitude faster than single-core simulated annealing on crafted instances, while noting that other classical heuristics were comparable. A 2014 study on random spin glasses found no speed-up, and a 2022 benchmark found a D-Wave Advantage scaling worse than every classical solver tested on 3-XORSAT.

Is simulated bifurcation a kind of quantum annealing?

No. Simulated bifurcation is a classical algorithm from Toshiba that integrates the motion of coupled oscillators on FPGAs or GPUs. It is called quantum-inspired because it is the classical limit of a proposed quantum machine made of Kerr-nonlinear parametric oscillators, but no quantum hardware is involved.

What is a digital annealer?

The Digital Annealer is Fujitsu's custom CMOS chip for QUBO problems. Its algorithm is based on simulated annealing, with many trial flips evaluated in parallel and a mechanism to escape local minima, and later versions use parallel tempering. Despite the name it is a classical machine, not a quantum annealer.

Which QUBO solver is the best?

None is best on every problem. In independent benchmarks the Fujitsu Digital Annealer led the QUBO solvers on 3-XORSAT (a GPU code written for that problem scaled slightly better) and on random satisfiability, Toshiba's simulated bifurcation machine led on Sherrington-Kirkpatrick spin glasses, and D-Wave's hybrid solver won the most instances from the MQLib library. Test the candidates on your own instances with the same time budget.

Is QAOA the same as quantum annealing?

They are related but not the same. Both start from a transverse-field state and steer it toward the problem's low-energy states. Quantum annealing does this as a continuous physical evolution on an annealer; QAOA does it with alternating layers of gates whose angles a classical optimiser tunes, on a gate-based quantum computer or simulator.

Sources

  1. S. Kirkpatrick, C. D. Gelatt, M. P. Vecchi. Optimization by simulated annealing. Science 220, 671–680, 1983.
  2. T. Kadowaki, H. Nishimori. Quantum annealing in the transverse Ising model. Physical Review E 58, 5355, 1998. arXiv:cond-mat/9804280.
  3. H. Goto, K. Tatsumura, A. R. Dixon. Combinatorial optimization by simulating adiabatic bifurcations in nonlinear Hamiltonian systems. Science Advances 5, eaav2372, 2019.
  4. H. Goto, K. Endo, M. Suzuki, Y. Sakai, T. Kanao, Y. Hamakawa, R. Hidaka, M. Yamasaki, K. Tatsumura. High-performance combinatorial optimization based on classical mechanics. Science Advances 7, eabe7953, 2021.
  5. M. Aramon, G. Rosenberg, E. Valiante, T. Miyazawa, H. Tamura, H. G. Katzgraber. Physics-inspired optimization for quadratic unconstrained problems using a digital annealer. Frontiers in Physics 7, 48, 2019. arXiv:1806.08815.
  6. T. Inagaki, Y. Haribara, K. Igarashi, T. Sonobe, S. Tamate, T. Honjo, et al. A coherent Ising machine for 2000-node optimization problems. Science 354, 603–606, 2016.
  7. E. Farhi, J. Goldstone, S. Gutmann. A quantum approximate optimization algorithm. arXiv preprint, 2014. arXiv:1411.4028.
  8. K. Boothby, P. Bunyk, J. Raymond, A. Roy. Next-generation topology of D-Wave quantum processors. arXiv preprint, 2020. arXiv:2003.00133.
  9. T. F. Rønnow, Z. Wang, J. Job, S. Boixo, S. V. Isakov, D. Wecker, J. M. Martinis, D. A. Lidar, M. Troyer. Defining and detecting quantum speedup. Science 345, 420, 2014. arXiv:1401.2910.
  10. V. S. Denchev, S. Boixo, S. V. Isakov, N. Ding, R. Babbush, V. Smelyanskiy, J. Martinis, H. Neven. What is the computational value of finite-range tunneling? Physical Review X 6, 031015, 2016. arXiv:1512.02206.
  11. R. Hamerly, T. Inagaki, P. L. McMahon, D. Venturelli, A. Marandi, et al. Experimental investigation of performance differences between coherent Ising machines and a quantum annealer. Science Advances 5, eaau0823, 2019. arXiv:1805.05217.
  12. M. Kowalsky, T. Albash, I. Hen, D. A. Lidar. 3-regular 3-XORSAT planted solutions benchmark of classical and quantum heuristic optimizers. Quantum Science and Technology 7, 025008, 2022. arXiv:2103.08464.
  13. H. Oshiyama, M. Ohzeki. Benchmark of quantum-inspired heuristic solvers for quadratic unconstrained binary optimization. Scientific Reports 12, 2146, 2022. arXiv:2104.14096.
  14. H. Munoz-Bauza, D. A. Lidar. Scaling advantage in approximate optimization with quantum annealing. Physical Review Letters 134, 160601, 2025. arXiv:2401.07184.
  15. M. P. Harrigan, K. J. Sung, M. Neeley, et al. Quantum approximate optimization of non-planar graph problems on a planar superconducting processor. Nature Physics 17, 332–336, 2021. arXiv:2004.04197.

Run it yourself with eigenkit

eigenkit is EigenOrbit's open-source Python library: write a circuit once, run it on any simulator or QPU, and test it against a classical baseline on your own data before you invest in it or rule it out.

$ pip install eigenkit