1 Introduction

Quantum annealing is a quantum computing paradigm that relies on quantum fluctuations to solve optimization problems [1,2,3,4,5,6,7,8,9,10]. In September 2020, D-Wave Systems has released a quantum annealer with a 5000+ qubit quantum processing unit (QPU) called Advantage [11]. This system has more than twice as many qubits as its predecessor D-Wave 2000Q and an increase in qubit connectivity from 6 to 15 by using the Pegasus topology [12,13,14,15]. High expectations have been placed on its computational power, and first independent studies have become available [16,17,18,19,20,21,22,23]. For such a rapidly developing technology, it is an important task for independent researchers to study progress and test new developments.

Conceptually, there are three classes of benchmarks for quantum annealers:

  1. (1)

    Comparison with detailed real-time simulations of quantum annealing systems based on solving the time-dependent Schrödinger equation [24, 25] or the time-dependent master equation [5, 6, 26,27,28,29,30].

  2. (2)

    Direct QPU benchmarks (including comparison with other quantum annealing systems and optimization problem solvers) for problems of intermediate size that may or may not need embeddings and solve either real-world or artificial problems [11, 31,32,33,34,35,36,37,38,39,40,41].

  3. (3)

    Benchmarks of hybrid solvers that use a combination of QPUs and CPUs or GPUs to solve large-scale application problems [12, 18, 42, 43].

The experiments reported in this paper focus on benchmarking the bare QPU performance, thus belonging to benchmarking class (2).

We assess the progress in quantum annealing technology by benchmarking both Advantage and D-Wave 2000Q with exact cover problems. The exact cover problem is an NP-complete problem [44] that has become a prominent application to study quantum annealing [45,46,47,48,49] and gate-based quantum computing [50,51,52,53,54]. In our case, the exact cover problems represent simplified aircraft scheduling scenarios. They are derived from the tail assignment problem [55] (see [53] for more information), which belongs to the class of set covering and partitioning problems that is covered in a tremendous amount of literature on both applications and solution techniques in operations research (see e.g., [56, 57]). Related aircraft assignment problems on quantum annealers have been studied in [58,59,60].

Fig. 1
figure 1

Visualization of the exact cover problem with 40 logical qubits (instance 0). a Boolean matrix A defining the exact cover problem instance (see Eq. (13)). For each of the 40 routes, the matrix A indicates all flights that are covered by this route. The exact cover problem is to find a selection of routes (i.e., a subset of rows of the matrix A) such that all 472 flights are covered exactly once (meaning that the sum of the selected rows contains only ones). For this problem, rows belonging to this solution are indicated with blue boxes. The full solution is given by the ground state \(\vert 0100001010010101000000101000100000000000\rangle \) (the rightmost qubit corresponds to route 0). (b) Coupler graph of the Ising formulation of this problem (cf. Eq. (2)), where each nonzero \(J_{ij}\) corresponds to a black line between the 40 qubits. With 711 out of all \(\left( {\begin{array}{c}40\\ 2\end{array}}\right) =780\) couplers being nonzero, this problem is almost fully connected (see also Appendix C). A distribution of the values of the Ising parameters of this problem is shown in (c) for the qubit biases \(h_i\) given by Eq. (10) and (d) for the couplers \(J_{ij}\) given by Eq. (11))

Full size image

The essence of the present exact cover problems is shown in Fig. 1a. In this case, we are given 40 flight routes. Each route contains several out of 472 flights. The task is to find a selection of flight routes such that all 472 flights are covered exactly once. On the quantum annealer, each route is represented by a qubit. If a route is to be selected, the corresponding qubit ends up in the state \(|1\rangle \) after the measurement. Eventually, each selected route shall be assigned to one airplane. We remark that solving the problem with a given set of routes is a simplification of the general case.

The difficulty of the problem for a quantum computer can be seen by the following counting argument: For 40 routes (i.e., 40 qubits), the number of possible selections is \(2^{40}\approx 10^{12}\). For 120 routes (i.e., 120 qubits), which are the largest problems that are solved in the present benchmark set, the number of selections already grows to \(10^{36}\). Hence, exact cover problems are well suited for benchmarking the Advantage and D-Wave 2000Q QPUs.

We find that Advantage outperforms D-Wave 2000Q on almost all problems in the present benchmark set up to 120 logical qubits. Advantage can embed and solve larger problems. Furthermore, the time-to-solution on Advantage is at least roughly a factor of two shorter. Advantage scores better success rates for all problems with almost all-to-all connectivity. Only some problems with a very sparse qubit connectivity have sometimes higher success rates on D-Wave 2000Q.

The remainder of this paper is structured as follows. In Sect. 2, we describe the mathematical details associated with the exact cover problems under investigation. In Sect. 3, we present and discuss the results that Advantage and D-Wave 2000Q have produced for both small-scale and large-scale exact cover problems. Section 4 contains our conclusions.

2 Methods

This section presents the mathematical details behind the problems under investigation and how they are solved on the quantum annealers. We first outline the type of optimization problems that Advantage and D-Wave 2000Q can solve, including the important distinction between physical and logical qubits. Then, we describe the tail assignment and exact cover problems under investigation, along with their formulation on the D-Wave 2000Q and Advantage QPUs.

2.1 QUBO and Ising problems

The QPUs produced by D-Wave Systems are designed to solve binary quadratic models (BQMs), i.e., quadratic optimization problems over discrete variables that can each take two different values. BQMs are typically formulated as quadratic unconstrained binary optimization (QUBO) models or Ising models:

$$\begin{aligned}&\text {QUBO}:&\min _{x_i=0,1} \left( \sum _{i\le j} x_i Q_{ij} x_j + C _1\right) , \end{aligned}$$
(1)
$$\begin{aligned}&\text {Ising}:&\min _{s_i=\pm 1} \left( \sum _{i} h_i s_i + \sum _{i<j} J_{ij} s_i s_j + C _2\right) , \end{aligned}$$
(2)

where the indices i and j range over all qubits. In the QUBO model in Eq. (1), the problem is defined by the QUBO matrix Q with values \(Q_{ij}\in \mathbb R\), and the binary problem variables are \(x_i=0,1\). In the Ising model in Eq. (2), the problem is defined by the biases \(h_i\in \mathbb R\) and the couplers \(J_{ij}\in \mathbb R\), and the problem variables are \(s_i=\pm 1\). An example distribution of \(h_i\) and \(J_{ij}\) of the problems under investigation is shown in Fig. 1c and d, respectively.

It is worth mentioning that on the D-Wave QPUs, all problems are internally converted into Ising models and (if the autoscaling feature is on) rescaled by a constant factor such that \(h_i\in [-2,2]\) and \(J_{ij}\in [-1,1]\) [61]. (Note that these ranges might be different for future QPUs.) To convert between QUBO and Ising formulation, we use the quantum annealing convention (see also [62])

$$\begin{aligned} x_i = \frac{1+s_i}{2}, \end{aligned}$$
(3)

so that \(x_i=0\) (\(x_i=1\)) maps to \(s_i=-1\) (\(s_i=1\)). Note that in the literature on gate-based quantum computing, also the alternative \(x_i=(1-s_i)/2\) is often used [63] (which would result in a change of sign for the qubit biases, \(h_i\mapsto -h_i\)).

The constants \(C _1\) and \(C _2\) in Eqs. (1) and (2) do not affect the solution of the problem. However, they can be used to shift the energy (i.e., the value of the objective function at the solution). We use it to shift the energy of the ground state to zero so that we can conveniently determine the success rate by counting all solutions with energy zero.

2.2 Physical and logical qubits

Many problems may require nonzero couplers \(Q_{ij}\) or \(J_{ij}\) between different qubits i and j that do not physically exist on the QPUs. For instance, the 40-qubit problem sketched in Fig. 1b has almost all-to-all connectivity. In this case, solving the problem on a QPU requires the concept of embedding the problem on a QPU.

Conventionally, the qubits that physically exist on a QPU are called physical qubits. On a D-Wave 2000Q QPU, the 2000+ physical qubits are connected in a Chimera topology [7]. This means that each physical qubit is connected to 6 other physical qubits on average. On an Advantage QPU, the Chimera topology has been upgraded to the Pegasus topology [11]. This means that nearly all of the 5000+ physical qubits are connected to 15 other physical qubits, increasing the connectivity by a factor of 2.5.

When problems require a larger connectivity than provided by the QPU, the effective connectivity between the qubits can be increased by combining several physical qubits into a logical qubit. The QUBO and Ising models in Eqs. (1) and (2) are typically formulated in terms of such logical qubits, and not the underlying physical qubits. The physical qubits that form a logical qubit are called a chain. To ensure that physical qubits within a chain function as a single logical qubit, the couplers \(J_{ij}\) between them are set to a reasonably large, negative value called \(\texttt {chain\_strength}\). If a chain between two qubits breaks (i.e., if the product of the Ising variables is \(s_is_j = -1\)), a penalty of \(2\times \texttt {chain\_strength}\) is added to the energy.

We define the \(\texttt {chain\_strength}\) in terms of the relative chain strength \(\mathrm {RCS\in [0,1]}\) according to

$$\begin{aligned} \texttt {chain\_strength} = \mathrm {RCS} \times \texttt {max\_strength}, \end{aligned}$$
(4)

where \(\texttt {max\_strength}=\max \left( \left\{ \left| h_i\right| \right\} \cup \left\{ \left| J_{ij}\right| \right\} \right) \) is the maximum absolute value of all \(h_i\) and \(J_{ij}\). For instance, for the problem sketched in Fig. 1, max_strength would be 119.5.

An embedding is a mapping from each logical qubit to a chain of physical qubits. An important property of embeddings are the chain lengths. In our experience, embeddings with too large chains may result in a poor quality of the solutions produced by a QPU. In Appendix C, we provide more details on the specific chains encountered in the embeddings (see Fig. 7).

The D-Wave Ocean SDK [64] provides algorithms to automatically generate qubit embeddings with a given value for \(\texttt {chain\_strength}\). Still, when using a QPU, finding and characterizing embeddings is an important step and may considerably affect the quality of the solution. For this reason, as part of the present benchmark, we systematically investigate different embeddings and relative chain strengths in Sect. 3.1.

Note that the Pegasus topology is an extension of the Chimera topology, so that a Chimera graph can be natively embedded in a Pegasus graph (see Appendix B and also [18]). We made use of this relation for the experiments presented in Appendix B. For all other experiments, the embeddings onto the Chimera and Pegasus topologies were generated independently using the D-Wave Ocean SDK.

2.3 Tail assignment problem

The problem instances considered in this work are derived from the tail assignment problem [55]. The tail assignment problem is a fundamental component of aircraft assignment problems, i.e., the problem of assigning flights to individual airplanes, identified by their tail number. The objective is to minimize the overall cost subject to certain constraints such as minimum connection times, airport curfews, maintenance, and preassigned activities. The general role of tail assignment problems in aircraft scheduling and their relation to the column generation technique [65] and the branch-and-price algorithm [66] is described in detail in [53].

We consider a simple form of the tail assignment problem given by

$$\begin{aligned}&\text {minimize}&\sum _{r=0}^{R-1} c_r x_r,&\end{aligned}$$
(5)
$$\begin{aligned}&\text {subject to}&\sum _{r=0}^{R-1} A_{rf} x_r = 1 \quad \forall f=0,\ldots ,F-1,&\end{aligned}$$
(6)

where \(r=0,\ldots ,R-1\) enumerates all flight routes, \(f=0,\ldots ,F-1\) enumerates the flights, \(x_r\in \{0,1\}\) are the Boolean problem variables with \(x_r=1\) if route r is to be selected, \(c_r\) is the cost of selecting route r, and \(A\in \{0,1\}^{R\times F}\) is the Boolean problem matrix with \(A_{rf}=1\) if flight f is contained in route r (see Fig. 1a for an example of the matrix A). Further models for the tail assignment problem can be found in [51, 53, 55, 60].

A BQM version of the tail assignment problem given by Eqs. (5) and (6) is

$$\begin{aligned}&\min _{x\in \{0,1\}^N}\left( \lambda c^Tx + \left( A^Tx - b\right) ^2\right) , \end{aligned}$$
(7)

where the number of qubits \(N=R\) is given by the number of flight routes, \(c = (c_0,\ldots ,c_{R-1})^T\) contains the costs, and \(b = (1,\ldots ,1)^T\) is an F-dimensional vector of ones. Note that the scaling factor \(\lambda \) in Eq. (7) is the inverse of the penalty multiplier that would determine the scale of the constraint \(Ax=b\) in Eq. (6). It was put in front of the objective function \(c^Tx\) so that the exact cover version of the problem corresponds to \(\lambda =0\) (see Sect. 2.4).

We obtain the QUBO formulation of the tail assignment problem by multiplying out the square in Eq. (7) and collecting all terms into the general QUBO model Eq. (1). An outline of the calculation is given in Appendix A. After doing this, we can read off the entries of the QUBO matrix as

$$\begin{aligned} Q_{ij}&= {\left\{ \begin{array}{ll} (2AA^T)_{ij} &{} (i<j) \\ (AA^T)_{ii} - (2Ab)_i + \lambda c_i &{} (i=j) \end{array}\right. }, \end{aligned}$$
(8)
$$\begin{aligned} C _1&= b^Tb. \end{aligned}$$
(9)

Note that the qubit indices \(i,j\in \{0,\ldots ,N-1\}\) correspond to the previous route index \(r\in \{0,\ldots ,R-1\}\).

Finally, the Ising formulation of the tail assignment problem is found by using Eq. (3) to replace the qubit variables \(x_i\) by spin variables \(s_i\) in the QUBO model Eq. (1). After collecting linear, quadratic, and constant terms, we find the expressions for the coefficients of the Ising model Eq. (2):

$$\begin{aligned} h_i&= \sum _j \frac{1}{2}(AA^T)_{ij} - (Ab)_i + \frac{1}{2} \lambda c_i, \end{aligned}$$
(10)
$$\begin{aligned} J_{ij}&= \frac{1}{2}(AA^T)_{ij}, \end{aligned}$$
(11)
$$\begin{aligned} C _2&= C _1 + \sum _{i< j} \frac{1}{2} (AA^T)_{ij} \nonumber \\&+ \sum _i \frac{1}{2} ((AA^T)_{ii} - (2Ab)_i + \lambda c_i). \end{aligned}$$
(12)

These expressions hold for general values of A and \(b\). (Details on the calculation can also be found in Appendix A.) A characteristic distribution of the biases \(h_i\) and the couplers \(J_{ij}\) is shown in Fig. 1c and d, respectively.

2.4 Exact cover problem

The exact cover problem is an NP-complete set partitioning problem [44]. In matrix form, it can be written as

$$\begin{aligned} \min _{x_r=0,1} \sum _{f=0}^{F-1} \left( \sum _{r=0}^{R-1}A_{rf}x_r - 1 \right) ^2, \end{aligned}$$
(13)

and its purpose can be directly understood from Fig. 1a: The selection of routes (i.e., rows) with \(x_r=1\) has to be such that each flight f in the problem matrix A is covered exactly once.

The exact cover problem corresponds to the feasibility version of the tail assignment problem given by the sole constraints in Eq. (6), without the objective function in Eq. (5). Formally, the exact cover problems are obtained by setting \(\lambda =0\) in Eq. (7), which yields Eq. (13).

Hence, the QUBO (Ising) coefficients of the exact cover problem are given by Eq. (8) (Eqs. (10) and (11)) for \(\lambda =0\) (see also [47]). Note that the explicit expressions for the constants in Eqs. (9) and (12) are not relevant for the solution of the problem on the quantum annealer. However, it is convenient to add them to the resulting energies to ensure that the energy minimum is zero, because then we can determine the success rate by counting the occurrences of samples with energy zero.

The problems have been generated using the column generation method as described in detail in [53]. Additional properties of the exact cover problems, including the number of logical couplers and physical qubits used in the generated embeddings, can be found in Appendix C.

Fig. 2
figure 2

Success rates for exact cover problems with 30–40 qubits (instance 0) as a function of the relative chain strength (bottom axis) on a D-Wave 2000Q and b Advantage with default annealing time \(20\,\mu \mathrm {s}\). The scan of the relative chain strength is repeated for 10 different, randomly generated embeddings (represented by different colors) and with 10 repetitions each to gather statistics. Markers indicate the corresponding standard deviation above and below the mean. Filled areas between the markers are guides to the eye. The curves for the success rates as a function of the RCS are representative of the other problem instances 1, 2, and 3 characterized in Appendix C

Full size image
Fig. 3
figure 3

Success rates as a function of the annealing time on a D-Wave 2000Q and b Advantage, averaged over 10 repetitions. Different markers indicate different problem instances: 0 (markers with error bars), 1 (pluses), 2 (crosses), 3 (circles). For each instance, the runs in each panel are performed with the same embeddings and relative chain strengths, characterized in Fig. 7 in Appendix C. For instance 0 (highlighted in color), these parameters correspond to the best configuration of the corresponding panel in Fig. 2. Additionally, for instance 0, the standard deviation from the 10 repetitions is indicated by the filled areas between the error bars

Full size image

3 Results

In this section, we present the benchmark results for Advantage and D-Wave 2000Q. We first consider small and intermediate exact cover problems with 30–40 logical qubits and almost full connectivity, with a focus on comparing different embeddings and annealing times. Then, we proceed to large exact cover problems with up to 120 logical qubits, with a focus on the success rate and the time that it takes the QPUs to solve the problems. We note that all considered problem sizes, although large for current quantum annealers, are still small for classical solvers and can be solved within a few minutes using, for instance, a linear programming solver. For the sake of completeness, we also present results for tail assignment problems with \(\lambda \ne 0\) (cf. Eq. (7)) in Appendix D.

The experiments reported in this section were performed in August and September 2020 using the solver DW_2000Q_VFYC_6 for D-Wave 2000Q and the solvers Advantage_beta and Advantage_system1.1 for Advantage. We stress that the goal of the present study was to compare the different QPUs under equivalent conditions. For this reason, we did not make use of any post-processing features to solely improve the quality of the results. Unless otherwise stated, all QPU settings were left at their default values.

3.1 Densely connected problems with 30–40 qubits

The problems with \(N=30,32,34,36,38,40\) qubits and almost full connectivity (cf. Appendix C) are exact cover problems from aircraft scheduling scenarios with 472 flights. Each qubit represents a flight route that contains some of the 472 flights (cf. Fig. 1). We remark that by construction, the ground state of each problem instance is unique and contains 9 qubits in state \(|1\rangle \). We obtain the success rate by counting the number of samples with energy zero. (Note that the value of the constant Eq. (12) can be used to shift the energy accordingly.)

3.1.1 Characterizing embeddings and chain strengths

For each problem with \(N=30,32,34,36,38,40\) qubits, we generate 10 different embeddings on both D-Wave 2000Q and Advantage. For each embedding, we scan the RCS (cf. Eq. (4)). The results are shown in Fig. 2.

We see that the results on Advantage (Fig. 2b) are generally much better than on D-Wave 2000Q, especially for larger problems. The reason for this is that the 30–40 qubit problems are almost fully connected (cf. Figs. 1b and 6). Hence, in such cases, the user can clearly profit from the much larger connectivity between qubits on the Pegasus topology.

This observation is in line with the fact that the chains on Advantage for the same problem are much shorter (see Fig. 7b). We also see that for increasing problem size, generating multiple embeddings and tuning the chain strengths is crucial to obtain good results when using the bare QPUs.

3.1.2 Varying the annealing time

Next, we study the influence of the annealing time on the success rate. The reason is that for a quantum annealer, the annealing time is a central parameter with an expectable influence on the success rate: From the adiabatic theorem [67], comparably long annealing times are expected to lead to reasonable success rates. Given that the system is not completely isolated from an environment, however, long annealing times may lead to an increasing influence of noise on the system and thus in fact degrade the performance. Therefore, experiments beyond the default annealing time are vital for reliable benchmarks of quantum annealers.

We first select the best embedding and relative chain strength for each problem, based on the results from the previous section. For problem instance 0, for example, this configuration corresponds to the position of the peaks in each panel in Fig. 2. For the selected configuration, we then replace the default annealing time of \(20\,\mu \mathrm {s}\) by 20 different, logarithmically spaced annealing times in the QPU annealing time range \([1\,{\mu }{\mathrm s}, 2000\,{\mu }{\mathrm s}]\). The results for the 30–40 qubit problems and all four problem instances are shown in Fig. 3.

Comparing D-Wave 2000Q and Advantage, we can make the following observations: First, Advantage reaches higher maximum success rates, typically for the longest annealing time. Second, the success rates on Advantage are already reasonable for short annealing times. This means that Advantage is typically faster than D-Wave 2000Q (see also the comparison of QPU access times in Fig. 4b in the following section). Finally, the fluctuations over 10 repetitions are smaller on Advantage (see the filled areas in Fig. 3). We note that this observation also holds for the three instances whose statistics are not indicated in Fig. 3 (gray markers). Thus, we conclude that Advantage shows a demonstrable advantage over D-Wave 2000Q.

3.2 Large problems with 50–120 qubits

In this section, we consider exact cover problems of a larger size with 50 to 120 logical qubits. The goal is to assess the performance of the QPUs for larger but more sparsely connected problem instances. For each problem size, we consider six problem instances. Each corresponds to an aircraft scheduling problem of the type sketched in Fig. 1a with 535 flights. As before, the ground state of these exact cover problems is unique and known. It has 40 qubits in state \(|1\rangle \).

These large problems require many more physical qubits so that their embedded versions can occupy a large part of the QPUs. For this reason, the success rates may be smaller, and especially D-Wave 2000Q may not be able to solve the largest problems anymore.

Fig. 4
figure 4

Solution of exact cover problems with 50–120 logical qubits on both D-Wave 2000Q (yellow) and Advantage (blue). Shown are (a) the success rate as a function of the number of logical qubits and (b) the QPU access time (which mainly consists of programming, annealing, and readout times [68]) . Different markers indicate different problem instances: 0 (crosses), 1 (squares), 2 (circles), 3 (up-pointing triangles), 4 (down-pointing triangles), 5 (diamonds). The embedding and the relative chain strength for these runs was selected using the same procedure as in Sect. 3.1

Full size image

Indeed, as Fig. 4a shows, only Advantage can solve five out of the six 120 qubit instances. The success rates for the 80 and 100 qubit instances are comparable for D-Wave 2000Q and Advantage. Only for the 50 and 60 qubit instances, we see that D-Wave 2000Q yields sometimes better success rates than Advantage.

The reason for this can be understood by studying the number of couplers required for these problems: As shown in Fig. 6 in Appendix C, the 50 and 60 qubit problems need comparably little couplers, so they have a sparser connectivity. Therefore, many of the physical couplers present in the Pegasus topology on Advantage are not required. In other words, the problems can already be embedded well enough on the Chimera topology. This observation is in line with a similar observation for Advantage reported in [18].

We conjecture that in this case, the additional unused connections between qubits on Advantage may disturb the annealing process (additional experimental evidence for this conjecture is given in Appendix B), because even if they are not used, they still exist physically. The reason for our conjecture lies in a well-known physical mechanism: If two subsystems of a larger system are coupled, the system’s Hamiltonian contains a corresponding interaction term. For flux qubits manufactured by D-Wave Systems, this interaction is often mediated by tunable interqubit inductive couplers [5]. Such an interaction is always on, due to the existence of a physical link between the subsystems, even if its scale is very weak or if the coupler is “programmed to zero.” During the time evolution (which may take comparably long for an annealing process), the wave function of the system typically develops an error over time, which may have a detrimental impact on the function of a quantum computer. The effect can be easily observed for a system of coupled spins, for example the NMR quantum computing system [69]. Furthermore, it can also be found in superconducting qubit systems; see, e.g., the cross talk experiment in [70], the performance of two-qubit gates mediated by a coupler such as the resonator in [71] or the flux-tunable transmon in [72], or the theoretical study on ZZ couplings in [73]. We leave a detailed simulation of the present system for future research.

Finally, the large problems may require more time on the QPUs, so that they are well suited to compare QPU run times between D-Wave 2000Q and Advantage. Unless mentioned otherwise, the same number of reads and annealing times have been used for D-Wave 2000Q and Advantage to ensure a fair comparison. The timing results are shown in Fig. 4b.

For 50–80 qubits, the annealing time is less than or equal to \(20\,{\mu }{\mathrm s}\), so for 1000 reads it does not make up a significant fraction of the QPU access time. Under these conditions, we see that D-Wave 2000Q still needs at least twice as long to solve the problem. Thus, we infer that it is a speedup in programming and readout times [68] that make Advantage faster than D-Wave 2000Q.

For problems with 100 and 120, the annealing time needs to be significantly increased to find the ground state, which is visible in the QPU access times. For instance, problem instance 5 for 100 qubits on D-Wave 2000Q (the single yellow diamond at 100 qubits in Fig. 4b) corresponds to an annealing time of \(200\,{\mu }{\mathrm s}\). The same annealing time is required for problem instances 1, 2, 3, and 5 for 120 qubits on Advantage (the blue cluster at 120 qubits in Fig. 4b). Still, Advantage solves these problems faster than D-Wave 2000Q solves the corresponding 100 qubit problem. Only instance 4 for 120 qubits stands out (the blue down-pointing triangle in Fig. 4b): Here, an annealing time of \(2000\,{\mu }{\mathrm s}\) was required to find a solution. In order not to exceed the maximum run time on Advantage, the number of reads was reduced to 400. Thus, \(400\times 2000\,{\mu }{\mathrm s}=0.8\,\mathrm s\) makes up the largest part of the QPU access time in this case.

4 Conclusion

In this paper, we have benchmarked the performance of the 2000+ qubit quantum annealer D-Wave 2000Q and the 5000+ qubit quantum annealer Advantage. The benchmark suite consists of intermediate and large exact cover problems from aircraft scheduling scenarios with both sparse and dense logical qubit connectivity.

We observed a considerable increase in performance on Advantage. First, Advantage was able to solve exact cover problems with up to 120 logical qubits that were unsolvable on D-Wave 2000Q. Second, the success rates produced by Advantage were almost always higher. Third, Advantage is approximately twice as fast as D-Wave 2000Q in terms of both programming and readout times. Additionally, the required annealing times to solve a problem on Advantage are often shorter. Finally, the fluctuations in success rates over several repetitions on Advantage were smaller.

A large part of the increase in performance can be attributed not only to the larger number of physical qubits, but rather to the increase in qubit connectivity: Every qubit in the Pegasus topology is connected to 15 other qubits, as compared to 6 other qubits in the Chimera topology used in D-Wave 2000Q. We observed chain lengths in the embeddings that were roughly smaller by a factor of two. We could only observe better performance on D-Wave 2000Q for problems with very sparse qubit connectivity or when using the same embedding on D-Wave 2000Q and Advantage. The sparse problems could already be well embedded on the Chimera topology. We conclude that increasing the number of couplers does not necessarily improve the performance of a quantum annealer. Instead, whether an improvement or a degradation is observed depends on the connectivity of each problem instance. On the one hand, increasing the number of couplers allows to embed bigger problem instances, and it may also reduce the required chain length and thus the chain strength. On the other hand, we observed that if only a few couplers are needed, it may be better to use a processor with lower connectivity (see also [18]).

We also observed that, when the same embedding is used on the D-Wave 2000Q processor with connectivity 6 and the Advantage processor with connectivity 15, the results on the D-Wave 2000Q processor are significantly better (see Appendix B), suggesting that the additional couplers introduce additional noise if they are not needed. In particular, this supports the conclusion that for our experiments, the better embeddings (and not a reduced noise level) are the reason for the improved performance of Advantage over D-Wave 2000Q. Thus, a higher connectivity only yields an advantage if the embedding of the problem can actually be improved. An interesting project for future work would be a systematic study of the question whether, for a given embedding, there is some “minimal improvement” required to overcome the influence of additional unused couplers to also yield an improvement in the success probability.

When using the bare QPUs, it is essential to scan several embeddings and chain strengths to find optimal results. Furthermore, it is important to tune the annealing time. We note that besides the bare QPUs, we have also submitted all exact cover problems of our benchmark set to D-Wave’s hybrid solver services, which use a combination of QPUs and classical solvers to solve much larger problems [12]. All exact cover problems could be solved by the hybrid solvers hybrid_v1 (using D-Wave 2000Q) and hybrid_binary_quadratic_model_version2 (using Advantage) on September 14, 2020. See [18] for more detailed benchmarks of the hybrid solvers with problems of up to 12000 variables.

Our benchmark study confirms the consistent increase in both size and performance of quantum annealers over the past years. For the future, it is an interesting question whether D-Wave Systems will prove capable of keeping up the steep progress of doubling qubit numbers and increasing performance and qubit connectivity at the same time.