The Future of Quantum Computing with Superconducting Qubits

Sergey Bravyi,1 Oliver Dial,1 Jay M. Gambetta,1 Darı́o Gil,1 and Zaira Nazario1
IBM Quantum, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598,
USA
(Dated: 16 September 2022)
For the first time in history, we are seeing a branching point in computing paradigms with the emergence of
quantum processing units (QPUs). Extracting the full potential of computation and realizing quantum algo-
rithms with a super-polynomial speedup will most likely require major advances in quantum error correction
technology. Meanwhile, achieving a computational advantage in the near term may be possible by combining
multiple QPUs through circuit knitting techniques, improving the quality of solutions through error suppres-
sion and mitigation, and focusing on heuristic versions of quantum algorithms with asymptotic speedups.
For this to happen, the performance of quantum computing hardware needs to improve and software needs
to seamlessly integrate quantum and classical processors together to form a new architecture that we are
arXiv:2209.06841v1 [quant-ph] 14 Sep 2022

calling quantum-centric supercomputing. Long term, we see hardware that exploits qubit connectivity in
higher than 2D topologies to realize more efficient quantum error correcting codes, modular architectures for
scaling QPUs and parallelizing workloads, and software that evolves to make the intricacies of the technology
invisible to the users and realize the goal of ubiquitous, frictionless quantum computing.

I. INTRODUCTION available along the way. For example, a quantum process-

ing unit (QPU) with two-qubit gate fidelity of 99.99%
can implement circuits with a few thousand gates to a
The history of computing is that of advances born out
fair degree of reliability without resorting to error correc-
of the need to perform ever more sophisticated calcula-
tion. Such circuits are strongly believed to be practically
tions. Increasingly advanced semiconductor manufactur-
impossible to simulate classically, even with the help of
ing processes have resulted in faster and more efficient
modern supercomputers. This suggests the possibility
chips—most recently the 2 nm technology node1 —and
that the first demonstrations of a computational quan-
special accelerators like the GPU, TPU, and AI proces-
tum advantage—where a computational task of business
sors2 have allowed more efficient computations on larger
or scientific relevance can be performed more efficiently,
data sets. These advances share the same model of
cost-effectively, or accurately using a quantum computer
computation dating back to 1936 with the origins of
than with classical computations alone—may be achieved
the Church-Turing thesis. Now for the first time in
without or with limited error correction.
history, the field of computing has branched with the
Three central questions need to be answered for this
emergence of quantum computers, which, when scaled,
to happen: (1) how to extract useful data from the out-
promise to implement computations intractable for con-
put of noisy quantum circuits in the weak noise regime,
ventional computers—from modeling quantum mechani-
(2) how to design quantum algorithms based on shal-
cal systems3 to linear algebra4 , factoring5 , search6 , and
low circuits that can potentially solve some classically
more.
hard problems, and (3) how to improve the efficiency of
Unlocking the full potential of quantum processors re- quantum error-correction schemes and use error correc-
quires the implementation of computations with a large tion more sparingly.
number of operations. Since quantum gates are consider- These questions and our approach are discussed in de-
ably less accurate than their classical counterparts, it is tail in Section II. As an illustration, we pick one of the
strongly believed that error correction will be necessary simplest scientifically relevant applications of quantum
to realize long computations with millions or billions of computers—simulating time evolution of a spin chain
gates. Accordingly, most of quantum computing plat- Hamiltonian7 . We discuss state-of-the-art quantum algo-
forms are designed with the long term goal of realizing rithms for this problem and highlight the cost of making
error-corrected quantum circuits. As the noise rate de- time evolution circuits fault-tolerant by encoding each
creases below a constant architecture-dependent thresh- qubit into the surface code8,9 , considered the best fit for
old, an arbitrarily long quantum circuit can be executed hardware with two-dimensional qubit connectivity. For
reliably by redundantly encoding each qubit and repeat- problem sizes of practical interest, error correction in-
edly measuring parity check operators to detect and cor- creases the size of quantum circuits by nearly six orders
rect errors. However, the number of qubits required to re- of magnitude, making it prohibitively expensive for near-
alize error-corrected quantum circuits solving classically term QPUs (see Section II A).
hard problems exceeds the size of systems available today Question (1) is approached in Sections II B and II C
by several orders of magnitude. through quantum error mitigation10,11 and circuit knit-
Meanwhile, as the quality and number of qubits in ting12–15 . These techniques extend the size of quantum
quantum computers continue to grow, we must be able circuits that can be executed reliably on a given QPU
to harvest the computational power of quantum circuits without resorting to error correction. We estimate the
 2

overhead introduced by state-of-the-art error mitigation do not have to manage the underlying infrastructure. We
methods and discuss recent ideas on how to combine conclude with a high level view from a developer’s/user’s
error correction and mitigation. Circuit knitting tech- lens.
niques exploit structural properties of the simulated sys- This paper offers a perspective of the future of quan-
tem, such as geometric locality, to decompose a large tum computing focusing on an examination of what it
quantum circuit into smaller sub-circuits or combine so- takes to build and program near-term superconducting
lutions produced by multiple QPUs. quantum computers and demonstrate their utility. Real-
The classical simulation algorithms used in computa- izing the computational power of these machines requires
tional physics or chemistry are often heuristics and work the concerted efforts of engineers, physicists, computer
well in practice, even though they do not offer rigor- scientists, and software developers. Hardware advances
ous performance guarantees. Thus, it is natural to ask will raise the bar of quantum computers’ size and fidelity.
whether rigorous quantum algorithms designed for simu- Theory and software advances will lower the bar for im-
lating time evolution admit less expensive heuristic ver- plementing algorithms and enable new capabilities. As
sions that are more amenable to near-term QPUs. We both bars converge in the next few years, we will start
discuss such algorithms in Section II D to address ques- seeing the first practical benefits of quantum computa-
tion (2). tion.
To approach question (3), we discuss generalizations
of the surface code known as low-density parity check
(LDPC) quantum codes16,17 . These codes can pack many II. TOWARDS PRACTICALLY USEFUL QUANTUM
more logical qubits into a given number of physical qubits CIRCUITS
such that, as the size of quantum circuits grows, only a
constant fraction of physical qubits is devoted to error Although in principle a quantum computer can repro-
correction (see Section II A for details). These more effi- duce any calculation performed on conventional classi-
cient codes need long-range connections between qubits cal hardware, the vast majority of everyday tasks are
embedded in a two-dimensional grid18 , but the efficiency not expected to benefit from quantum-mechanical effects.
benefits are expected to outweigh the long-range connec- However, using quantum mechanics to store and process
tivity costs. information can lead to dramatic speedups for certain
We then focus on quantum-centric supercomputing, carefully selected applications. Of particular interest are
which is a new architecture for realizing error mitiga- tasks that admit a quantum algorithm with the run-
tion, circuit knitting, and heuristic quantum algorithms time scaling as a small constant power of the problem
with substantial classical calculations. At the heart of size n—e.g., as n2 or n3 —whereas the best known clas-
this architecture is classical and quantum integration and sical algorithm solving the problem has runtime grow-
n
modularity. We need classical integration at real-time to ing
√ faster than any constant power of n—e.g., as 2 or
n
enable conditioning quantum circuits on classical compu- 2 . We define runtime as the number of elementary
tations (dynamic circuits), at near-time to enable error gates in a circuit (or circuits) implementing the algo-
mitigation and eventually error correction, and at com- rithm for a given problem instance. As the problem size
pile time to enable circuit knitting and advanced com- n grows, the more favorable scaling of the quantum run-
piling. We need modulairty to enable scaling and speed- time quickly compensates for a relatively high cost and
ing up workflows by using parallelization. We first start slowness of quantum gates compared with their classical
in Section III by focusing on superconducting computing counterparts. These exponential or, formally speaking,
hardware and we introduce a series of schemes—which we super-polynomial speedups are fascinating from a purely
denote m, l, c, and t couplers—that give us the amount theoretical standpoint and provide a compelling practical
of flexibility needed for realizing LDPC codes, scaling reason for advancing quantum technologies.
QPUs, and enabling workflows that take advantage of lo- Known examples of tasks with an exponential quan-
cal operations and classical communication (LOCC) and tum speedup include simulation of quantum many-body
parallelization. In Section IV, we discuss the require- systems19 , number theoretic problems such as integer
ments on the quantum stack by defining different lay- factoring5 , solving certain types of linear systems20 , es-
ers for integrating classical and quantum computations, timation of Betti numbers used in topological data anal-
which define requirements on latency, parallelization, and ysis21–23 , and computing topological invariants of knots
the compute instructions. From this, we can define a and links24 . (We leave aside speedups obtained in the so-
cluster-like architecture that we call quantum-centric su- called Quantum RAM model25 , for although it appears
percomputer. It consists of many quantum computation to be more powerful than the standard quantum circuit
nodes comprised of classical computers, control electron- model, it is unclear whether a Quantum RAM can be
ics, and QPUs. A quantum runtime can be executed on efficiently implemented in any real physical system.)
a quantum-centric supercomputer, working in the cloud Simulation of quantum many-body systems has re-
or other classical computers to run many quantum run- ceived the most attention due to its numerous scientific
times in parallel. Here we propose that a serverless model and industrial applications, and for being the original
should be used so that developers can focus on code and value proposition for quantum computing26 . The ground
 3

state and thermal-equilibrium properties of many-body

systems can often be understood, at least qualitatively, first order
using classical heuristic algorithms such as dynamical 108 fourth order
mean-field theory (DMFT) or perturbative methods. sixth order

CNOT count
However, understanding their behavior far from equi-
librium in the regime governed by coherent dynamics
or performing high-precision ground state simulations 107
for strongly-interacting electrons—e.g., in the context of
quantum chemistry—is a notoriously hard problem for
classical computers. 106
As a simple illustration, consider a spin chain com-
posed of n quantum spins (qubits or qudits) with Hamil-
tonian
20 30 40 60 80 100
n−1
X Qubits n
H= Hj,j+1 ,
j=1
FIG. 1. Estimated number of CNOT gates required to ap-
where Hj,j+1 is a two-spin nearest-neighbor interaction. proximate the unitary evolution operator e−iHt for the n-
qubit Heisenberg chain with t = n and approximation error
The Schrödinger equation
0.001 using randomized k-th order product formulas (k =
d|ψ(t)i 1, 4, 6). The presented data is based on empirical estimates of
i = H|ψ(t)i ref. 39, see Eq. (70) therein, assuming that exponentiating a
dt single term in the Hamiltonian costs 3 CNOTs.
governs the coherent time evolution of the system from
some fixed initial state |ψ(0)i.
Suppose our goal is to compute the expected value of power of n; e.g., t ∼ n, this constitutes an exponential
some local observable on the time-evolved state |ψ(t)i = quantum speedup.
e−iHt |ψ(0)i. Such expected values are of great interest A natural question is what are the minimum quan-
for understanding, among other things, thermalization tum resources; i.e., qubit and gate counts, required to
mechanisms in closed quantum systems27 . Transforming convincingly demonstrate a quantum advantage for sim-
the time-dependent expected values into the frequency ulating coherent dynamics. Childs, Maslov, et. al. pro-
domain provides valuable information about the excita- posed a concrete benchmark problem for this, simulating
tion spectrum of the system28,29 . A slightly modified ver- the time evolution of the spin-1/2 Heisenberg chain with
sion of this problem that involves measuring each qubit of n = t = 100 and approximation Pn−1error 0.0017 . P
The Hamil-
n
|ψ(t)i is known to be BQP-complete30–33 , meaning that tonian has the form H = σj ~σj+1 + j=1 hj σjz ,
j=1 ~
it is essentially as hard as simulating a universal quantum where hj ∈ [−1, 1] are randomly chosen magnetic fields.
computer. Fig. 1 shows the gate count estimates for the benchmark
The known classical algorithms for simulating the co- problem obtained by Childs, Ostrander, and Su39 , sug-
herent time evolution of a quantum spin chain have run- gesting that about 107 CNOT gates (and a comparable
time min (2O(n) , 2O(vt) ), where v ∼ maxj kHj,j+1 k is the number of single-qubit gates) are needed. This exceeds
Lieb-Robinson velocity, which controls how fast informa- the size of quantum circuits demonstrated experimen-
tion propagates through the system34 . For simplicity, we tally to date by several orders of magnitude. As we move
ignore factors polynomial in n and t. The runtime 2O(n) from simple spin chain models to more practically rel-
can be achieved using a standard state vector simulator evant Hamiltonians, the gate count required to achieve
while the runtime 2O(vt) can be achieved by approximat- quantum advantage increases dramatically. For exam-
ing |ψ(t)i with Matrix Product States35,36 or by restrict- ple, simulating the active space of molecules involved in
ing the dynamics to a light cone37 . In general, the linear catalysis problems may require about 1011 Toffoli gates40 .
growth of the entanglement entropy with time appears The only viable path to reliably implementing circuits
to be an insurmountable obstacle for classical simulation with 107 gates or more on noisy quantum hardware is
algorithms. quantum error correction.
Haah, Kothari, et. al. recently found a nearly optimal
quantum algorithm for simulating the time evolution of
spin chain Hamiltonians38 with runtime Õ(nt), where Õ A. Quantum error correction
hides factors logarithmic in n, t, and the inverse approx-
imation error. This algorithm works by approximating One reason why conventional classical computers be-
the time evolution operator e−iHt by a product of sim- came ubiquitous is their ability to store and process infor-
pler unitaries describing forward and backward time evo- mation reliably. Small fluctuations of an electric charge
lution of small blocks of spins of length O(log nt). Assum- or current in a microchip can be tolerated due to a highly
ing that the evolution time t scales as a small constant redundant representation of logical 0 and 1 states by
 4

a collective state of many electrons. Quantum error- offered by the code. In the sub-threshold regime, one
correcting codes provide a similar redundant representa- must be able to realize arbitrarily precise logical gates
tion of quantum states that protects them from certain from some universal gate set by choosing a large enough
types of errors. A single logical qubit can be encoded code distance.
into n physical qubits by specifying a pair of orthogonal The 2D surface code8,9 has so far been considered
n-qubit states |0i and |1i called logical-0 and logical-1. an uncontested leader in terms of the error thresh-
A single-qubit state α|0i + β|1i is encoded by the logical old—close to 1% for the commonly studied depolarizing
state α|0i + β|1i. A code has distance d if no operation noise48–50 —yet has two important shortcomings. First,
affecting fewer than d qubits can distinguish the logical allocating a roughly d×d patch of physical qubits for each
states |0i and |1i or map them into each other. More logical qubit incurs a large overhead. Unfortunately, it
generally, a code may have k logical qubits encoded into was shown51 that any 2D stabilizer code has encoding
n physical qubits and the code distance d quantifies how rate k/n = O(1/d2 ) which vanishes for large code dis-
many physical qubits need to be corrupted before the log- tance. This means that as one increases the degree of
ical (encoded) state is destroyed. Thus good codes have protection offered by the surface code, quantified by the
a large distance d and a large encoding rate k/n. code distance d, its encoding rate approaches zero. That
Stabilizer-type codes41,42 are by far the most studied is, as the size of quantum circuits grows, the vast ma-
and promising code family. A stabilizer code is defined by jority of physical qubits are devoted to error correction.
a list of commuting multi-qubit Pauli observables called This is a known fundamental limitation of all quantum
stabilizers such that logical states are +1 eigenvectors codes that can be realized locally in the 2D geometry.
of each stabilizer. One can view stabilizers as quantum To make error correction more practical and minimize
analogues of classical parity checks. Syndrome measure- qubit overhead, codes with a large encoding rate k/n
ments aim to identify stabilizers whose eigenvalue has are preferable. For example, quantum LDPC codes can
flipped due to errors. The eigenvalue of each stabilizer achieve a constant encoding rate independent of the code
is repeatedly measured and the result—known as the er- size44 . In fact, the encoding rate can be arbitrarily close
ror syndrome—is sent to a classical decoding algorithm. to one16 . A recent breakthrough result52 demonstrated
Assuming that the number of faulty qubits and gates is the existence of so-called good quantum LDPC codes
sufficiently small, the error syndrome provides enough in- that combine a constant encoding rate k/n (which can
formation to identify the error (modulo stabilizers). The be close to 1/2) and a linear distance d ≥ cn for some
decoder can then output the operation that needs to be constant c > 0. For comparison, the 2D surface code
applied to recover the original logical state. has an asymptotically vanishing encoding rate and has
√
Most of the codes designed for quantum computing distance at most n. Certain LDPC codes have a favor-
are of the LDPC type16,43,44 , meaning that each stabi- able property known as single-shot error correction53,54 .
lizer acts on only a small number of qubits and each qubit They provide a highly redundant set of low-weight Pauli
participates in a small number of stabilizers, where small observables (known as gauge operators) that can be mea-
means a constant independent of the code size. The main sured to obtain the error syndrome more efficiently. This
advantage of quantum LDPC codes is that the syndrome reduces the number of syndrome measurement cycles per
measurement can be performed with a simple constant- logical gate from O(d) to O(1) and hence enables very
depth quantum circuit. This ensures that the syndrome fast logical gates. The syndrome measurement circuit
information can be collected frequently enough to cope for a quantum LDPC code requires a qubit connectivity
with the accumulation of errors. Furthermore, errors in- dictated by the structure of stabilizers, i.e., one must be
troduced by the syndrome measurement circuit itself are able to couple qubits that participate in the same stabi-
sufficiently benign since the circuit can propagate errors lizer. Known examples of LDPC codes with a single-shot
only within a “light cone” of constant size. error correction require 3D or 4D geometry53–55 .
A code must satisfy several requirements to have appli- The second shortcoming of the surface code is the diffi-
cations in quantum computing. First, it must have a high culty of implementing a computationally universal set of
enough error threshold—the maximum level of hardware logical gates56 . The surface code and its variations such
noise that it can tolerate. If the error rate is below the as the honeycomb code57 or folded surface code58 offer
threshold, the lifetime of logical qubit(s) can be made ar- a low-overhead implementation of logical Clifford gates
bitrarily long by choosing a large enough code distance. such as CNOT, Hadamard H, and phase shift S. These
Otherwise, errors can accumulate faster than the code gates can be realized by altering the pattern of stabilizers
can correct them and logical qubits can become even less measured at each time step using the code deformation
reliable than the constituent physical qubits. Second, one method. However, Clifford gates are not computationally
needs a fast decoding algorithm to perform error correc- universal on their own. A common strategy for achiev-
tion in real time as the quantum computation proceeds. ing universality is based on√preparation of logical ancil-
This may be challenging since the decoding problem for lary states (|0i + eiπ/4 |1i)/ 2 known as magic states. A
general stabilizer codes is known to be NP-hard in the magic state is equivalent (modulo Clifford operations) to
worst case45–47 . Third, one must be able to compute on a single-qubit gate T = diag(1, eiπ/4 ). The Clifford+T
the logical qubits without compromising the protection gate set is universal and has a rich algebraic structure
 5

enabling efficient and nearly-optimal compiling of quan- putations. A limited form of correction for shallow quan-
tum algorithms59,60 . Unfortunately, the overhead for tum circuits can be achieved by combining the out-
distilling high-fidelity magic states is prohibitively large. comes of multiple noisy quantum experiments in a way
O’Gorman and Campbell61 performed a careful examina- that cancels the contribution of noise to the quantity of
tion of available distillation methods and their overhead, interest10,11 . These methods, collectively known as error
considering the implementation of a logical Clifford+T mitigation, are well suited for the QPUs available today
circuit of size N with an overall fidelity of 90%. Assum- because they introduce little to no overhead in terms of
ing a physical error rate of 10−3 , the following fitting the number of qubits and only a minor overhead in terms
formulas were found for the space-time volume ((physi- of extra gates. However, error mitigation comes at the
cal qubits) × (syndrome measurement cycles)) associated cost of an increased number of circuits (experiments) that
with a single logical gate: need to be executed. In general, this will result in an
exponential overhead; however, the base of the exponent
Logical gate Physical space-time volume can be made close to one with improvements in hardware
2.77 and control methods, and each experiment can be run in
CNOT 1610 + 45 (log10 N )
T -gate 3.13 + 3220 (log10 N )
3.20 parallel. Furthermore, known error mitigation methods
apply only to a restricted class of quantum algorithms
The space-time volume roughly quantifies the number of that use the output state of a quantum circuit to esti-
physical gates required to implement a single logical gate. mate the expected value of observables.
As an example, consider the Heisenberg benchmark Probabilistic error cancellation (PEC)10,70 aims to ap-
problem described above with 100 logical qubits. The proximate an ideal quantum circuit via a weighted sum of
desired time evolution operator can be approximated us- noisy circuits that can be implemented on a given quan-
ing about 107 CNOTs and single-qubit gates (see fig. 1). tum computer. The weights assigned to each noisy circuit
However, each single-qubit gate needs to be compiled us- can be computed analytically if the noise in the system
ing the logical-level gate set {H, S, T }. In total, this re- is sufficiently well characterized or learned by mitigating
quires roughly 109 T -gates and a comparable number of errors on a training set of circuits that can be efficiently
Clifford gates7 . Accordingly, the physical space-time vol- simulated classically71 . We expect that the adoption of
umes of a single logical CNOT and T -gate are roughly PEC will grow due to the recent theoretical and exper-
2 × 104 and 4 × 106 , respectively. (In fact, this underesti- imental advances in quantum noise metrology72–74 . For
mates the error correction overhead since the Heisenberg example, ref. 74 shows how to model the action of noise
benchmark problem requires logical circuit fidelity 0.999 associated with a single layer of two-qubit gates by a
rather than 0.9, as considered in ref. 61.) Markovian dynamics with correlated Pauli errors. This
The large overhead associated with logical non-Clifford model can be described by a collection of single-qubit
gates may rule out the near-term implementation of and two-qubit Pauli errors P1 , . . . , Pm and the associ-
error-corrected quantum circuits, even if fully function- ated error rate parameters λ1 , . . . , λm ≥ 0 such that the
ing logical qubits based on the surface code become avail- combined noise channel acting on a quantum register has
able soon. There have been several strategies proposed the form Λ(ρ) = exp [L](ρ), where L is a Lindblad gen-
erator, L(ρ) = i=1 λi (Pi ρPi† − ρ). The unknown error
Pm
recently for reducing this overhead, including high-yield
magic state distillation methods62,63 , better strategies for rates λi can be learned to within several digits of preci-
preparing “raw” noisy magic states that reduce the re- sion by repeating the chosen layer of gates many times
quired number of distillation rounds64 , and better sur- and measuring the decay of suitable observables74 . The
face code implementations of distillation circuits65–68 . error mitigation overhead (as measured by the number of
A recent breakthrough result by Benjamin Brown69 circuit repetitions) per layer of gates scales as γ 2 , where
shows how to realize a logical non-Clifford gate CCZ !
m
(controlled-controlled-Z) in the 2D surface code archi- X
tecture without resorting to state distillation. This ap- γ = exp 2 λi .
i=1
proach relies on the observation that a 3D version of the
surface code enables an easy (transversal) implementa- For a circuit composed of d > 1 layers, the error rates
tion of a logical CCZ and a clever embedding of the 3D λi may be layer-dependent and have to be learned sepa-
surface code into a 2+1 dimensional space-time. It re- rately for each layer. As observed in ref. 74, this model
mains to be seen whether this method is competitive can approximate the actual hardware noise very well us-
compared with magic state distillation. ing only m = O(n) elementary Pauli errors Pi supported
on edges of the qubit connectivity graph, where n is the
total number of qubits in the circuit. In general, the run-
B. Error mitigation time for getting a noise-free estimate will depend on the
circuit implemented and the noise model used.
Although error correction is vital for realizing large- A large class of quantum algorithms that can benefit
scale quantum algorithms with great computational from PEC is based on the so-called hardware-efficient
power, it may be overkill for small or medium size com- circuits75 . A depth-d hardware-efficient circuit con-
 6

to scale up to 27 qubits and still reconstruct observables.

Whether this method can be combined with PEC, which
gives an unbiased estimation, remains an open question.
More general (non-Markovian) noise can be mitigated
using the virtual distillation technique78,79 . It works
by combining two copies of a noisy output state ρ in a
way that enables measurements of observables on a state
ρ2 /Tr(ρ2 ). Assuming that ρ has most of its weight on
the ideal output state, virtual distillation can quadrati-
cally suppress the contributions of errors. However, this
method introduces at least a factor of two overhead in
the number of qubits and gates. A review of the exist-
ing error mitigation proposals can be found in Endo, et.
al.80 .
We anticipate error mitigation to continue to be rele-
FIG. 2. Runtime scaling (number of circuit instances) needed
to error mitigate 100 and 1000 Trotter steps in a circuit of 100 vant when error-corrected QPUs with a hundred or more
qubits and layers of non-overlapping two-qubit gates, each logical qubits become available. As discussed in Section
gate affected by a local depolarizing two-qubit error. The II A, the first generation of error-corrected quantum chips
red dotted line identifies 100 million circuits, the daily limit based on 2D stabilizer codes may not be able to execute
assuming a repetition rate of 1 ms. So far the only architec- universal computations. Such QPUs are likely to offer
tures that have achieved this speed are solid-state based. The only high-fidelity Clifford gates such as the Hadamard or
number of circuit instances dramatically decreases with slight CNOT, which can all be efficiently simulated on a classi-
improvements in the error rate of the physical gates. cal computer. Meanwhile, logical non-Clifford gates such
as the T -gate may remain out of reach due to the need
to perform magic state distillation. This leads to the
sists of d layers of two-qubit gates such that all gates interesting possibility of combining error correction and
within the same layer are non-overlapping and couple mitigation. A concrete proposal by Piveteau, et al.81
qubits that are nearest-neighbors in the QPU connec- leverages the ability to realize noisy logical T -gates with
tivity graph.
PmDenoting the average error rate per qubit fidelity comparable to or exceeding that of physical (un-
λ̄ = (1/n) i=1 λi , averaged over all d layers, the overall encoded) gates. Applying error mitigation protocols at
PEC overhead scales as (γ̄)dn , where γ̄ = exp (4λ̄). This the logical level to cancel errors introduced by noisy T -
allows a simple formula for estimating the runtime, J, gates enables one to simulate universal logical circuits
for a noise-free estimate from a quantum circuit of depth without resorting to state distillation. This may consid-
d and width n to be erably reduce the hardware requirements for achieving
a quantum advantage. However, error mitigation comes
J = d(γ̄)dn β, (1) at the cost of an increased number of circuit executions.
Assuming a physical gate fidelity of 99.9% and a bud-
where β is the average time to run a single layer of the
get of 1, 000 circuit executions, Piveteau, et al.81 esti-
circuit. One can view β as a measure of the “speed” and
mate that logical Clifford+T circuits with about 2,000
γ̄ as a hardware-dependent parameter that quantifies the
T -gates can be realized reliably. This is far beyond the
average “quality” of gates across the entire QPU.
limit of existing classical algorithms that can simulate
For a spin chain of n = 100 qubits, the size of our
Clifford+T circuits with about 50 T -gates82,83 . Similar
benchmark problem, fig. 2 shows the number of circuit
ideas for combining error correction and mitigation are
instances that need to be sampled to perform PEC for
discussed in refs. 84 and 85.
100 and 1000 Trotter steps using the decomposition in fig.
3A of ref. 74. Current hardware runs of up to 108 circuits
daily (red dashed line) and error rates of 10−3 have been
demonstrated. Hence, we anticipate that with a cou- C. Circuit Knitting
ple orders of magnitude improvement this becomes pos-
sible. Furthermore, this runtime can be further reduced We can extend the scope of near-term hardware to
with the quantum-centric supercomputing architecture compensate for other shortcomings such as a limited
that allows parallelized execution of quantum circuits. number of qubits or qubit connectivity by using circuit
We can also measure the quantity of interest at several knitting techniques. This refers to the process of sim-
different values of the noise rate and perform an extrapo- ulating small quantum circuits on a quantum computer
lation to the zero-noise limit10,11,76 . This method cancels and stitching their results into an estimation of the out-
the leading-order noise contribution as long as the noise is come of a larger quantum circuit. As was the case with
weak and Markovian. Unlike PEC, this method is biased error mitigation, known circuit knitting techniques apply
and heuristic but may require fewer circuits for the re- to a restricted class of quantum algorithms that aim to
construction. This method was recently demonstrated77 estimate the expected value of observables.
 7

The most well-known example is circuit cutting12–15 . the expected energy of a given quantum Hamiltonian or
In this method, a large quantum circuit is approximated a classical cost function encoding the problem of interest.
by a weighted sum of circuits consisting of small iso- The trial state is usually defined as the output state of
lated sub-circuits. Each sub-circuit can be executed sep- a shallow quantum circuit. Rotation angles that define
arately on a small QPU. The overhead introduced by individual gates serve as variational parameters. These
this method (as measured by the number of circuit rep- parameters are adjusted via a classical feedback loop to
etitions) scales exponentially with the number of two- optimize the chosen cost function.
qubit gates or qubit wires that need to be cut in order At present, there is no mathematical proof that VQA
to achieve the desired partition of the circuit. Surpris- can outperform classical algorithms in any task. In fact,
ingly, it was recently shown that the circuit cutting over- it is known that VQA based on sufficiently shallow (con-
head can be substantially reduced by running the iso- stant depth) variational circuits with 2D or 3D qubit
lated sub-circuits in parallel using non-interacting QPUs connectivity can be efficiently simulated on a classical
that can only exchange classical data86 . This approach computer101,102 . This rules out a quantum advantage.
requires hardware capable of implementing dynamic cir- Meanwhile, the performance of VQA based on deep vari-
cuits1 , where the control electronics is extended to in- ational circuits is severely degraded by noise103 . How-
clude independent QPUs. ever, as the error rates of QPUs decrease, we should be
A second example is entanglement forging87 , where ei- able to execute VQA in the intermediate regime where
ther an entangled variational state is decomposed into quantum circuits are already hard to simulate classically
a weighted sum of product states or the entanglement but the effect of noise can still be mitigated.
between a pair of qubit registers is converted into time- As a concrete example, let us discuss possible appli-
like correlations within a single register88 . The overhead cations of VQA to the problem of simulating coherent
of this method typically scales exponentially with the time evolution of quantum spin chains. It is commonly
amount of entanglement across the chosen partition of believed38 that approximating the time evolution oper-
the system. ator e−iHt for a Hamiltonian H describing a 1D chain
A third example, closely related to circuit knitting, of n qubits requires a quantum circuit of size scaling at
uses embedding methods to decompose the simulation least linearly with the space-time volume nt. Meanwhile,
of a large quantum many-body system into smaller sub- the best known rigorous quantum algorithms based on
systems that can be simulated individually on a QPU. product formulas39 or Lieb-Robinson bounds38 require
The interactions between subsystems are accounted for circuits of size O(n2 t) or O(nt · polylog(nt)), respectively.
by introducing an effective bath that could be either a A natural question is whether VQA can reduce the cir-
classical environment or another small quantum system. cuit size to what is believed to be optimal, that is, linear
The decomposition of the original system and the op- in nt. If this is indeed the case, a QPU with gate fidelity
timization of the bath parameters are performed on a around 99.99% may be able to solve the classically hard
classical computer that can exchange classical data with problem instances described above with space-time vol-
the QPU. Well-known examples of quantum embedding ume nt ∼ 104 using existing error mitigation techniques.
methods that build on their classical counterparts are Variational quantum time evolution (VarQTE) algo-
dynamical mean-field theory89–91 , density-matrix embed- rithms, pioneered by Li and Benjamin11 , could be an
ding92–94 , and density-functional embedding95 . alternative to simulate the time evolution of these classi-
cally hard instances given near-term noisy QPUs. These
algorithms aim to approximate the time-evolved state
D. Heuristic quantum algorithms |ψ(t)i = e−iHt |ψ(0)i by a time-dependent variational
ansatz |φ(θ)i = U (θ)|0n i, where U (θ) is a parameter-
Heuristic quantum algorithms can be employed near- ized quantum circuit with a fixed layout of gates and
term to solve classical optimization96 , machine learn- θ = θ(t) is a vector of variational parameters. The ini-
ing97 , and quantum simulation98 problems. These tial state |ψ(0)i is assumed to be sufficiently simple so
fall into two categories—algorithms that use kernel that the variational ansatz for |ψ(0)i is easy to find. The
methods97 and variational quantum algorithms (VQA). goal is to find a function θ(t) such that the variational
Quantum kernel methods have also been found that lead state |φ(θ(t))i approximates the time evolved state |ψ(t)i
to provable speedups99 and expand to a class of kernels for all t in the chosen time interval. As shown in ref.
for data with group structure100 . For VQA, the basic pro- 11, the desired function θ(t) can be efficiently computed
posal is appealingly simple: an experimentally-controlled using the stationary-action principle with a Lagrangian
trial state is used as variational wavefunction to minimize L(t) = hφ(θ(t))|d/dt + iH|φ(θ(t))i. This yields a first-
order differential equation98 q Mp,q θ̇q = Vp where
P

1 Dynamic circuits are computational circuits that combine quan- Mp,q = Im(h∂p φ(θ)|∂q φ(θ)i),
tum and classical operations, using the outcome of classical com-
putations to adapt subsequent quantum operations. For more
information, see section IV. Vp = −Re(h∂p φ(θ)|H|φ(θ)i),
 8

and ∂p ≡ ∂θ∂p . As shown in [11], the entries of M and III. THE PATH TO LARGE QUANTUM SYSTEMS
V can be efficiently estimated on a quantum computer.
A comprehensive review of VarQTE algorithms can be The perspective above leads to a challenge in quantum
found in refs. 98 and 104. hardware. We believe there will be near-term advan-
The fact that VarQTE algorithms are heuristics and tage using a mixture of error mitigation, circuit knitting
therefore lack rigorous performance guarantees raises the and heuristic algorithms. On a longer time frame, par-
question of how to validate them. This becomes par- tially error-corrected systems will become critical to run-
ticularly important for large problem sizes where veri- ning more advanced applications and further down the
fying a solution of the problem on a classical computer line, fault-tolerant systems running on not-as-yet fully
becomes impractical. Ref. 105 recently developed a ver- explored LDPC codes with non-local checks will be key.
sion of VarQTE based on McLachlan’s variational princi- The first steps for all of these approaches are the same:
ple that comes with efficiently computable bounds on the we need hardware with more qubits capable of higher fi-
distance between the exact time-evolved state |ψ(t)i and delity operations. We need tight integration of fast clas-
the approximate variational state found by the VarQTE sical computation to handle the high run-rates of circuits
algorithms. Thus, although VarQTE lacks a rigorous jus- needed for error mitigation and circuit knitting, and the
tification, one may be able to obtain a posteriori bounds classical overhead of the error correction algorithm after-
on its approximation error for some specific problems of wards. This drives us to identify a hardware path that
practical interest. starts with the early heuristic small quantum circuits and
grows until reaching an error-corrected computer.

E. Summary
A. Cycles of Learning
To summarize, the Heisenberg chain example illus-
trates what we believe are general guidelines for designing The first step in this path is to build systems able to
near-term quantum algorithms. demonstrate near-term advantage with error mitigation
First, our best chance of attaining a quantum advan- and limited forms of error correction. Just a few years
tage is by focusing on problems that admit an exponen- ago, QPU sizes were limited by control electronics cost
tial (super-polynomial) quantum speedup. Even though and availability, I/O space, quality of control software,
a quantum algorithm that achieves such speedup with and a problem referred to as “breaking the plane”106 ,
formal proof may be out of reach for near-term hard- i.e., routing microwave control and readout lines to qubits
ware, its mere existence serves as compelling evidence in the center of dense arrays. Today, solutions to these
that quantum-mechanical effects such as interference or direct barriers to scaling have been demonstrated, which
entanglement are beneficial for solving the chosen prob- has allowed us to lift qubit counts beyond 100—above the
lem. threshold where quantum systems become intractably
Second, the only known way to realize large-scale quan- difficult to simulate classically and examples of quantum
tum algorithms relies on quantum error-correcting codes. advantage become possible. The next major milestones
The existing techniques based on the surface code are are (1) increasing the fidelity of QPUs enough to allow
not satisfactory due their poor encoding rate and high exploration of quantum circuits for near-term quantum
cost of logical non-Clifford gates. Addressing these short- advantage with limited error correction and (2) improv-
comings may require advances in quantum coding theory ing qubit connectivity beyond 2D—either through modi-
such as developing high-threshold fault-tolerant protocols fied gates, sparse connections with non-trivial topologies,
based on quantum LDPC codes and improving the qubit and/or increasing the number of layers for quantum sig-
connectivity of QPUs beyond the 2D lattice. Supple- nals in 3D integration—to enable the longer term explo-
menting error correction with cheaper alternatives such ration of efficient non-2D LDPC error-correction codes.
as error mitigation and circuit knitting may provide a These developments are both required for our longer term
more scalable way of implementing high-fidelity quantum vision, but can be pursued in parallel.
circuits. Work on improving the quality of quantum systems by
Third, near-term quantum advantage should be pos- improving gate fidelities involves many cycles of learn-
sible by exploring less expensive, possibly heuristic ver- ing, trying coupling schemes, process changes, and in-
sions of the algorithm considered. Those heuristic quan- novations in controlling coupling and crosstalk. Scal-
tum algorithms lack rigorous performance guarantees, ing this work to large QPUs capable of demonstrating
but they may be able to certify the quality of a solu- quantum advantage, and ultimately to the extreme sys-
tion a posteriori and offer a way to tackle problems that tem scales we anticipate in the distant future, involves
cannot be simulated classically. integrating different technologies with enough reliability
We believe these general guidelines define the future and skill to make size be limited by cost and need, not
of quantum computing theory and will guide us to im- by technological capability. This adds challenges in re-
portant demonstrations of its benefits for the solution of liability, predictability, and manufacturability of QPUs
scientifically important problems in the next few years. while continuing to incorporate improved technologies
 9

FIG. 3. An example of a scheme that allows breaking the plane for signal delivery compatible with the integration of hundreds
of qubits. It is composed of technologies adapted from conventional CMOS processing.

into these complex systems. Meanwhile, the increased eter in fixed frequency systems. We can take advan-
development, fabrication, and test times for larger sys- tage of these statistical correlations wherever they exist
tems creates a lag in cycles of innovation that must be for rapid progress in parts of our process110 or in post-
overcome. process tuning111 . However, reliably establishing these
The manufacturing cycle time increases with QPU so- correlations requires measuring hundreds or thousands
phistication. Many simple transmon QPUs require just a of devices, a nontrivial feat.
single level of lithography and can be easily fabricated in Absent these correlations, we can use simplified test ve-
a day or two. Even the 5- and 16-qubit QPUs that were hicles; for example, rather than using the entire compli-
IBM’s initial external cloud quantum systems involved cated signal delivery stack when trying to improve qubit
only two lithography steps and took a week to fabricate. coherence, we can use a simplified device designed to ob-
Compare this to more advanced packaging schemes like tain good statistics and fast processing112 . Still, identi-
those at MIT Lincoln Laboratory107–109 or IBM’s newer fying specific steps leading to increased coherence is non-
“Eagle” QPUs (fig. 3), which involve dozens of lithog- trivial. It is rarely possible to change just one parameter
raphy steps and slow process steps, and take months to in materials processing. Changing a metal in a qubit
build at a research-style facility with one-of-a-kind tools. may also change etch parameters, chemicals compatible
This increased cycle time makes it harder to reach the fi- with the metal for subsequent processing, and even al-
delities and coherence times needed as well as debug the lowed temperature ranges113 . Once an improved process
manufacturing and assembly for reliable QPU yield. is found, it is hard to identify exactly which steps were
Reliability in semiconductor manufacturing is not a critical vs. simply expedient.
new problem. In general, among the unique component We must gather sufficient statistics when performing
challenges faced in building a scaled machine, the conven- materials research for the results be meaningful and pro-
tional semiconductor technologies integrated on chip are vide enough certainty114 . We should carefully document
the most well studied. Incorporating them in supercon- process splits wherever relevant, and we should publish
ducting technologies is more a matter of ensuring that the changes in materials processes that lead to neutral or
associated processes are compatible with each other than even negative results, not just just publish highly suc-
inventing new approaches. However, the rapid growth of cessful work.
volume we anticipate being needed is a major challenge. Similar difficulties occur in non-material based re-
Many failure modes in superconducting quantum sys- search on devices. Some gates work well between pairs
tems are not detectable until the QPUs are cooled to of qubits yet exhibit strong couplings that make them
their operating temperature, sub-100 mK. This is a se- unsuitable for larger QPUs or compromise single-qubit
vere bottleneck that renders in-line test (where a device performance. Three- and four-qubit experiments are no
sub-component is tested for key metrics before the QPU longer challenging from a technical or budgetary perspec-
build finishes) and process feed-forward (where future tive. To be relevant to larger QPUs, research needs to
process steps are modified to correct for small deviations move away from two-qubit demos, especially hero exper-
in early steps and stabilize total device performance) dif- iments between a single pair of qubits in which many
ficult or impossible. There are exceptions where it is pos- critical defects can be masked by luck.
sible to tightly correlate an easy measurement at room A mixture of long cycle-time complex devices and short
temperature with ultimate QPU performance: for ex- cycle-time test vehicles for sub-process development and
ample, resistance measurements of Josephson junctions quantum operations is key to continuing improvements in
can accurately predict their critical currents and hence, the quality of QPUs and provides a recipe for continued
the frequency of qubits made with them—a key param- R&D contributions as the largest QPUs begin to exceed
 10

the capabilities of smaller groups and labs. Nonethe- thousands of qubits and beyond. On top of this, the soft-
less, reductions in long cycle times are needed. Some ware that translates a quantum circuit into the low-level
of this will come naturally—first-of-a-kind processes and representation of this control hardware is becoming in-
QPUs usually take longer as they tend to include ex- creasingly complex and expensive to produce. Reducing
tra steps, inspections, and in-line tests that, while sug- cost favors a common control platform with customized
gested by general best practices, may not be necessary. analog front ends. Open-specification control protocols
While counterproductive from a cost viewpoint, building like OpenQASM3119 are already paving the way for this
the “same” QPU repeatedly to iron out manufacturing transformation.
problems and speed up cycles of innovation will likely be
a successful strategy for the largest QPUs with the most
complex fabrication flows. C. Classical parallelization of quantum processors

Reaching near-term quantum advantage will require

B. Supporting Hardware taking advantage of techniques like circuit knitting and
error mitigation that effectively stretch the capabilities of
Scaling to larger systems also involves scaling classical QPUs—trading off additional circuit executions to emu-
control hardware and the input/output (I/O) chain in late more qubits or higher fidelities. These problems can
and out of the cryostat. This I/O chain, while still need- be pleasingly parallel, where individual circuits can ex-
ing substantial customization for the exact QPU being ecute totally independently on multiple QPUs, or may
controlled, consists of high volumes of somewhat more benefit from the ability to perform classical communi-
conventional devices; for example, isolators, amplifiers, cation between these circuits that span multiple QPUs.
scaled signal delivery systems, and more exotic replace- Introduction of control hardware that is able to run mul-
ments such as non-ferrite isolators and quantum limited tiple QPUs as if they were a single QPU with shared
amplifiers that may offer performance, cost, or size im- classical logic, or split a single QPU into multiple virtual
provements. These components have enormous potential QPUs to allow classical parallelization of quantum work-
for being shared between various groups pursuing quan- loads is an important near-term technology for stretching
tum computing, and in some instances can be purchased this advantage to the limit. Longer term, these technolo-
commercially already. However, assembling these sys- gies will play a critical enabling role as we begin to build
tems at the scale required today, let alone a few years quantum systems that span mutliple chips and multiple
time, requires a high volume cryogenic test capability cryostats, i.e., modular quantum systems.
that does not currently exist in the quantum ecosystem,
creating a short-term need for vertically-integrated man-
ufacturing of quantum systems. The challenge here is es- D. Modularity
tablishing a vendor and test ecosystem capable of scaled,
low-cost production—a challenge made difficult by the The introduction of modular quantum systems will
fact that the demand is somewhat speculative. be key to bootstrapping ourselves from near-term quan-
There are also one-off components per system; for ex- tum advantage towards long-term error-corrected quan-
ample, each quantum computer we deploy only requires tum systems. These are systems with repeating unit cells
a single dilution refrigerator, or in many cases a frac- that can be replaced if defective, with quantum links be-
tion thereof. The dilution refrigerator manufacturer ef- tween the chips to entangle unit cells or perform remote
fectively acts as a systems integrator for cryo-coolers, gates. This approach simplifies QPU design and test, and
wiring solutions, pumping systems, and even some ac- allows us to scale quantum systems at will.
tive electronics. Maintaining the flexibility we need to In the near term, given limited or no error correction,
change quickly as the systems scale will be most easily the unit cells will require high-bandwidth and high fi-
attainable if we can standardize many of these interfaces delity links to connect them—there is not enough time
so that, for example, moving to a more scalable cooling to use complex protocols such as entanglement distilla-
technology at 4K doesn’t require redesigning the entire tion. The simplest proposals to accomplish this extend
refrigeration infrastructure. quantum busses off chip, allowing the same gates between
Currently, each group building large QPUs has their distant chips as on a single processor120,121 . This “dense
own bespoke control hardware. Given the radically dif- modularity”, which we denote m, effectively extends the
ferent control paradigms and requirements115–118 , it is chip size. This requires linking adjacent chips with ultra
unlikely that the analog front-ends of these systems could low loss, low cross-talk lines that are short enough to be
ever be shared. However, there is a common need for effectively single-mode—the distance between chips has
sequencing logic (branching, local and non-local condi- to be of the order of the distance between qubits on a
tionals, looping) at low-cost and low-power for all types single chip. Several technologies from classical computa-
of quantum computers, not just solid-state. These will tional hardware may be adaptable to this problem but
likely need to be built into a custom processor—an Appli- adding the flexibility to replace individual units will re-
cation Specific Integrated Circuit or ASIC—as we scale to quire other alternatives122 .
 11

(a)p type modularity for classical parallelization of QPUs (b)Dense modularity m and on-chip non-local couplers c for LDPC
codes for creating a single QPU from multiple chips

(c)Long-range l type modularity to enable quantum parallelization (d)l, m, p schemes can be combined to extend the scale of
of multiple QPUs hardware to thousands of qubits.

(e)t type modularity involves microwave-to-optical transduction to

link QPUs in different dilution refrigerators.

FIG. 4. Beyond classical parallelization of QPUs, shown in (a), long-range quantum connections carry a high penalty in gate
speed and fidelity. As shown in (b)-(e), a high fidelity, large quantum system will likely involve three levels of modularity—a
very short-range modularity m that allows breaking a QPU into multiple chips with minimal cost in gate speed and fidelity, a
longer range connection l for use within a single cryogenic environment to both get around I/O bottlenecks and allow non-trivial
topologies or routing, and a very long-range optical “quantum network” t to allow nearby QPUs to work together as a single
quantum computational node (QCN). We will also need on-chip non-local couplers c as shown in (b) for the exploration of
LDPC codes. In this figure, pink lines represent quantum communication and purple lines represent classical communication.
 12

Type Description Use

p Real-time classical communication Classical parallelization of QPUs
m Short range, high speed, chip-to-chip Extend effective size of QPUs
l Meter-range, microwave, cryogenic Escape I/O bottlenecks, enabling multi-QPUs
c On-chip non-local couplers Non-planar error-correcting code
t Optical, room-temperature links Ad-hoc quantum networking

TABLE I. Types of modularity in a long-term scalable quantum system

The high density of qubits in this “dense modular- five types of modularity—classical parallelization, dense
ity” creates a spatial bottleneck for classical I/O and chip-to-chip extension of 2D lattices of qubits (m), sparse
cooling. Proposals to ameliorate this near term include connections with non-trivial topology within a dilution
the development of high-density connectors and cables refrigerator (l), non-local on-chip couplings for error cor-
to route classical signals on and off the chip123,124 , and rection (c), and long-range fridge-to-fridge quantum net-
the addition of time- and frequency-domain multiplexing working (t) (Table I). The optimal characteristic size of
of controls. A longer term approach to address this is each level of modularity is an open question. The individ-
to improve qubit connectivity through the use of a modi- ual “chip-to-chip” modules will still be made as large as
fied gate performed over a long conventional cable125–127 , possible, maximizing fidelity and connection bandwidth.
called l modularity. Beyond allowing us to escape con- Performing calculations on a system like this with multi-
trol and cooling bottlenecks, these long-range couplers ple tiers of connectivity is still a matter of research and
enable the realization of non-2D topologies, thereby not development129,130 .
only reducing the average distance between qubits but Modularity needs to happen not just at the scale of
also opening the door to the exploration of more effi- the QPU, but at all levels of the system. Modular clas-
cient non-2D LDPC error correction codes128 . Develop- sical control systems allow for easy subsystem testing,
ing these long-range couplers thus not only allows us to replacement, and assembly. It’s much easier to build a
scale our near-term systems, but begins to form the basis test infrastructure for a large number of small modules
for how to build quantum systems with mulitple QPUs. each year than a single, re-workable monolith. The same
The technologies that enable both dense modularity can be said of refrigeration, with the added benefit that
and long-range couplers, once developed and optimized, shipping and deploying monolithic large refrigeration sys-
will ultimately be ported back into the qubit chip to en- tems is impractical. A large number of our current fail-
able non-local, non-2D connectivity. These on-chip non- ure points come in I/O and signal delivery, so modular
local c couplers will ultimately allow implementation of solutions where sub-assemblies can be swapped out are
high-rate LDPC codes, bringing our long-term visions to essential. The challenge here is moving the replaceable
completion. unit from a single unit (a cable) to a larger unit (a flexible
Finally, connecting multiple quantum computers in an ribbon cable or other cable assembly).
ad-hoc way will allow us to create larger systems as While the jury is still out on module size and other
needed. In this “quantum networking” approach, the sig- hardware details, what is certain is that the utility of
nals are typically envisioned to leave the dilution refriger- any quantum computer is determined by its ability to
ator, enabled by long-term technological advancements in solve useful problems with a quantum advantage while its
microwave-to-optical transduction using photonic t links adoption relies on the former plus our ability to separate
between different fridges. its use from the intricacies of its hardware and physics-
With these four forms of modularity, we can redefine level operation. Ultimately, the power provided by the
“scale” for a quantum system by hardware is accessed through software that must enable
flexible, easy, intuitive programming of the machines.
n = ([(q m) l]t) p

where n is the number of qubits in the entire modular

and parallelized quantum system. The system is com- IV. THE QUANTUM STACK
prised of QPUs made from m chips, each QPU having
q × m qubits. The QPUs can be connected with l t quan- For quantum computing to succeed in changing what it
tum channels (quantum parallelization), with l of them means to compute, we need to change the architecture of
being microwave connections and t optical connections. computing. Quantum computing is not going to replace
Finally, to enable things like circuit cutting and speed- classical computing but rather become an essential part
ing up error mitigation, each of these multi-chip QPUs of it. We see the future of computing being a quantum-
can support classical communication, allowing p classical centric supercomputer where QPUs, CPUs, and GPUs all
parallelizations. work together to accelerate computations. In integrating
A practical quantum computer will likely feature all classical and quantum computation, it is important to
 13

FIG. 5. Circuits can be represented at various levels. Unitary blocks represent circuits from libraries. These can be decomposed
into parameterized circuits using the universal set of gates. Parameterized physical circuits use the physical gates supported
by the hardware, while scheduled circuits specify timing, calibrations, and pulse shapes.

identify (1) latency, (2) parallelism (both quantum and Before we go into the stack, we need to redefine a quan-
classical), and (3) what instructions should be run on tum circuit. Here we define a quantum circuit as follows:
quantum vs. classical processors. These points define
different layers of classical and quantum integration.
A quantum circuit is a computational routine consisting
 14

of coherent quantum operations on quantum data, such

as qubits, and concurrent (or real-time) classical com-
putation. It is an ordered sequence of quantum gates,
measurements, and resets, which may be conditioned on
and use data from the real-time classical computation. If
it contains conditioned operations, we refer to it is as a
dynamic circuit. It can be represented at different levels
of detail, from defining abstract unitary operations down
to setting the precise timing and scheduling of physical
operations.

This is general enough to represent the circuit

model131 , the measurement model132 , and the adiabatic
model133 of computation, and special routines such as
FIG. 6. The quantum software stack is comprised of four
teleportation. Furthermore, it can represent the circuit
layers, each targeting the most efficient execution of jobs at
at various levels: unitary (unitary block that could rep- different levels of detail. The bottom layer focuses on the ex-
resent circuit libraries such as quantum phase estima- ecution of quantum circuits. Above it, the quantum runtime
tion, classical functions, etc.), standard decomposition efficiently integrates classical and quantum computations, ex-
(reduced to a universal set of gates or expressing the clas- ecutes primitive programs, and implements error mitigation
sical functions as reversible gates), parameterized phys- or correction. The next layer up (quantum serverless) pro-
ical circuits (using the physical gates supported by the vides the seamless programming environment that delivers
hardware, possibly including ancilla qubits not used in integrated classical and quantum computations through the
the circuit, or parameters that are easy to update in cloud without burdening developers with infrastructure man-
real-time), and scheduled circuits (complete timing in- agement. Finally, the top layer allows users to define work-
flows and develop software applications.
formation, calibrated gates, or gates with assigned pulse
shape) (see fig. 5). OpenQASM119 is an example inter-
mediate representation for this extended quantum circuit
and can represent each of these various abstractions. tribution of the output. The estimator allows users to
With this extended quantum circuit definition, it is efficiently calculate expectation values of observables.
possible to define a software stack. Fig. 6 shows a high The circuit sent to the runtime would be a parameter-
level view of the stack, where we have defined four impor- ized physical circuit. The software would perform a run-
tant layers: dynamic circuits, quantum runtime, quan- time compilation and process the results before returning
tum serverless, and software applications. At the low- the corrected outcome. The runtime compilation would
est level, the software needs to focus on executing the update the parameters, add error suppression techniques
circuit. At this level, the circuit is represented by con- such as dynamical decoupling, perform time-scheduling
troller binaries that will be very dependent on the super- and gate/operation parallelization, and generate the con-
conducting qubit hardware, supported conditional oper- troller code. It would also process the results with error
ations and logic, and the control electronics used. It will mitigation techniques, and in the future, error correc-
require control hardware that can move data with low tion. Both today’s error mitigation and tomorrow’s er-
latency between different components while maintain- ror correction will place strong demands on the classical
ing tight synchronization. For superconducting qubits, computing needed inside these primitive programs. The
real-time classical communication will require a latency circuit execution time could be as low as 100 microsec-
of ∼100 nanoseconds. To achieve this latency, the con- onds (maybe even 1 microsecond for error correction),
trollers will be located very close to the QPU. Today, the which is not possible over the cloud. It will need to be
controllers are built using FPGAs to provide the flexibil- installed as part of the quantum computer. Fortunately,
ity needed, but as we proceed to larger numbers of qubits error mitigation is pleasingly parallel, thus using multi-
and more advanced conditional logic, we will need ASICs ple QPUs to run a primitive will allow the execution to
or even cold CMOS. be split and done in parallel.
We refer to the next level up as the quantum runtime At the third level, we imagine software that can com-
layer. This is the core quantum computing layer. In the bine advanced classical calculations with quantum calcu-
most general form, we expect a quantum computer to lations. As described earlier in this paper, introducing
run quantum circuits and generate non-classical proba- classical computing can enable ideas such as circuit knit-
bility distributions at their outputs. Consequently, much ing. Here we need to be able to call quantum primitive
of the workloads are sampling from or estimating prop- programs as well as perform classical calculations such as
erties of distributions. The quantum runtime thus needs circuit partitions. We call this a workflow (fig. 7 shows
to include at least two primitive programs: the sampler examples of workflows for circuit knitting). We refer to
and the estimator. The sampler collects samples from quantum serverless as the software architecture and tool-
a quantum circuit to reconstruct a quasi-probability dis- ing that supports this in a way that allows developers to
 15

FIG. 7. Example of a quantum serverless architecture integrating quantum and classical computations. Quantum runtimes
are illustrated by estimator primitives. Cloud computing is illustrated by general classical computing. Specialized classical
computing such as high precision computing (HPC) or graphics processing units (GPUs) could be integrated into the serverless
architecture. In circuit cutting, a larger circuit is split into many smaller circuits using a specialized classical computer. For each
of the smaller circuits, an estimator primitive is executed (E1 , · · ·, EN ) and if needed, a classical computing routine could be
used to condition future circuits on the results of previous estimators. The process can be repeated as needed. In entanglement
forging, a 2N-qubit wavefunction is decomposed into a larger number of N-qubit circuits. The entanglement synthesis may need
to be offloaded to specialized classical processors. For each N-qubit circuit, an estimator EN is executed and combined to give
the global outcome. This process could be repeated if used in a variational algorithm. Quantum embedding separates sub-parts
of a problem that can be simulated classically from those computationally most costly and requiring quantum computations.
A specialized classical computer could be used to condition the problem on previous outcomes. The quantum simulations
employ estimators EN running on QPUs. The estimators can condition quantum circuits on previous outcomes with classical
calculations run on the general classical processors. Collectively, this set of tools allows larger systems to be simulated with
higher accuracy.

focus only on code and not on the classical infrastruc- mitigation, and eventually, error correction into a run-
ture. Along with circuit knitting, this layer will also allow time environment that returns a simplified application
advanced circuit compiling that could include synthesis, programming interface (API) to the next layer.
layout and routing, and optimization—all of which are The algorithm developer combines quantum runtime
parts of the circuit reduction that should happen before with classical computing, implements circuit knitting,
sending the circuit to execute. and builds heuristic quantum algorithms and circuit li-
Finally, at the highest level of abstraction, the com- braries. The purpose is to enable quantum advantage.
puting platform must allow users to efficiently develop Finally, as we demonstrate examples of quantum advan-
software applications. These applications may need ac- tage, the model developer will be able to build software
cess to data and to resources not needed by the quantum applications to find useful solutions to complex problems
computation itself but needed to provide the user an an- in their specific domain, enabling enterprises to get value
swer to a more general problem. from quantum computing. Fig. 8 summarizes the types
Each layer of the software stack we just described of developers addressed by each layer of the software
brings different classical computing requirements to stack and the time scales involved depending on the type
quantum computing and defines a different set of needs of job being executed and how close to the hardware each
for different developers. Quantum computing needs to developer is working.
enable at least three different types of developers: ker- In putting all of this together and scaling to what
nel, algorithm, and model developers. Each developer we call a quantum-centric supercomputer, we do not see
creates the software, tools, and libraries that feed the quantum computing integrating with classical computing
layers above, thereby increasing the reach of quantum as a monolithic architecture. Instead, fig. 9 illustrates an
computing. architecture for this integration as a cluster of quantum
The kernel developer focuses on making quantum cir- computational nodes coupled to classical computing or-
cuits run with high quality and speed on quantum hard- chestration. The darker the color, the closer the classical
ware. This includes integrating error suppression, error and quantum nodes must be located to reduce latency.
 16

FIG. 8. The time scales and resources involved in quantum computing depend on the needs of the different types of developers
and the level of abstraction at which they work. Quantum researchers and kernel developers work closer to the hardware while
model developers require the highest level of software abstraction.

V. CONCLUSION

In conclusion, we have charted how we believe that

quantum advantage in some scientifically relevant prob-
lems can be achieved in the next few years. This mile-
stone will be reached through (1) focusing on problems
that admit a super-polynomial quantum speedup and
advancing theory to design algorithms—possibly heuris-
tic—based on intermediate depth circuits that can out-
perform state-of-the-art classical methods, (2) the use of
a suite of error mitigation techniques and improvements
in hardware-aware software to maximize the quality of
the hardware results and extract useful data from the
FIG. 9. Model of a cluster-like architecture integrating classi- output of noisy quantum circuits, (3) improvements in
cal processors with QPUs to address latency, parallelization, hardware to increase the fidelity of QPUs to 99.99% or
and the distribution of instructions among classical and quan- higher, and (4) modular architecture designs that allow
tum processors. The darker the color, the lower the latency parallelization (with classical communication) of circuit
required. execution. Error mitigation techniques with mathemati-
cal performance guarantees, like PEC, albeit carrying an
exponential classical processing cost, provide a mean to
quantify both the expected run time and the quality of
processors needed for quantum advantage. This is the
near-term future of quantum computing.
Progress in the quality and speed of quantum systems
Threaded runtimes can execute primitives on multiple will improve the exponential cost of classical processing
controllers. Classical communication in real time be- required for error mitigation schemes, and a combina-
tween the controllers can be used to enable things like tion of error mitigation and error correction will drive a
circuit cutting. The figure also shows how future QPUs gradual transition toward fault-tolerance. Classical and
with quantum parallelization (l and t couplers) can be quantum computations will be tightly integrated, orches-
controlled by a single controller. We imagine that there trated, and managed through a serverless environment
could be workloads that need near-time classical com- that allows developers to focus only on code and not in-
munication (i.e., calculations based on the outcome of frastructure. This is the mid-term future of quantum
circuits that must complete in around 100 microseconds) computing.
or to share states between the primitives, enabled by a Finally, we have seen how realizing large-scale quan-
data fabric. Finally, the orchestration would be respon- tum algorithms with polynomial run times to enable the
sible for workflows, serverless, nested programs (libraries full range of practical applications requires quantum er-
of common classical+quantum routines), the circuit knit- ror correction, and how error correction approaches like
ting toolbox, and circuit compilation. the surface code fall short of the long term needs owing to
 17

their inefficiency in implementing non-Clifford gates and 16 D. Gottesman, “Fault-tolerant quantum computation with con-
poor encoding rate. We outlined a way forward provided stant overhead,” arXiv preprint arXiv:1310.2984 (2013).
17 N. P. Breuckmann and J. N. Eberhardt, “Quantum low-density
by the development of more efficient LDPC codes with
parity-check codes,” PRX Quantum 2, 040101 (2021).
a high error threshold, and the need for modular hard- 18 N. Baspin and A. Krishna, “Quantifying nonlocality: How out-
ware with non-2D topologies to allow the investigation performing local quantum codes is expensive,” arXiv preprint
of these codes. This more efficient error correction is the arXiv:2109.10982 (2021).
19 S. Lloyd, “Universal quantum simulators,” Science , 1073–1078
long-term future of quantum computing.
(1996).
20 A. W. Harrow, A. Hassidim, and S. Lloyd, “Quantum algorithm

for linear systems of equations,” Physical Review Letters 103,

ACKNOWLEDGMENTS 150502 (2009).
21 S. Lloyd, S. Garnerone, and P. Zanardi, “Quantum algorithms

for topological and geometric analysis of data,” Nature commu-

We thank Kristan Temme, Abhinav Kandala, Ewout nications 7, 1–7 (2016).
22 C. Gyurik, C. Cade, and V. Dunjko, “Towards quan-
van den Berg, Jerry Chow, Antonio Córcoles, Ismael
tum advantage via topological data analysis,” arXiv preprint
Faro, Blake Johnson, Tushar Mittal, and Matthias Stef- arXiv:2005.02607 (2020).
fen for their assistance reviewing this manuscript. 23 S. Ubaru, I. Y. Akhalwaya, M. S. Squillante, K. L. Clark-

son, and L. Horesh, “Quantum topological data analysis

1 “IBM
with linear depth and exponential speedup,” arXiv preprint
unveils world’s first 2 nanometer chip technology, opening arXiv:2108.02811 (2021).
a new frontier for semiconductors,” https://newsroom.ibm.com/ 24 D. Aharonov, V. Jones, and Z. Landau, “A polynomial quan-
2021-05-06-IBM-Unveils-Worlds-First-2-Nanometer-Chip- tum algorithm for approximating the Jones polynomial,” Algo-
Technology-Opening-a-New-Frontier-for-Semiconductors. rithmica 55, 395–421 (2009).
2 A. Reuther, P. Michaleas, M. Jones, V. Gadepally, S. Samsi,
25 V. Giovannetti, S. Lloyd, and L. Maccone, “Quantum random
and J. Kepner, “Survey of machine learning accelerators,” in access memory,” Physical Review Letters 100, 160501 (2008).
2020 IEEE High Performance Extreme Computing Conference 26 R. Feynman, “Simulating physics with computers,” Interna-
(HPEC) (IEEE, 2020) pp. 1–12. tional Journal of Theoretical Physics 21, 467–488 (1982).
3 D. S. Abrams and S. Lloyd, “Quantum algorithm providing ex-
27 A. M. Kaufman, M. E. Tai, A. Lukin, M. Rispoli, R. Schit-
ponential speed increase for finding eigenvalues and eigenvec- tko, P. M. Preiss, and M. Greiner, “Quantum thermalization
tors,” Physical Review Letters 83, 5162–5165 (1999). through entanglement in an isolated many-body system,” Sci-
4 A. W. Harrow, A. Hassidim, and S. Lloyd, “Quantum algorithm
ence 353, 794–800 (2016).
for linear systems of equations,” Physical Review Letters 103, 28 O. Shtanko and R. Movassagh, “Unitary subharmonic response
150502 (2009). and floquet Majorana modes,” Physical Review Letters 125,
5 P. W. Shor, “Algorithms for quantum computation: Discrete
086804 (2020).
logarithms and factoring,” in Proceedings 35th annual sympo- 29 I. Aleiner, F. Arute, K. Arya, J. Atalaya, R. Babbush, J. C.
sium on foundations of computer science (Ieee, 1994) pp. 124– Bardin, R. Barends, A. Bengtsson, S. Boixo, A. Bourassa, et al.,
134. “Accurately computing electronic properties of materials using
6 L. K. Grover, “A fast quantum mechanical algorithm for
eigenenergies,” arXiv preprint arXiv:2012.00921 (2020).
database search,” in Proceedings of the Twenty-eighth Annual 30 K. G. Vollbrecht and J. I. Cirac, “Quantum simulators,
ACM Symposium on Theory of Computing, STOC ’96 (ACM, continuous-time automata, and translationally invariant sys-
New York, NY, USA, 1996) pp. 212–219. tems,” Physical Review Letters 100, 010501 (2008).
7 A. M. Childs, D. Maslov, Y. Nam, N. J. Ross, and Y. Su, 31 D. Nagaj and P. Wocjan, “Hamiltonian quantum cellular au-
“Toward the first quantum simulation with quantum speedup,” tomata in one dimension,” Physical Review A 78, 032311
Proceedings of the National Academy of Sciences 115, 9456– (2008).
9461 (2018). 32 A. Kay, “Computational power of symmetric hamiltonians,”
8 A. Y. Kitaev, “Fault-tolerant quantum computation by anyons,”
Physical Review A 78, 012346 (2008).
Annals of Physics 303, 2–30 (2003). 33 B. A. Chase and A. J. Landahl, “Universal quantum walks
9 S. B. Bravyi and A. Y. Kitaev, “Quantum codes on a lattice
and adiabatic algorithms by 1d hamiltonians,” arXiv preprint
with boundary,” arXiv preprint quant-ph/9811052 (1998). arXiv:0802.1207 (2008).
10 K. Temme, S. Bravyi, and J. M. Gambetta, “Error mitigation 34 S. Bravyi, M. B. Hastings, and F. Verstraete, “Lieb-robinson
for short-depth quantum circuits,” Physical Review Letters 119, bounds and the generation of correlations and topological quan-
180509 (2017). tum order,” Physical Review Letters 97, 050401 (2006).
11 Y. Li and S. C. Benjamin, “Efficient variational quantum simu-
35 T. J. Osborne, “Efficient approximation of the dynamics of one-
lator incorporating active error minimization,” Physical Review dimensional quantum spin systems,” Physical Review Letters
X 7, 021050 (2017). 97, 157202 (2006).
12 S. Bravyi, G. Smith, and J. A. Smolin, “Trading classical 36 U. Schollwöck, “The density-matrix renormalization group in
and quantum computational resources,” Physical Review X 6, the age of matrix product states,” Annals of Physics 326, 96–
021043 (2016). 192 (2011).
13 T. Peng, A. W. Harrow, M. Ozols, and X. Wu, “Simulating
37 M. Hastings, “Observations outside the light cone: Algorithms
large quantum circuits on a small quantum computer,” Physical for nonequilibrium and thermal states,” Physical Review B 77,
Review Letters 125, 150504 (2020). 144302 (2008).
14 W. Tang, T. Tomesh, M. Suchara, J. Larson, and M. Martonosi, 38 J. Haah, M. B. Hastings, R. Kothari, and G. H. Low, “Quantum
“Cutqc: using small quantum computers for large quantum algorithm for simulating real time evolution of lattice hamilto-
circuit evaluations,” in Proceedings of the 26th ACM Interna- nians,” SIAM Journal on Computing , FOCS18–250 (2021).
tional Conference on Architectural Support for Programming 39 A. M. Childs, A. Ostrander, and Y. Su, “Faster quantum sim-
Languages and Operating Systems (2021) pp. 473–486. ulation by randomization,” Quantum 3, 182 (2019).
15 K. Mitarai and K. Fujii, “Constructing a virtual two-qubit gate
40 D. W. Berry, C. Gidney, M. Motta, J. R. McClean, and R. Bab-
by sampling single-qubit operations,” New Journal of Physics bush, “Qubitization of arbitrary basis quantum chemistry lever-
23, 023021 (2021).
 18

aging sparsity and low rank factorization,” Quantum 3, 208 that created it,” New Journal of Physics 17, 023037 (2015).
(2019). 65 A. G. Fowler and S. J. Devitt, “A bridge to lower overhead
41 D. Gottesman, “Class of quantum error-correcting codes satu- quantum computation,” arXiv preprint arXiv:1209.0510 (2012).
rating the quantum Hamming bound,” Physical Review A 54, 66 D. Litinski, “A game of surface codes: Large-scale quantum

1862 (1996). computing with lattice surgery,” Quantum 3, 128 (2019).

42 A. R. Calderbank, E. M. Rains, P. W. Shor, and N. J. Sloane, 67 A. Paetznick and A. G. Fowler, “Quantum circuit optimization

“Quantum error correction and orthogonal geometry,” Physical by topological compaction in the surface code,” arXiv preprint
Review Letters 78, 405 (1997). arXiv:1304.2807 (2013).
43 R. Gallager, “Low-density parity-check codes,” IRE Transac- 68 D. Litinski, “Magic state distillation: Not as costly as you

tions on Information Theory 8, 21–28 (1962). think,” Quantum 3, 205 (2019).

44 J.-P. Tillich and G. Zémor, “Quantum LDPC codes with posi- 69 B. J. Brown, “A fault-tolerant non-Clifford gate for the surface

tive rate and minimum distance proportional to the square root code in two dimensions,” Science Advances 6, eaay4929 (2020).
of the blocklength,” IEEE Transactions on Information Theory 70 S. Endo, S. C. Benjamin, and Y. Li, “Practical quantum error

60, 1193–1202 (2013). mitigation for near-future applications,” Physical Review X 8,

45 M.-H. Hsieh and F. Le Gall, “NP-hardness of decoding quantum 031027 (2018).
error-correction codes,” Physical Review A 83, 052331 (2011). 71 A. Strikis, D. Qin, Y. Chen, S. C. Benjamin, and Y. Li,
46 K.-Y. Kuo and C.-C. Lu, “On the hardness of decoding quan- “Learning-based quantum error mitigation,” PRX Quantum 2,
tum stabilizer codes under the depolarizing channel,” in 2012 040330 (2021).
International Symposium on Information Theory and its Ap- 72 R. Harper, S. T. Flammia, and J. J. Wallman, “Efficient learn-

plications (IEEE, 2012) pp. 208–211. ing of quantum noise,” Nature Physics 16, 1184–1188 (2020).
47 P. Iyer and D. Poulin, “Hardness of decoding quantum stabilizer 73 S. T. Flammia, “Averaged circuit eigenvalue sampling,” arXiv

codes,” IEEE Transactions on Information Theory 61, 5209– preprint arXiv:2108.05803 (2021).
5223 (2015). 74 E. van den Berg, Z. Minev, A. Kandala, and K. Temme, “Prob-
48 R. Raussendorf and J. Harrington, “Fault-tolerant quantum abilistic error cancellation with sparse Pauli-Lindblad models
computation with high threshold in two dimensions,” Physical on noisy quantum processors,” arXiv preprint arXiv:2201.09866
Review Letters 98, 190504 (2007). (2022).
49 A. G. Fowler, A. M. Stephens, and P. Groszkowski, “High- 75 A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink,

threshold universal quantum computation on the surface code,” J. M. Chow, and J. M. Gambetta, “Hardware-efficient vari-
Physical Review A 80, 052312 (2009). ational quantum eigensolver for small molecules and quantum
50 D. S. Wang, A. G. Fowler, and L. C. Hollenberg, “Surface code magnets,” Nature 549, 242–246 (2017).
quantum computing with error rates over 1%,” Physical Review 76 A. Kandala, K. Temme, A. D. Córcoles, A. Mezzacapo, J. M.

A 83, 020302 (2011). Chow, and J. M. Gambetta, “Error mitigation extends the com-
51 S. Bravyi, D. Poulin, and B. Terhal, “Tradeoffs for reliable putational reach of a noisy quantum processor,” Nature 567,
quantum information storage in 2D systems,” Physical Review 491–495 (2019).
Letters 104, 050503 (2010). 77 Y. Kim, C. J. Wood, T. J. Yoder, S. T. Merkel, J. M. Gambetta,
52 P. Panteleev and G. Kalachev, “Asymptotically good quantum K. Temme, and A. Kandala, “Scalable error mitigation for
and locally testable classical LDPC codes,” STOC 2022: Pro- noisy quantum circuits produces competitive expectation val-
ceedings of the 54th Annual ACM SIGACT Symposium on The- ues,” (2021).
ory of Computing , 375 (2022). 78 W. J. Huggins, S. McArdle, T. E. O’Brien, J. Lee, N. C. Ru-
53 H. Bombı́n, “Single-shot fault-tolerant quantum error correc- bin, S. Boixo, K. B. Whaley, R. Babbush, and J. R. McClean,
tion,” Physical Review X 5, 031043 (2015). “Virtual distillation for quantum error mitigation,” Phys. Rev.
54 A. Kubica and M. Vasmer, “Single-shot quantum error correc- X 11, 041036 (2021).
tion with the three-dimensional subsystem toric code,” arXiv 79 B. Koczor, “Exponential error suppression for near-term quan-

preprint arXiv:2106.02621 (2021). tum devices,” Phys. Rev. X 11, 031057 (2021).
55 E. T. Campbell, “A theory of single-shot error correction for 80 S. Endo, Z. Cai, S. C. Benjamin, and X. Yuan, “Hybrid
adversarial noise,” Quantum Science and Technology 4, 025006 quantum-classical algorithms and quantum error mitigation,”
(2019). Journal of the Physical Society of Japan 90, 032001 (2021).
56 S. Bravyi and R. König, “Classification of topologically pro- 81 C. Piveteau, D. Sutter, S. Bravyi, J. M. Gambetta, and
tected gates for local stabilizer codes,” Physical Review Letters K. Temme, “Error mitigation for universal gates on encoded
110, 170503 (2013). qubits,” Physical Review Letters 127, 200505 (2021).
57 M. B. Hastings and J. Haah, “Dynamically generated logical 82 S. Bravyi and D. Gosset, “Improved classical simulation of quan-

qubits,” Quantum 5, 564 (2021). tum circuits dominated by Clifford gates,” Physical Review Let-
58 J. E. Moussa, “Transversal Clifford gates on folded surface ters 116, 250501 (2016).
codes,” Physical Review A 94, 042316 (2016). 83 S. Bravyi, D. Browne, P. Calpin, E. Campbell, D. Gosset, and
59 V. Kliuchnikov, D. Maslov, and M. Mosca, “Fast and efficient M. Howard, “Simulation of quantum circuits by low-rank stabi-
exact synthesis of single-qubit unitaries generated by Clifford lizer decompositions,” Quantum 3, 181 (2019).
and T gates,” Quantum Information and Computation 13, 607 84 M. Lostaglio and A. Ciani, “Error mitigation and quantum-

(2013). assisted simulation in the error corrected regime,” Phys. Rev.

60 N. J. Ross and P. Selinger, “Optimal ancilla-free Clifford+T Lett. 127, 200506 (2021).
approximation of z-rotations,” Quantum Information and Com- 85 Y. Suzuki, S. Endo, K. Fujii, and Y. Tokunaga, “Quantum

putation 16, 901 (2016). error mitigation for fault-tolerant quantum computing,” PRX
61 J. O’Gorman and E. T. Campbell, “Quantum computation with Quantum 3, 010345 (2022).
realistic magic-state factories,” Physical Review A 95, 032338 86 C. Piveteau and D. Sutter, “Circuit knitting with classical com-

(2017). munication,” arXiv preprint arXiv:2205.00016 (2022).

62 J. Haah, M. B. Hastings, D. Poulin, and D. Wecker, “Magic 87 A. Eddins, M. Motta, T. P. Gujarati, S. Bravyi, A. Mezzacapo,

state distillation with low space overhead and optimal asymp- C. Hadfield, and S. Sheldon, “Doubling the size of quantum
totic input count,” Quantum 1, 31 (2017). simulators by entanglement forging,” PRX Quantum 3, 010309
63 M. B. Hastings and J. Haah, “Distillation with sublogarithmic (2022).
overhead,” Physical Review Letters 120, 050504 (2018). 88 P. Huembeli, G. Carleo, and A. Mezzacapo, “Entanglement
64 Y. Li, “A magic state’s fidelity can be superior to the operations Forging with generative neural network models,” arXiv preprint
 19

arXiv:2205.00933 (2022). and packaging,” IEEE Microwave Magazine 21, 72 (2020).

89 B. Bauer, D. Wecker, A. J. Millis, M. B. Hastings, and 110 J. M. Kreikebaum, K. P. O’Brien, A. Morvan, and I. Siddiqi,
M. Troyer, “Hybrid quantum-classical approach to correlated “Improving wafer-scale Josephson junction resistance variation
materials,” Phys. Rev. X 6, 031045 (2016). in superconducting quantum coherent circuits,” Superconductor
90 J. M. Kreula, L. Garcı́a-Álvarez, L. Lamata, S. R. Clark, Science and Technology 33, 06LT02 (2020).
111 J. B. Hertzberg, E. J. Zhang, S. Rosenblatt, E. Magesan, J. A.
E. Solano, and D. Jaksch, “Few-qubit quantum-classical sim-
ulation of strongly correlated lattice fermions,” EPJ Quantum Smolin, J.-B. Yau, V. P. Adiga, M. Sandberg, M. Brink, J. M.
Technology 3, 1–19 (2016). Chow, and J. S. Orcutt, “Laser-annealing Josephson junctions
91 S. Bravyi and D. Gosset, “Complexity of quantum impurity for yielding scaled-up superconducting quantum processors,”
problems,” Communications in Mathematical Physics 356, 451– npj Quantum Inf 7, 129 (2021).
112 J. M. Gambetta, C. E. Murray, Y.-K.-K. Fung, D. T. Mc-
500 (2017).
92 G. Knizia and G. K.-L. Chan, “Density matrix embedding: A Clure, O. Dial, W. Shanks, J. W. Sleight, and M. Steffen,
simple alternative to dynamical mean-field theory,” Physical re- “Investigating surface loss effects in superconducting transmon
view letters 109, 186404 (2012). qubits,” IEEE Transactions on Applied Superconductivity 27,
93 G. Knizia and G. K.-L. Chan, “Density matrix embedding: A 1–5 (2017).
113 A. P. M. Place, L. V. H. Rodgers, P. Mundada, B. M. Smitham,
strong-coupling quantum embedding theory,” Journal of chem-
ical theory and computation 9, 1428–1432 (2013). M. Fitzpatrick, Z. Leng, A. Premkumar, J. Bryon, A. Vraji-
94 L. Mineh and A. Montanaro, “Solving the Hubbard model using toarea, S. Sussman, G. Cheng, T. Madhavan, H. K. Babla,
density matrix embedding theory and the variational quantum X. H. Le, Y. Gang, B. Jäck, A. Gyenis, N. Yao, R. J. Cava,
eigensolver,” Physical Review B 105, 125117 (2022). N. P. de Leon, and A. A. Houck, “New material platform for
95 H. Ma, M. Govoni, and G. Galli, “Quantum simulations of ma- superconducting transmon qubits with coherence times exceed-
terials on near-term quantum computers,” npj Computational ing 0.3 milliseconds,” Nature Communications 2021 12:1 12, 1–6
Materials 6, 1–8 (2020). (2021).
96 E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approx- 114 C. R. H. McRae, G. M. Stiehl, H. Wang, S. X. Lin, S. A. Cald-

imate optimization algorithm applied to a bounded occurrence well, D. P. Pappas, J. Mutus, and J. Combes, “Reproducible co-
constraint problem,” arXiv:1412.6062 (2014). herence characterization of superconducting quantum devices,”
97 V. Havlı́ček, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kan- Applied Physics Letters 119, 100501 (2021).
115 L. Geck, A. Kruth, H. Bluhm, S. van Waasen, and S. Heinen,
dala, J. M. Chow, and J. M. Gambetta, “Supervised learning
with quantum-enhanced feature spaces,” Nature 567, 209–212 “Control electronics for semiconductor spin qubits,” Quantum
(2019). Science and Technology 5, 015004 (2019).
98 X. Yuan, S. Endo, Q. Zhao, Y. Li, and S. C. Benjamin, “Theory 116 X. Xue, B. Patra, J. P. G. van Dijk, N. Samkharadze, S. Sub-

of variational quantum simulation,” Quantum 3, 191 (2019). ramanian, A. Corna, B. Paquelet Wuetz, C. Jeon, F. Sheikh,
99 Y. Liu, S. Arunachalam, and K. Temme, “A rigorous and ro- E. Juarez-Hernandez, B. P. Esparza, H. Rampurawala, B. Carl-
bust quantum speed-up in supervised machine learning,” Nature ton, S. Ravikumar, C. Nieva, S. Kim, H.-J. Lee, A. Sammak,
Physics 17, 1013 (2021). G. Scappucci, M. Veldhorst, F. Sebastiano, M. Babaie, S. Peller-
100 J. R. Glick, T. P. Gujarati, A. D. Corcoles, Y. Kim, A. Kandala, ano, E. Charbon, and L. M. K. Vandersypen, “CMOS-based
J. M. Gambetta, and K. Temme, “Covariant quantum kernels cryogenic control of silicon quantum circuits,” Nature 593, 205–
for data with group structure,” arXiv:2105.03406 (2021). 210 (2021).
101 S. Bravyi, D. Gosset, and R. Movassagh, “Classical algorithms 117 C. A. Ryan, B. R. Johnson, D. Ristè, B. Donovan, and T. A.

for quantum mean values,” Nature Physics 17, 337–341 (2021). Ohki, “Hardware for dynamic quantum computing,” Review of
102 N. J. Coble and M. Coudron, “Quasi-polynomial time approx- Scientific Instruments 88, 104703 (2017).
118 J. M. Pino, J. M. Dreiling, C. Figgatt, J. P. Gaebler, S. A.
imation of output probabilities of geometrically-local, shallow
quantum circuits.” 2021 IEEE 62nd Annual Symposium on Moses, M. S. Allman, C. H. Baldwin, M. Foss-Feig, D. Hayes,
Foundations of Computer Science (FOCS) , 598 (2022). K. Mayer, C. Ryan-Anderson, and B. Neyenhuis, “Demonstra-
103 D. Stilck França and R. Garcı́a-Patrón, “Limitations of opti- tion of the trapped-ion quantum CCD computer architecture,”
mization algorithms on noisy quantum devices,” Nature Physics Nature 592, 209–213 (2021).
119 A. Cross, A. Javadi-Abhari, T. Alexander, N. de Beaudrap, L. S.
17, 1221 (2021).
104 S. Barison, F. Vicentini, and G. Carleo, “An efficient quan- Bishop, S. Heidel, C. A. Ryan, J. Smolin, J. M. Gambetta, and
tum algorithm for the time evolution of parameterized circuits,” B. R. Johnson, “OpenQASM 3: A broader and deeper quantum
arXiv preprint arXiv:2101.04579 (2021). assembly language,” arXiv:2104.14722 (2021).
105 C. Zoufal, D. Sutter, 120 A. Gold, J. Paquette, A. Stockklauser, M. J. Reagor, M. S.
and S. Woerner, “Error bounds
for variational quantum time evolution,” arXiv preprint Alam, A. Bestwick, N. Didier, A. Nersisyan, F. Oruc, A. Razavi,
arXiv:2108.00022 (2021). B. Scharmann, E. A. Sete, B. Sur, D. Venturelli, C. J. Winkle-
106 J. M. Gambetta, J. M. Chow, and M. Steffen, “Building logical black, F. Wudarski, M. Harburn, and C. Rigetti, “Entangle-
qubits in a superconducting quantum computing system,” npj ment across separate silicon dies in a modular superconducting
Quantum Information 2017 3:1 3, 1–7 (2017). qubit device,” (2021), arXiv:2102.13293.
107 D. R. W. Yost, M. E. Schwartz, J. Mallek, D. Rosenberg, 121 C. Conner, A. Bienfait, H.-S. Chang, M.-H. Chou, É. Dumur,

C. Stull, J. L. Yoder, G. Calusine, M. Cook, R. Das, A. L. J. Grebel, G. Peairs, R. Povey, H. Yan, Y. Zhong, et al., “Super-
Day, E. B. Golden, D. K. Kim, A. Melville, B. M. Niedziel- conducting qubits in a flip-chip architecture,” Applied Physics
ski, W. Woods, A. J. Kerman, and W. D. Oliver, “Solid-state Letters 118, 232602 (2021).
qubits integrated with superconducting through-silicon vias,” 122 M. P. Larsson and S. Lucyszyn, “A micromachined separable

npj Quantum Information 6, 59 (2020). RF connector fabricated using low-resistivity silicon,” Journal
108 S. K. Tolpygo, V. Bolkhovsky, T. J. Weir, L. M. Johnson, M. A. of Micromechanics and Microengineering 16, 2021–2033 (2006).
Gouker, and W. D. Oliver, “Fabrication process and properties 123 D. B. Tuckerman, M. C. Hamilton, D. J. Reilly, R. Bai, G. A.

of fully-planarized seep-submicron Nb/Al–AlOx /Nb Josephson Hernandez, J. M. Hornibrook, J. A. Sellers, and C. D. Ellis,
junctions for VLSI circuits,” IEEE Transactions on Applied Su- “Flexible superconducting Nb transmission lines on thin film
perconductivity 25, 1 (2015). polyimide for quantum computing applications,” Superconduc-
109 D. Rosenberg, S. Weber, D. Conway, D. Yost, J. Mallek, tor Science and Technology 29, 084007 (2016).
G. Calusine, R. Das, D. Kim, M. Schwartz, W. Woods, J. L. 124 D. J. Reilly, “Engineering the quantum-classical interface of

Yoder, and W. D. Oliver, “Solid-state qubits: 3D integration solid-state qubits,” npj Quantum Information 2015 1:1 1, 1–10
 20

(2015). wards fault-tolerant universal quantum computation,” Nature

125 Y. Zhong, H.-S. Chang, A. Bienfait, É. Dumur, M.-H. Chou, 2017 549:7671 549, 172–179 (2017).
129 N. H. Nickerson, Y. Li, and S. C. Benjamin, “Topological quan-
C. R. Conner, J. Grebel, R. G. Povey, H. Yan, D. I. Schuster,
and A. N. Cleland, “Deterministic multi-qubit entanglement in a tum computing with a very noisy network and local error rates
quantum network,” Nature 2021 590:7847 590, 571–575 (2021). approaching one percent,” Nature Communications 2013 4:1 4,
126 P. Kurpiers, P. Magnard, T. Walter, B. Royer, M. Pechal, 1–5 (2013).
130 C. Monroe, R. Raussendorf, A. Ruthven, K. R. Brown,
J. Heinsoo, Y. Salathé, A. Akin, S. Storz, J.-C. Besse, S. Gas-
parinetti, A. Blais, and A. Wallraff, “Deterministic quantum P. Maunz, L.-M. Duan, and J. Kim, “Large-scale modular
state transfer and remote entanglement using microwave pho- quantum-computer architecture with atomic memory and pho-
tons,” Nature 558, 264–267 (2018). tonic interconnects,” Physical Review A 89, 022317 (2014).
127 N. Leung, Y. Lu, S. Chakram, R. K. Naik, N. Earnest, R. Ma, 131 D. P. DiVincenzo, “Quantum compu-
K. Jacobs, A. N. Cleland, and D. I. Schuster, “Deterministic tation,” Science 270, 255–261 (1995),
bidirectional communication and remote entanglement genera- https://www.science.org/doi/pdf/10.1126/science.270.5234.255.
132 M. A. Nielsen, “Quantum computation by measurement and
tion between superconducting qubits,” npj Quantum Informa-
tion 2019 5:1 5, 1–5 (2019). quantum memory,” Physics Letters A 308, 96–100 (2003).
128 E. T. Campbell, B. M. Terhal, and C. Vuillot, “Roads to- 133 E. Farhi, J. Goldstone, S. Gutmann, and M. Sipser, “Quantum

computation by adiabatic evolution,” arXiv:0001106 (2000).