A systems perspective of quantum computing

Anne Matsuura, Sonika Johri, and Justin Hogaboam

Citation: Physics Today 72, 3, 40 (2019); doi: 10.1063/PT.3.4163

View online: https://doi.org/10.1063/PT.3.4163
View Table of Contents: https://physicstoday.scitation.org/toc/pto/72/3
Published by the American Institute of Physics

ARTICLES YOU MAY BE INTERESTED IN

Machine learning meets quantum physics

Physics Today 72, 48 (2019); https://doi.org/10.1063/PT.3.4164

Quantum computing with neutral atoms

Physics Today 70, 44 (2017); https://doi.org/10.1063/PT.3.3626

Quantum computing ramps up in private sector

Physics Today 73, 22 (2020); https://doi.org/10.1063/PT.3.4612

Quantum foundations
Physics Today 72, 50 (2019); https://doi.org/10.1063/PT.3.4141

Majorana qubits for topological quantum computing

Physics Today 73, 44 (2020); https://doi.org/10.1063/PT.3.4499

Quantum Computing: A Gentle Introduction

Physics Today 65, 53 (2012); https://doi.org/10.1063/PT.3.1442
ISTOCK.COM/NIPLOT
 Anne Matsuura is the director of quantum
applications and architecture, Sonika Johri is a
quantum algorithms research scientist, and Justin
Hogaboam is a senior quantum systems architect
at Intel Labs in Hillsboro, Oregon.

A systems perspective of
QUANTUM
COMPUTING
Anne Matsuura, Sonika Johri,
and Justin Hogaboam

Quantum architects, with knowledge of both

physics and computer engineering, are key

Q
to developing scalable quantum computers.

uantum computing research is coming of age. Recent engineering

advances in quantum bit, or qubit, systems have led to a steady
stream of successful physics experiments that demonstrate
the computational capabilities of small numbers of qubits.
Companies are tackling the difficult task of advancing beyond
small, lab-scale, proof-of-principle devices to build scalable quantum computing
systems that utilize those capabilities. The ultimate goal of a full quantum computing
system of commercially relevant scale—thousands to millions of logical qubits—
is to harness the power of quantum mechanics inside a technology capable of solving
real-world problems that are intractable for classical computers.
Historically, quantum computing research has qubit systems and, more recently, on real qubit
existed in the realm of physics and mathematics systems.3 However, a quantum computing system
and has concentrated on two fundamental areas: will be composed of far more than algorithms and
qubits and quantum algorithms. Qubit research has qubits, just as a classical computer is made of more
focused on the creation, operation, and perfor- than software programs and transistors.
mance benchmarking of qubits in experimental How does one create a quantum computing
devices. Those efforts have been guided by the system that takes a quantum algorithm as input
DiVincenzo criteria, which list the operational and automatically performs a computation on
conditions necessary to demonstrate quantum qubits? Researchers have now begun to define
computing in a physical system.1 Quantum algo- the essential functionalities of the architectural
rithm research has focused on developing algo- layers4,5 without specifying how to build the
rithms2 that can be implemented on abstract ideal remaining components of a fully operating and

MARCH 2019 | PHYSICS TODAY 41

QUANTUM COMPUTING

reproducible quantum computing system. Examples of a few many gates can be run in series within the fidelity, or error,
of those functionalities are illustrated in figure 1. budget?
Opportunities are available for those willing to acquire the After choosing the appropriate logical operations and type
cross-disciplinary skills needed to help build a fully scalable of qubit, the researcher applies the sequence of native gates to
quantum computer. Designers of such a system are asking real physical qubits. From an algorithm designer’s perspec-
thought-provoking questions that cross the boundaries between tive, a two-qubit gate can be applied between any pair of avail-
physics, engineering, and computer architecture. able qubits. However, in many types of systems, those two
qubits must be physically near one another to perform the op-
Application-driven design eration. Thus the algorithm needs to be mapped to the parts of
How does one program and run a quantum algorithm on real the qubit grid capable of executing operations and be se-
qubits? First it is decomposed into a quantum circuit (see box 1) quenced into a schedule of operations that can be optimally im-
comprising a series of logical operations, or gates. Designers plemented, sometimes in parallel, within the limitations of the
find symmetries in the physics that simplify the circuits and quantum device.
enable the algorithms to run on the few physical qubits avail- Mapping and scheduling the algorithm operations onto the
able today. However, quantum algorithms cannot be optimized qubits is difficult because most qubit devices have limited,
completely by computers yet, and so those painstaking calcu- often nearest-neighbor, connectivity, which places constraints
lations must be done entirely by hand. Aided by software tools, on where particular qubit operations can be implemented. Op-
the designer then estimates the number of qubits and gates erations between more distant qubits may require shuttling or
required by analyzing how the accu-
racy of the result is affected by param-
eters such as coherence times, gate
fidelities, and approximations in the
Optimization
algorithm. Optimization tools should Applications/
be developed that eliminate the need Compilers Mapping and scheduling
to simplify algorithms by hand. For
algorithms Fault-tolerant operations
example, they should be able to quickly
generate various circuits that achieve
the same targeted unitary transforma-
tion and then allow the circuit designer
to pick the best one.
Today a number of programming
environments allow compilers to per-
form rudimentary optimization of
Digital and analog control
algorithms and prepare them for re- Quantum Control Error correction
source analysis, simulation, or execu-
tion on a particular type of qubit.
chip electronics Input–output to qubits
Basically, the compilers take a quan-
tum circuit and translate it into a se-
quence of logical quantum instruc-
tions known as quantum assembly
FIGURE 1. FUNCTIONALITIES NECESSARY FOR A QUANTUM COMPUTER are shown here.
language (QASM). The text-format
An application-driven design will help determine how to create a system that will accelerate
language represents the quantum
algorithms from a particular application area. (Applications/algorithms: Jay M. Gambetta, Jerry
circuit as a series of operations. The
M. Chow, and Matthias Steffen, CC BY 4.0; compilers: iStock.com/Bet_Noire; control electronics:
instruction sequence describes what Steve Jurvetson, CC BY 2.0; quantum chip: Yurchanka Siarhei/Shutterstock.com.)
needs to happen to each qubit in order
to run the given algorithm.
From the algorithm designer’s perspective, the QASM code quantum teleportation, which rapidly depletes the quantum
sequence is now ready to execute on a real qubit system. resources available to run the algorithm.6 Mapping and sched-
Unfortunately, all QASM-defined quantum logic gates are not uling protocols are made even more challenging because phys-
necessarily available on every type of qubit system. Qubits can ical qubits are notoriously fragile and noisy, so protocols need
be created out of many materials, and the choice determines to be created that can adapt to those imperfections.
which physical quantum gates are available. Such gates, Ensuring that a quantum computer is initialized to a state
called native gates, are specific to the underlying qubit sys- that is as similar as possible for all algorithmic runs is also im-
tem. The logical QASM operations must therefore be trans- portant because quantum algorithms must be executed repeat-
lated into the corresponding sets of native gates. However, in edly to achieve a statistically meaningful result. The process
doing so, researchers need to consider factors that influence must be performed in a rapid and accurate manner so that it
system design and architecture, such as whether the logical does not introduce an additional source of error into poten-
qubit operations used in the algorithm can run directly on the tially thousands of runs of the quantum circuit.
qubit system. If they can’t, what equivalent gates or gate se- As in classical computing, error correction will be critical
quences can be run? What gates can be run in parallel? How for the reliable functioning of a quantum machine, but classical
42 PHYSICS TODAY | MARCH 2019
error-correcting protocols do not work directly for quantum Since the number of physical qubits is currently limited,
computing. Like bits in classical computers, the information error-mitigation techniques, as opposed to full, rigorous error
encoded in individual qubits can be destroyed by environmen- correction, need to be designed for the NISQ-era devices. One
tal noise. However, quantum computing error sources are more such technique is the removal of errors using additional or
numerous, and error rates are much higher. Quantum error postselected measurements.8 The hope is to eventually build a
correction involves encoding a logical qubit state in multiple fully fault-tolerant quantum computer that is robust to physi-
physical qubits and using measurements and classical comput- cal errors. Such a computer requires qubits with long lifetimes,
ing resources to detect errors quickly enough to correct them high-fidelity gates, and fast feedback—all challenges for the re-
during computation. In today’s Noisy Intermediate-Scale Quan- search field.
tum (NISQ) era,7 the short qubit coherence times and low gate The first quantum computers probably will be coprocessors
fidelities require a high ratio of physical to logical qubits. The coupled with a classical central processing unit. In such a model,
lack of active feedback in most current qubit devices means the quantum computer acts as an accelerator for a particular
that state-of-the-art ones still cannot apply a nondestructive part of the algorithm. In the short term, when error correction
error measurement to a series of qubits, identify the source of is not available, the most promising algorithms are quantum–
errors, determine a correction, and implement the correction classical hybrids, which require an intimate coupling between
within the lifetime of the physical qubits. the classical and quantum processors. Thus architectural design

BOX 1. CODESIGN: QUANTUM FOURIER TRANSFORM

In a quantum circuit, as shown on the left Each SWAP gate, denoted by a verti- of 1.25. (For details, see A. Holmes et al.,
below, each line corresponds to a partic- cal line at the far right of the circuit, https://arxiv.org/abs/1811.02125.) The
ular qubit. Each box represents an opera- exchanges the quantum states of two pseudocode for the algorithm at the bot-
tion, and the chronological order pro- qubits. tom right has an outer loop that executes
ceeds from left to right. Quantum gates The quantum circuit is executed on n n times. Within the outer loop are two
represent the unitary operations that need qubits, and the total number of gates loops that can each be executed in paral-
to be applied on the qubits and are de- scales as O(n2). The O notation classifies lel, which implies O(1) time steps. That al-
noted by specific symbols. Quantum gates algorithms according to how their run- gorithm construction also eliminates the
are the building blocks of a quantum cir- ning time or space requirements grow need in the end for an extra time step
cuit, just as classical logic gates are the with the input size. However, the circuit with SWAP gates.
building blocks of digital circuits in clas- depth—the longest path between algo-
sical computing. This particular diagram rithm input to output—can be calcu-
describes a quantum Fourier-transform lated as O(n) in the optimal, theoretical
(QFT) algorithm. Here H is the Hadamard case, when every qubit is connected
gate, a one-qubit rotation that maps the to every other qubit, as shown on the
qubit basis states |0> and |1> to two su- right. The green circles represent indi-
perposition states: |0> maps to |0> + |1> vidual qubits, and the black lines indi-
and |1> maps to |0> − |1>, where cate where two-qubit gates are possible.
It would be much easier to manufacture
1 1
[ ]
H = 1/√2 1 −1 . a linear array of qubits with nearest-
neighbor connectivity as shown below
A controlled phase gate, C–Pn, acts on on the left. A surprising result is that
two qubits and applies a phase change the QFT can be scheduled to run even on
when they are both in the |1> state: a linear array of qubits with a circuit
C–Pn |j〉 ⊗|k〉 = exp ( 2πi2(kj))|j〉 ⊗|k〉.
n
depth that also has O(n) scaling and
only a small constant-factor overhead

MARCH 2019 | PHYSICS TODAY 43

QUANTUM COMPUTING

features throughout the computing system will likely be a mix help researchers understand the appropriate layers of func-
of classical and quantum pieces. In the design of the instruction tionality required to run the algorithm on real qubits and to de-
sequence, for instance, some researchers advocate for a hybrid sign a scalable quantum computing system that incorporates
of quantum and classical instructions,9,10 while others propose those layers. Even with system noise, the knowledge gained by
that the instruction sequence should be purely quantum.11 running small algorithmic building blocks can help researchers
Whereas the efficient use of coprocessors in computer archi- improve the system organization and architecture, including
tecture is a well understood benefit of modern design, how to the optimal connectivity of the qubits and the appropriate
implement hybrid architecture for quantum computing remains qubit grid organization. By keeping the number of gates low,
an open research question. researchers can run small algorithms on as few as five to seven
Classical electronics for controlling qubits will be an inte- qubits12,13 without using quantum error correction. If all the
gral part of any quantum computing system. Currently, even commonly used components of algorithms from a particular
small qubit systems require racks of laboratory electronics and application area can be run separately, the hope is that full-
numerous wires for controlling and operating qubits in a cryo- scale quantum algorithms can be run on the larger-scale qubit
genic refrigerator or ultrahigh vacuum chamber. As the num- system as the number of qubits and executable circuit depth
ber of qubits scales up, the increase in on-chip and input– are scaled up. An application-driven design will thus result in
output wiring interconnects will introduce more heat and a system architecture that will serve as an accelerator for the
noise to the qubit system. Noise is also easily added during the particular application area. Although difficulties will arise as
often long delay times necessary to send electrical signals to successively larger qubit systems are created, running algo-
the qubits in the chamber. Current qubit hardware uses be- rithms on each generation of quantum hardware will allow re-
tween five and seven input–output cables for each qubit. How- searchers to learn from each new system and solve scaling
ever, that arrangement does not scale beyond a few tens of problems in a systematic manner.
qubits before manufacturers would need to build larger, cus-
tom dilution refrigerators. The problem of interconnect scala- System versus qubit performance
bility for qubit control will be critical to any quantum comput- Traditionally, quantum computing performance has been pri-
ing system of useful size. marily focused only on qubits. The most important quantifica-
Different qubit connectivity layouts and engineering con- tion metrics have been the physical characteristics of the qubit
straints, such as the number of control lines relative to the technologies themselves, such as whether they satisfy the Di-
number of qubits and the parallelization and selectivity for Vincenzo criteria by concentrating on longer qubit coherence
qubit control operations, introduce further restrictions. How- times and on qubit gate performance as measured by gate error
ever, such restrictions offer opportunities for optimization. At rates, gate execution speed, and interconnectivity. However,
Intel Labs, for example, research has shown that the quantum for quantum computing to move beyond physics research to a
Fourier-transform algorithm can be scheduled to execute on a computer technology, researchers need to start thinking in terms
linear array of qubits with nearest-neighbor connectivity al- of overall system performance, which is ultimately what end
most as efficiently as a fully connected qubit system, as shown users will care about most.
in box 1. Research that compares the hardware architectures of two
The key to designing a quantum computing system is to de- systems built from different qubit technologies14 illustrates is-
velop a library of algorithms that are small building blocks of sues from a hardware perspective. However, that perspective
larger, real-world applications in areas such as quantum chem- leaves out many components that are critical to a complete
istry and condensed-matter physics. Those algorithms may quantum computing system beyond individual qubit perfor-
then be used to drive the design of the full quantum computer mance. The field is beginning to move toward approaches such
system down to the physical qubits. Such an approach will as IBM’s quantum volume metric that defines a family of quan-

FIGURE 2. RESEARCHERS CONSIDER FOUR INTERACTING CATEGORIES OF FUNCTIONALITIES when designing a system capable
of running a quantum algorithm automatically on real qubit devices. For example, the connectivity of the qubits, as shown on the right,
may influence the number of control lines that connect to the qubits in the cryogenic refrigerator. The left side shows how the quantum
coprocessor may interact with the classical processor.

44 PHYSICS TODAY | MARCH 2019

 ∣q0〉

BOX 2. CHOOSING A CNOT ∣q1〉

GATE DECOMPOSITION
A logical controlled-NOT,
or CNOT, gate is an en- ∣q0〉 Z RZ(−π/2)
tangling operation that ∣q0〉 RX(π) RX(π)
flips the target qubit be- ½ ½
tween 1 and 0 if the con-
∣q1〉 RX(π/2) Z RX(π/2)
trol qubit is in the 1 state. ∣q1〉 RY(π/2) RZ(−π/2) RY(−π/2)
It must be decomposed
into a sequence of native
quantum gates for the qubit technology formance simulation could provide met- to choose only one to implement the log-
to perform the gate operation on the rics to help choose which CNOT to incor- ical CNOT in the system. Depending on
specific qubit system. Two possible de- porate into the specific design. the performance of the qubits available
compositions are shown, where RX, RY, Depending on the fidelity of the sin- at a particular point in the execution of
and RZ denote rotations around the x-, y- gle- and two-qubit gates in the circuits the algorithm, it is also possible to choose
and z-axes, respectively. A system per- and on their speed, researchers may want a different logical CNOT sequence.

tum circuits for measuring system performance.15 Challenges tem performance simulator before any hardware is built. In the
in the field remain, such as defining a set of broader system case of quantum computing, creating a system performance
performance metrics appropriate for a larger-scale computing simulator is a more computationally expensive task because of
technology and building on the lessons learned from running the superposition of states and the entanglement of qubits. In
realistic algorithms for useful application areas on small qubit addition, the physics of the qubits themselves affects the func-
systems. tioning of the quantum computing system. The challenge is to
As quantum computing systems scale beyond small num- construct a simplified Hamiltonian for a grid of qubits that
bers of qubits running small algorithms, researchers may be when incorporated into the system performance simulator pro-
able to leverage well-established engineering techniques from vides predictive insights for system design decisions without
classical computing to learn how to design and scale quantum exceeding the running time and memory constraints of the
computers to sizes that are useful for more complex quantum simulation.
algorithms. A few of the trade-offs architects consider when de- A system performance simulator for quantum computing
signing a quantum computing system are shown in figure 2. involves two simulator classes. The first, a system simulator,
Algorithm execution time refers to the amount of resources models the software architecture, hardware architecture, and
that must be devoted to compiling and scheduling a sequence control electronics from the compiler down to the classical con-
of quantum gates to run on the specific qubit device technology trol pulses and interacts with the second, a quantum device
and qubit plane organization provided by the hardware. Hard- simulator that mimics the Hamiltonians of few-qubit systems
ware control and execution flow consist of the set of available and the interface with classical control.
quantum gates and the parallelism and degree of individual The first kind of simulator captures basically everything not
qubit control afforded by classical control electronics that will quantum about the quantum computer, such as programming
be key to determining how a system performs. Qubit perfor- languages, the compiler, control schemes, and qubit connectiv-
mance metrics, such as one- and two-qubit gate fidelity, state- ity. Such a simulator may help address quantum architecture
preparation and measurement errors, and coherence times, in- questions such as the following: How many qubits are needed
fluence the maximal executable circuit depth, fault tolerance to enable useful applications? What number of qubits should
approach, and other system performance characteristics. the quantum architecture be able to handle in the next 10–20
Qubit plane organization considers how connected the phys- years? Where should the division between room temperature
ical qubits are to each other. If the qubit plane provides near- and cryogenic control be? Should different elements of the ar-
est-neighbor, two-dimensional planar connectivity between chitecture be constructed of different qubit types?
qubits, it will be possible to implement topological error- The second kind of simulator will consist of the Hamilto-
correction codes. If that degree of connectivity cannot be pro- nian of a small number of qubits. It takes as inputs device-level
vided, then a repetition code or another error-correcting ap- metrics, such as the coherence times, one- and two-qubit gate
proach must be taken, which will affect the amount of fault fidelities, electromagnetic cross talk between qubits, qubit con-
tolerance that the system can provide. nectivity, and the electronics that are used to control the gates.
Since a qubit is often an approximation for a multilevel quan-
The new multidisciplinary field of quantum architecture tum system, a low-level simulator may involve extra levels that
In classical computing, software tools are used to model and have perturbative but nonvanishing effects on the system. The
simulate the functioning of all the components in the system, simulator may also incorporate noise such as charge-trapping
with its limitations and constraints, to enable better hardware defects for semiconductor dots and will allow researchers
design. Numerous alternate designs are first modeled in a sys- to perform functions like optimizing the implementation of
MARCH 2019 | PHYSICS TODAY 45
QUANTUM COMPUTING

common quantum computing algorithms, pinpointing the throughout the quantum computing system require re-
most damaging sources of noise, and debugging quantum searchers with knowledge of both physics and computer archi-
hardware. tecture design. The new field of quantum architecture stands
It would be more powerful to combine the two simulator poised to be an exciting career choice for a new generation of
approaches into one overall quantum computer system per- physicists. It bridges the boundary between quantum and clas-
formance simulator. The joint simulator could run small algo- sical computing and will be key to building truly useful quan-
rithms on a model with, for instance, a specific compiler, qubit tum computers in the future.
control system, type of qubit, and qubit connectivity. It could
also analyze modifications to any of those components and re- The authors gratefully acknowledge contributions from Jim Held and
veal how the changes would affect overall computation. Com- Xiang Zou.
puter architects then could make better design choices at both
classical and quantum levels of the quantum computing sys-
tem. Furthermore, experimentalists and qubit designers could
REFERENCES
1. D. P. DiVincenzo, Fortschr. Phys. 48, 771 (2000).
analyze the impact of their choices regarding qubit type, con- 2. A. Montanaro, npj Quantum Inf. 2, 15023 (2016).
nectivity, and control constraints and better understand how 3. R. Van Meter, C. Horsman, Commun. ACM 56, 84 (2013).
their qubits might be used in a full system environment. 4. F. Chong, D. Franklin, M. Martonosi, Nature 549, 180 (2017).
Box 2 details a particular system-level design choice that 5. N. C. Jones et al., Phys. Rev. X 2, 031007 (2012).
6. M. Oskin et al., in Proceedings of the 30th Annual International Sym-
could be facilitated by a system performance simulation. It posium of Computer Architecture, ACM (2003), p. 374.
shows two ways to decompose a common logical operation, 7. J. Preskill, Quantum 2, 79 (2018).
the CNOT gate—a two-qubit gate that performs a NOT op- 8. J. I. Colless et al., Phys. Rev. X 8, 011021 (2018).
eration on the second qubit only when the first qubit is in 9. R. S. Smith, M. J. Curtis, W. J. Zeng, https://arxiv.org/abs/1608
state ∣1〉—into physical qubit operations and rotations. For a .03355.
10. A. W. Cross et al., https://arxiv.org/abs/1707.03429.
specific quantum system, the best choice for the physical gate 11. A. JavadiAbhari et al., in Proceedings of the 11th ACM Conference
decomposition will depend on many factors, such as qubit co- on Computing Frontiers, ACM (2014), art. no. 1.
herence times, gate fidelity, the execution time of the gate op- 12. N. M. Linke et al., Phys. Rev. A 98, 052334 (2018).
eration, and qubit connectivity. A simulator that models the en- 13. J. S. Otterbach et al., https://arxiv.org/abs/1712.05771.
14. N. M. Linke et al., Proc. Natl. Acad. Sci. USA 114, 3305 (2017).
tire quantum computing system will help with those design 15. L. S. Bishop et al., Quantum Volume, IBM technical report (4 March
decisions. 2017), available at https://pdfs.semanticscholar.org/650c/3fa2a
The classical–quantum design choices that need to be made 231cd77cf3d882e1659ee14175c01d5.pdf. PT