You are on page 1of 10

review articles

doi:10.1145/ 2494568
system itself, the algorithmic work-
Quantum computer architecture holds the key loads for which it is designed, its speed
and capabilities in meeting those
to building commercially viable systems. goals, its interfaces to the classical con-
trol logic, and the design of the classi-
By Rodney Van Meter and Clare Horsman cal control systems are all the responsi-
bility of quantum computer architects.

A Blueprint
In this article, we review the prog-
ress that has been made in developing
architectures for full-scale quantum
computers. We highlight the process

for Building
of integrating the basic elements that
have already been developed, and in-
troduce the challenges that remain
in delivering on the promise of quan-

a Quantum
tum computing.
The most famous development to
date in quantum algorithms is Shors
algorithm for factoring large numbers

Computer
in polynomial time.33 While the ver-
nacular press often talks of factoring
large numbers in seconds using a
quantum computer, in reality it is not
even possible to discuss the prospec-
tive performance of a system without
knowing the physical and logical clock
speed, the topology of the interconnect
among the elements, the number of
logical quantum bits (qubits) available
Sma ll -sc ale quantum c o mputing devices built on in the system, and the details of the al-
gorithmic implementationin short,
a variety of underlying physical implementations without specifying the architecture.
exist in the laboratory, where they have been evolving Figure 1 illustrates the impact that ar-
chitecture can have on the bottom-line
for over a decade, and have demonstrated the viability of a quantum computer; here,
fundamental characteristics necessary for building
systems. The challenge lies in extending these systems
key insights
to be large enough, fast enough, and accurate enough
to solve problems that are intractable for classical 
General-purpose quantum computers
capable of efficiently solving difficult
systems, such as the factoring of large numbers and problems will be physically large,
comprising millions or possibly billions of
the exact simulation of other quantum mechanical quantum bits in distributed systems.

systems. The architecture of such a computer will 


Quantum computer architecture matches
available physical qubit technologies to
be key to its performance. Structurally, when built, applications.
a quantum computer will in fact be a hybrid 
The driving force of an architecture is
Illustration by Coh erent Im ages

quantum error correction, guarding


device, with quantum computing units serving as against loss of fragile quantum data.
coprocessors to classical systems. The program, 
Even with quantum computers, constant
much control circuitry, and substantial pre- and post- factors matter as much as asymptotic
computational class; researchers
processing functions will reside on the classical side worldwide are working to bring error-
corrected clock speeds up to operationally
of the system. The organization of the quantum attractive levels.

84 communicatio ns o f th e acm | octo ber 201 3 | vo l . 5 6 | no. 1 0


cred it t k

octob er 2 0 1 3 | vo l. 56 | n o. 1 0 | c om m u n icat ion s of t he ac m 85


review articles

Figure 1. Scaling the classical number field sieve (NFS) vs. Shors quantum algorithm for factoring.37

The horizontal axis is the length of the number to be factored. The steep curve is NFS, with the marked point
at L = 768 requiring 3,300 CPU-years. The vertical line at L = 2048 is NISTs 2007 recommendation for RSA
key length for data intended to remain secure until 2030. The other lines are various combinations of quantum
computer logical clock speed for a three-qubit operation known as a Toffoli gate (1Hz and 1MHz), method of
implementing the arithmetic portion of Shors algorithm (BCDP, D, and F), and quantum computer architecture
(NTC and AC, with the primary difference being whether or not long-distance operations are supported). The
assumed capacity of a machine in this graph is 2L2 logical qubits. This figure illustrates the difficulty of making
pronouncements about the speed of quantum computers.

1 billion years

NIST recommended RSA key length


ar
Shor, 1Hz, BCDP

U -ye
CP
ro n
1 million years
Time to Factor an L-bit Number

pte
zO
GH
1 thousand years
n1
Shor, 1MHz, BCDP
e, i

100 years
v
sie

10 years
eld

one year
er fi

one month Shor, 1MHz, alg. F, NTC arch.


mb
Nu

one day

one hour
Shor, 1MHz, alg. D, AC arch.
100 seconds

one second

100 1000 10000 100000


L (bits)

the architecture used can make the dif- rent classical capabilities, for certain possible even with imperfect systems,
ference between an interesting proof- systems designed to solve certain but our concern here is the design of
of-concept device and an immediate problems. As we begin to design subsystems for executing QEC, which
threat to all RSA encryption. larger quantum computers, it must can be called the quantum computer
In developing a quantum computer be recognized that large systems are microarchitecture.26 Recent prog-
architecture we have much to learn not simply larger versions of small ress in experimentally demonstrated
from classical computer architec- systems. The conceptual stack of building blocks and the implementa-
ture, but with a few important cave- subfields that must all contribute tion of QEC are the first two topics ad-
ats. Foremost among these caveats is to a scalable, real-world machine dressed in this article.
that the delicate nature of quantum is shown in Figure 2, divided into a At the top of the stack, machines
information demands that memory set of layers. In this article, we dis- will be designed for specific work-
elements be very active. Second, long cuss the elements of this structure in loads, to run certain algorithms (such
wires or long-distance connections turn, all leading to the central core of as factoring or simulation) that exist
inside a quantum computer are either quantum computer architecture. in computational complexity classes
nonexistent, requiring nearest neigh- At the bottom of the stack we have believed to be inaccessible to classical
bor, cellular automaton-like transfer the technologies for storing individu- computers. Without these algorithms,
of data, or are at best poor quality, re- al qubits, and processing or transport- there will be no economic incentive to
quiring much effort to transfer even ing them to take part in a larger com- build and deploy machines.
a single qubit from place to place us- putation. How small groups of qubits With context established at both
ing quantum teleportation and error will interconnect is the first problem the top and bottom of the stack, we
management techniques. Thus, the quantum computer architecture must present progress that has been made
principles of classical computer ar- solve. Given the fragility of quantum toward integrated architectures, and
chitecture can be applied, but the an- data, how can many qubits be kept finish with a detailed example of the
swers arrived at likely will differ sub- alive long enough to complete a immense scale-up in size and slow-
stantially from classical architectures. complex computation? The solution down in speed arising from the error
Quantum computer architecture to this problem, the field of quantum correction needs of a full-scale, digi-
as a field remains in its infancy, but error correction (QEC), began with tal quantum computer. We note that
carries much promise for produc- studies demonstrating that arbitrarily the process of writing this article has
ing machines that vastly exceed cur- accurate computation is theoretically been made substantially easier by the

86 communicatio ns o f th e acm | octo ber 201 3 | vo l . 5 6 | no. 1 0


review articles

appearance in the last few years of ex- quantum spin of an electron, atom, for quantum computing over useful
cellent reviews on many architecture- nucleus, or quantum dot, through the lifetimes is an immense experimental
relevant subfields.1,3,4,9,13,19,28 magnetic flux of a micron-scale cur- challenge. While experimental prog-
rent loop, to the state of a photon or ress has been impressive in the last
Qubit Technologies photon group (its polarization, posi- decade, moving away from one- and
At the lowest level of our Figure 2, tion or timing). two-qubit demonstrations, we are still
we have the technological building The accompanying table lists a a very long way away from entangling,
blocks for the quantum computer. selection of qubit technologies that storing, and manipulating qubits on
The first significant attempt to char- have been demonstrated in the labo- anything like the scale of classical
acterize the technology needed to ratory, with examples of the mate- computing and bits. Here, we are able
build a computer came in the mid- rial and the final device given for each to discuss only selected examples of
1990s, when DiVincenzo listed crite- state variable. the various technologies on offer; for
ria that a viable quantum computing Controlling any kind of physical more information, we recommend the
technology must have: (1) a two-lev- system all the way down to the quan- recent survey by Ladd et al.19
el physical system to function as a tum level is difficult, interacting qubits Ion trap devices and optical sys-
qubit;a (2) a means to initialize the with each other but not anything else tems currently lead in the number of
qubits into a known state; (3) a uni- is even harder, and control of systems qubits that can be held in a device,
versal set of gates between qubits; (4)
measurement; and (5) long memory Figure 2. Quantum computer architecture among some sub-fields of quantum computation.
QEC is quantum error correction; FT is fault tolerant.
lifetime.10 These criteria were later
augmented with two communication
criteria, the ability to convert between
Quantum complexity theory
stationary and flying qubits, and the Quantum
ability to transmit the latter between computing
Quantum algorithms theory
two locations.
In any qubit technology, the first Architecture-aware Quantum
algorithm implementation
criterion is the most vital: What is the programming
QEC theory
state variable? Equivalent to the elec- Programming languages
trical charge in a classical computer,
what aspect of the physical system Quantum
System Organization computer
encodes the basic 0 or 1 state? architecture
The initialization, gates, and measure-
ment process then follow from this QEC and FT implementation Quantum
computer
basic step. Many groups worldwide microarchitecture
Classical Interconnection topologies
are currently working with a range of control and floor planning
state variables, from the direction of
Qubit interconnect technologies
Technology
building blocks
a DiVincenzos original criteria was written with
Qubit storage and gate technologies
qubits, on which we concentrate in this article,
but tri-level qutrits or higher-dimension qudit
data elements are also possible.

A few of the many types of qubit technologies available. Different technologies rely on the same or different state variables to hold
quantum data, implemented in different materials and devices. The equivalent technology for standard classical computing is given.

Gate-based Superconducting
Scaled CMOS superconducting circuit (adiabatic
Type (classical) Ion trap Quantum dot Optical circuit circuit computation)
State variable Electrical charge Ion spin Electron spin, energy Photon polarization, Magnetic flux, charge, Magnetic flux
level, or position time, or position or current phase
Material Doped silicon Atoms in free-space Solid-state Optical waveguides, Superconducting Superconducting
electromagnetic field semi-conductor for example, etched in Josephson junction at Josephson junction at
at cryogenic silicon cryogenic temperatures cryogenic temperatures
temperatures
Device gate MOSFET Laser- or vibrational- Laser- or electrically- Beam splitters and Electrically-driven Electrically-controlled
mediated interaction driven exchanges and photon detectors exchanges and couplers
rotations rotations
Maximum > 109 transistors 14 3 8 2 full qubits + 2 special- 8 coupled, 50
demonstrated per chip, 1016 per purpose memories functional?
variables supercomputing
system

octob er 2 0 1 3 | vo l. 56 | n o. 1 0 | c om m u n icat ion s of t he ac m 87


review articles

and controlled and entangled. Ions current in a loop of superconducting silicon crystal lattice cell size. In con-
are trapped in a vacuum by electro- wire (Figure 3). trast, although the atoms themselves
magnetic potentials, and 14-qubit In solid-state technologies, the ex- are of course tiny, ion traps are limited
entanglement has been demonstrat- perimental focus has been on improv- to inter-atom spacing of perhaps tens
ed (the largest entangled state in ing the control of individual qubits of microns for RF and optical control.
any form shown to date).27 Complex rather than growing their numbers, but Nanophotonic systems will require
bench-top linear optical circuits are that focus has begun to shift. Jaw-Shen components tens of microns across,
capable of entangling eight-photon Tsai has noted that superconducting to accommodate the 1.5m wave-
qubit states, and have been shown to qubit memory lifetime has more than length light that is desirable for tele-
perform nontrivial computation over doubled every year since 1999, and has communications and silicon optics.
short timescales.40 Both of these ap- now reached the point where quantum Superconducting flux qubits require
proaches do not scale in those forms, error correction becomes effective. a current ring microns across. All of
but scalable approaches are also un- Overall, the prospects are very good these technologies result in qubit de-
der development. Groups headed by for systems consisting of tens of qubits vices that are macroscopic, or nearly
Wineland, Monroe, Chuang, and oth- to appear in multiple technologies over so, with areal densities a million times
ers have demonstrated the necessary the next few years, allowing experimen- less than computer chips. This fact
building blocks for ion traps.21 Inte- tal confirmation of the lower reaches of will have enormous impact on large-
grated nanophotonics (using photons the scaling behavior of quantum algo- scale architectures, as we will see.
as qubits) made on silicon chips pro- rithms and the effectiveness of quan- First steps in quantum architec-
vides the route to getting optics off of tum error correction. ture. At this lowest level, the job of
the lab bench and into easily scalable However, one factor that is of- quantum architecture is to determine
systems, and is making substantial ten unappreciated when looking at how individual qubits or qubit blocks
progress in groups such as OBriens.19 these qubit technologies is the physi- interconnect and communicate
Solid-state electronic devices for cal scale of the devices, particularly in order to process data. There are
gate-based computation, while cur- in comparison with classical digital three main areas where experimental
rently trailing in terms of size of en- technologies. Many people associate groups have begun to consider archi-
tangled state demonstrated, hold out quantum effects with tiny objects, but tectural implications in designing
great promise for mass fabrication most of these technologies use de- their systems.
of qubits.19 By trapping a single extra vices that are enormous compared to Heterogeneity. Some researchers
electron in a 3D block of semicon- the transistors in modern computer have been investigating technological
ducting material, a quantum dot has a chips. Transistors really are reaching heterogeneity by using combinations
quantum spin relative to its surround- down to atomic scales, with vendors of electron spin, nuclear spin, mag-
ing that can hold a qubit of data. For having shipped chips fabricated with netic flux, and photon polarization in
flux qubits, the state variable is the a 22-nanometer process at the end of a single system.5 It is, however, equally
quantum state of the magnetic flux 2011. In these chips, the smallest fea- important to consider structural het-
generated by a micron-scale ring of tures will be only about 40 times the erogeneity, both in operational capa-
bility and in interconnect speed or
Figure 3. The Josephson junction superconducting flux qubit. fidelity. Martiniss group has recently
demonstrated a chip with a functional
Here, tiny gaps in a superconducting wire (marked with Xs), known as Josephson junctions, force distinction between two flux qubits
the current in the loop to quantum tunnel across the gap, causing the current to be quantized. and memory storage elements, lead-
We can use a counterclockwise current (referred to as spin up, as shown by the large arrow) ing them to refer to it as a quantum
to be a zero, and a clockwise current (spin down) to be a one.
von Neumann architecture.24 In many
Qubit is state of single magnetic flux quantum technologies, including some forms
of quantum dots and Josephson junc-
tion qubits, measurement of a qubit
requires an additional physical device,
consuming die space and making lay-
out of both classical and quantum
components more complex.
Integration and classical control.
Loop of super-conducting wire
Increasing the number of on-chip de-
vices will require improving on-chip
integration of control circuits and
multiplexing of I/O pins to get away
from multiple rack-mount units for
Josephson junctions controlling each individual qubit. Ear-
force current to be quantized ly efforts at considering the on-chip
classical control overhead as systems
grow include Oskins design,30 and

88 com municatio ns o f th e acm | octo ber 201 3 | vo l . 5 6 | no. 1 0


review articles

recent work by Levy et al. details the


on-chip hardwares impact on error
correction.22 Kim uses his expertise in
micro-mirrors to focus on the classi-
Approaches to Quantum
cal control systems for systems requir-
ing switched optical control.16,18 Con-
Information Processing
trolling meso-scale systemstens to Data within a quantum computer is typically stored as quantum bits or qubits. Like a
low hundreds of physical qubitswill classical bit, a qubit has two states 0 and 1 but, unlike a classical bit, a qubit may be in a
superposition of the two states, being in a certain sense in both 0 and 1 at the same time.
require substantial investment in sys- A quantum gate operation that changes the state of this qubit can act on both values
tems engineering. Experimental labo- simultaneously. Each element in the superposition has a (complex) weight, say for 0
ratories likely will have to contract or and for 1. When measuring a superposed state, only a single result (0 or 1) is returned,
hire professional staff with mixed- but the probability of measuring 0 is ||2 and of measuring 1 is ||2. We cannot predict
which outcome we will see, only its relative likelihood.
signal circuit design experience to The power of superposition extends when we consider qubit registers: by analogy
do extensive control circuits, and for with a classical register, many qubits act together to store data as bit strings. In the
many technologies cryogenic circuits quantum case, the register can be in a superposition of all possible register values. For
are required. example, a 3-qubit register can be in the superposed state of all eight values 000 to 111,
all with different weights. In some such cases, when the superposition contains more
Interconnects. Even within individu- than a single qubit, the qubits can be entangled with each other. The individual qubits
al chips, we are already seeing the be- no longer act independently, and exhibit much more strongly correlated behavior than is
ginning of designs with multiple types possible for classical systems. As with a single qubit, when quantum gates are performed
on a register, operations are performed on all values simultaneously.
of interconnects. Integration levels will Extracting the relevant data is the difficult part of quantum computing. Only one
likely reach hundreds to low thousands element of a superposition can ever be measured, and we cannot choose which one it
of qubits in a single chip or subsystem, is. Algorithm designers aim to manipulate the weights so that, when the time comes to
but reaching millions to billions of de- measure the result, the majority of the weight is given to a state that is a solution to the
input problem.
vices in a system will require intercon- Several different methods have been developed to use these fundamental elements
nects that remain only on the drawing of quantum computing. The most widely considered is circuit-based computation.
board. In some proposals, the inter- Directly analogous with classical digital computation, data is stored in qubits and
subsystem connections are physically manipulated by the application of gate operations. In general, the first step of a circuit-
based computation is to create an equal superposition of all register states. Gate
similar to intra-subsystem connec- operations between qubits then change the weights in the superposition, usually creating
tions, while in others they can be radi- entanglement in the process.
cally different. One ion trap proposal, A separate approach is adiabatic quantum computation. As with the circuit model,
the output state is measured to give the final answer. In this case, however, the state
for example, uses shared, quantum, vi- is designed to be the low energy ground state of a quantum system in the quantum
brational modes for intra-node opera- computer. The key to the computation is to adjust the coupling between quantum
tions and optical connections between systems in the device to allow it to relax into that specific ground state.
separate ion traps.29 Other approaches include measurement-based quantum computation, in which a
large entangled state is reduced to the desired output state simply by carefully choosing
how to measure the qubits, and direct simulation, in which the quantum states are
Error Correction designed to model a different physical system, rather than calculate a value numerically.
The heroic efforts of experimentalists
have brought us to the point where ap-
proximately 10 qubits can be controlled tion known as the no-cloning theo- sands. A key property of a code is its
and entangled. Getting to that stage has rem that means standard, classical, threshold, the accuracy to which all
been a monumental task as the fragile methods of controlling error cannot be physical qubit operations must per-
quantum system must be isolated from implemented.39 Quantum data cannot form in order for the code to work.
the environment, and its state protect- be backed-up, or copied for simple Once enough physical qubits can be
ed from drifting. What will happen, repetition code processing to detect interacted to make interacting logical
then, when we push to hundreds, thou- and correct for errors. qubits, with all physical device opera-
sands, or millions of qubits? The possibility of performing tions accurate to below the threshold,
This brings us to the next level of quantum computation is saved, how- then error correction will keep the
Figure 2, the microarchitecture for ever, by quantum error correction. quantum computer error-free for the
quantum computers. If a quantum ar- Some techniques are based on classi- runtime of the algorithm.
chitecture were designed where gates cal error correction and erasure cor- When assessing a classical error
of an algorithm were run directly on rection, while others are based on correcting code, an important aspect
the types of individual physical qubits uniquely quantum approaches.9 In all is the code rate, the ratio of delivered,
that we have been describing, it would cases, a number of physical qubits are corrected symbols to the number
not work. Both the accuracy of the gate combined to form one or more logical of raw symbols used. High rates are
operations and the degree of isolation qubits. The number of physical qubits achieved in part by using very large
from the environment required to per- per logical qubit is determined by the blocks that encode many bits. Block-
form robust computation of any sig- quantum operation error rates, the based codes have been explored in
nificant size lie far outside the realm physical memory lifetime, and the ac- quantum codes, but suffer from the
of feasibility. To make matters worse, curacy required of the algorithm, and drawback that logical operations on
quantum data is subject to a restric- can vary from tens to possibly thou- logical qubits within the block are dif-

octob er 2 0 1 3 | vo l. 56 | n o. 1 0 | c om m u n icat ion s of t he ac m 89


review articles

ficult to execute fault tolerantly. When types of noise. Techniques known as of the base physical error rate on the
selecting a quantum code, the rate is spin echo and dynamic decoupling architecture and performance, and
important, but the demands made on similarly can be used to partially re- will be discussed later.
the implementing hardware and the verse the impact of systematic ef- Efficient ancilla factories. Most nu-
ease of logical operations are critical. fects on memory, going with the meric quantum algorithms depend
Performing error correction is error for a while and against it for a heavily on a three-qubit gate called
not a simple task; in fact, the vast while, canceling out the effect. Puri- a controlled-controlled NOT gate, or
majority of the processing power ficationerror detection for specific Toffoli gate. In most quantum error
of a universal quantum computer stateswill be especially useful for correction paradigms, direct execu-
will be used to correct errors in the communication, either system inter- tion of a Toffoli gate on encoded logi-
quantum state. Application algo- nal or external. cal qubits is not possible. Instead, the
rithms and data processing make The implementation of error cor- Toffoli gate is performed using sev-
their appearance only at a level rection is perhaps the key near-term eral operations, including one that
well above the real-time, (physical) experimental goal. As experimental consumes a specially prepared an-
qubit-by-qubit work of error correc- capabilities have grown, groups have cilla (temporary variable) state. The
tion. An architecture for a universal begun competing to demonstrate ancilla is created using distillation (a
quantum computer will therefore quantum error correction in increas- quantum error detection code), which
have as its primary goal the execution ingly complete forms. Blatts group takes noisy physical states and builds
of specific types of error correction. performed multiple rounds of an error more accurate logical states. Cre-
The earliest ideas for QEC natu- correction-like circuit that detects and ation of these states may dominate
rally took advantage of classical error corrects certain types of errors using the workload of the machine, and re-
correction techniques. After solving certain simplifications.32 Pans group cent work has assumed that 75%90%
the problems of measuring error syn- has recently shown an eight-photon of the machine is dedicated to their
dromes without destroying the quan- entangled state related to the unit cell production. Isailovic et al. referred to
tum state and computing on encoded of 3D surface error correction.40 this need as running quantum appli-
states without inadvertently spreading Microarchitectures for error correc- cations at the speed of data, that is,
errors (known in QC literature as fault tion. As in classical computer architec- producing the generic ancilla states
tolerance, referring to runtime errors ture, microarchitecture is the bridge rapidly enough that they are not the
rather than mid-computation failure between physical device capabilities performance bottleneck.14
of hardware components), application and the architecture. Microarchitec- Balancing error management mecha-
of classical error correction became ture in the quantum case can be un- nisms. A functioning quantum com-
relatively straightforward.9 derstood to be the level dedicated to puter almost certainly will not rely on
A promising form of error correc- efficient execution of quantum error a single type of error correction, but
tion is surface code computation, management, while the system archi- will incorporate different forms of
which grew out of work by Kitaev and tecture is the organization of micro- error correction/detection/suppres-
others on topological quantum com- architecture blocks into a complete sion at different levels. Different error
puting. Raussendorf and collabora- system. There are several specifically management techniques have differ-
tors created the 2D and 3D versions quantum elements that must be con- ent strengths and weaknesses, and the
suitable for solid-state and photonic sidered for this microarchitecture. combinatorial space for integrating
systems, respectively.11,31 Fowler, De- Clock speed. The conversion fac- multiple types is large.
vitt, and others have extended the tor from physical gate cycle to logical Defects. A quantum architecture
practicality of these results, including gate cycle has a strong, underappreci- must also take into account that fab-
implementing the real-time error pro- ated impact on the performance of an rication will inevitably be an imper-
cessing necessary to determine that algorithm. It depends on a number of fect process. Qubits may be declared
the classical half of the machine is a architectural features, as well as the defective because they fail to cor-
tractable engineering problem.8 The error correction code itself. For the rectly hold the correct state variable
code rate of surface codes is poor, but ion trap-based system analyzed by carrier (for example, to trap a single
their requirement only for nearest- Clark et al.,6,26 a 10sec physical gate electron), because memory lifetime
neighbor connections will allow them results in a 1.6msec error correction is short or gate control imprecise, or
to work at a higher physical error rate cycle time using the basic form of because they fail to couple properly
than other methods on some attractive Steanes error correcting code, which to other qubits. For gate-based, error-
hardware platforms. encodes one logical qubit in seven corrected systems, calculations show
Beyond digital quantum error physical ones. The code will need to that a stringent definition of declar-
correction for arbitrary states, other be applied in recursive fashion, re- ing a device to be functional pays
approaches can be used to (partial- sulting in growth of the physical sys- for itself in reduced error correction
ly) isolate qubits from undesirable tem by an order of magnitude and an demands.36 A systems resilience to
interactions. Decoherence-free sub- increase in the logical clock cycle to low yield is very microarchitecture-
spaces encode a logical qubit in the 260msec, not far below the topmost dependent. Alternatively, the digital
phase difference of two physical qu- quantum line in Figure 1. This dra- quantum error correction itself can
bits, suppressing the effect of certain matic increase illustrates the effect be adapted to tolerate loss.34

90 communicatio ns o f th e acm | octo ber 201 3 | vo l . 5 6 | no. 1 0


review articles

Workloads quantum chemistry is growing. In a rection required and type of error man-
So far we have looked at the lower two somewhat unexpected result, Browns agement applied, the corresponding
levels of Figure 2. Before investigating group has shown that certain quantum execution time of logical Toffoli gates
a complete quantum computer archi- simulation algorithms expected to be (including the ancilla distillation dis-
tecture, we need to consider the algo- computationally tractable on quan- cussed earlier), and the microarchitec-
rithms and programs that will be run tum computers are turning out to have ture of individual areas of the system.
on the physical hardwarethe work- dismayingly large resource require- The DiVincenzo criteria are funda-
load for the quantum computer. We ments.6 However, the field of quantum mental and necessary, but not sufficient
therefore skip to the top of the stack: simulation is varied, and these simula- to build a practical large-scale system.
quantum programming and quantum tors remain attractive; they are simply Considering instead the issues of quan-
algorithms. We are still in the time going to take more quantum computa- tum computer architecture results in a
of Babbage, trying to figure out what tional resources (hence, more calendar different focus, highlighting such mun-
Knuth, Lampson, and Torvalds will do years to develop and more dollars to dane engineering criteria as being large
with a quantum computer. It has been deploy) than originally hoped. enough and fast enough to be useful,
widely believed that Shors factoring al- A key element of the process of de- and small enough and cheap enough
gorithm33 and quantum simulation3,4,20 veloping applications will be program- to be built. Very loosely, meeting the
will provide the two driving reasons to ming tools for quantum computers, DiVincenzo criteria can be viewed as the
build machines. There are, however, a and enough language designs were responsibility of experimental physi-
number of other useful and interesting under way by 2005 to warrant a survey cists, while the latter criteria are the re-
quantum algorithms, seven of which with a couple of hundred references.13 sponsibility of computer engineers.
are being investigated by teams in- In what are arguably the first examples Small-scale quantum architecture
volved in IARPAs Quantum Computer of true quantum programming, as can be said to have begun with Lloyds
Science Program.b Bacon and van Dam1 distinct from quantum algorithm de- 1993 proposal for a molecular chain
and Mosca28 have published surveys sign, shortly after the publication of computer,23 the first for a potentially
covering quantum random walks, game Shors factoring algorithm, Vedral et al. buildable device. The word scalable
theory, linear algebra, group theory, and Beckman et al. produced detailed attained a permanent position in the
and more. Our understanding of how descriptions of the circuits (sequences lexicon with Kielpinski et al.s 2002 pro-
to design new quantum algorithms of gate operations) necessary for the posal for an ion trap that can shuffle and
that asymptotically outperform classi- modular exponentiation portion of the manipulate individual atoms,17 an ap-
cal ones continues to grow, though the algorithm, which appeared as a single proach that continues to pay dividends.
number of people who can put the con- line in Shors original paper.2,38 The These ideas and many others for multi-
cepts into practice remains small. next step in implementation is to adapt qubit devices, such as the quantum
Given the applications we have, such a description for execution on von Neumann approach or scalable ion
how large a computer is needed to run a particular machine, as in the block traps with distinct operation and stor-
them, and how should it be structured? marked Architecture-aware algorithm age sites, sketch local areas of the system
Only a few quantum algorithms have implementation in Figure 2. Match- using the technological building blocks,
been evaluated for suitability for actual ing implementation choices to the but provide no direct guidance on how
implementation. Shors algorithm is strengths of the machine, including to organize large systems that meet the
commonly used as a benchmark, both choosing adder circuits that match the goal of solving one or more problems
for its importance and clarity, and be- application-level system interconnect, that are classically intractable.
cause the arithmetic and quantum Fou- and trading off time and space, will be When considering the macro archi-
rier transform on which it is founded a collaboration between the program- tecture, certain aspects of the design
are valuable building blocks for other mer and the tools.37 Maslov and others become clear. Because all of memory
algorithms.7,12,26,36 Unfortunately, the have studied efficient architecture- is expected to be active, the system
size and speed of a machine needed to aware compilation,25 an important ele- will probably not consist of separate
run the algorithm has been widely mis- ment in the IARPA program. Compiler CPU and memory banks connected via
understood. Architects have suggested backends for specific experimental wide, shared buses. A more uniform
a physical machine comprised of high hardware configurations will soon be array of microarchitecture error cor-
millions to billions of qubits to factor important, as will methods for debug- rection building blocks is the obvious
a 2,048-bit number, a size that experi- ging quantum programs in situ. approach, tempered by issues such as
mental physicists find staggering.7,16,26,36 defective devices and the needs of the
In part because designs for a Shor Quantum System Architectures classical control subsystems. Each
machine have proven to be intimidat- Finally we come to the central element of these microarchitecture building
ingly large, consensus is building that in Figure 2. A complete system design blocks may utilize heterogeneous tech-
a Shor machine will not be the first will specify everything from the busi- nology with an internal storage/com-
commercially practical system, and ness-level requirements for an algo- putation distinction.24 Individual chips
interest in designing machines for rithm and machine capable of outper- or ion traps will not be large enough
forming classical computers, through to execute some algorithms (notably
b http://www.iarpa.gov/solicitations.qcs.html. the details of the algorithms imple- Shor) at scale, likely forcing the adop-
The IARP QCS was terminated in April 2013. mentation, the strength of error cor- tion of a multicomputer structure.15,29,36

octob er 2 0 1 3 | vo l. 56 | n o. 1 0 | c om m u n icat ion s of t he ac m 91


review articles

Large-scale designs are going to be number, and position of the singular


difficult to create and evaluate without factories is dependent on the physical
the appropriate tools. Further invest- error rate and the size of the computa-
ment in automated tools for co-design tion to be performed, and the type of
of internally heterogeneous hardware
and compilation of software is critical. When will a interconnects available. The space for
wiring is a surface code-dependent fac-
One good example of this practice is quantum computer tor, not required when using other error
Svore and Cross, working with Chuang,
who have developed tool chains with do science, rather correction methods, and is probably a
substantial underestimate, though
round-trip engineering and error cor-
rection in mind.35
than be science? researchers are currently looking for
compact compilations of programs on
Architectural analyses exist for ion the surface code that will minimize this
trap systems using Steane error correc- factor. The chosen code distance d is
tion, and multiple, distinct forms of strongly dependent on the application
nanophotonic and solid-state systems algorithm itself and on the physical
using the surface code.7,16,26,36 We next gate error rate. Shors algorithm for L
take up one moderately complete ar- = 2,048 demands that, roughly speak-
chitecture as an example. ing, we must be able to run 1015 logical
operations with a high probability of
An Architecture at Scale correctly executing them all. This work
We can use error management and was done assuming a physical error
application workloads to determine rate of 0.2%, which is not very far below
the broad outlines of a computer that the surface code threshold of 0.75%.
could run a useful algorithm at full These two factors determine the large
scale. The size of a quantum computer distance of 56, and in the case of the
grows depending on the algorithms surface code required resources grow
demand for logical qubits, the quan- as d2, giving the high scale-up factor.11
tum error correction scheme, the gate The final factor of four is strongly de-
and memory error rate, and other fac- pendent on the details of the microar-
tors such as the yield of functional qu- chitecture and the yield.
bits in the system. Overall, including This results in a final system size of
space for various temporary variables six billion physical qubits for the main
and the ancilla state distillation, the quantum state itself, each of which
scale-up factor from logical to physi- must be independently controllable. In
cal qubits can reach half a million. this particular architecture, this stag-
As a specific example, the resource gering number must be augmented
growth in one architecture36 can be as- with additional qubits for communica-
signed approximately as follows: tions, on-chip optical switches, delay
Establish a substantially post- lines, and many external supporting
classical goal of factoring an L= lasers, optical switches, and measure-
2,048-bit number using Shors algo- ment devices, all deployed in a large
rithm, requiring multicomputer configuration.
6L logical qubits to run a time-effi- The performance of the system is
cient form of the algorithm, growing by determined by the error correction
8 to build singular factories for time and the complexity of executing
the state distillation process, allowing the application gates on top of the er-
the algorithm to run at the speed of data, ror correction code. The surface code
1.33 to provide wiring room to cycle time on this architecture for mea-
move logical qubits around within the suring all error syndromes is ~50sec,
system, far slower than the 100psec planned
10, 000 to run the Raussendorf- for physical gates. A Toffoli gate will re-
Harrington-Goyal form of the surface quire ~50mseca factor of 1,000 from
code31 with an error correcting code QEC cycle to logical gate, for this code
distance d = 56, and finally distance. Demonstrating how system-
4 to work around a yield of func- level issues affect performance, the
tional devices of 40%. QEC cycle time is limited by conten-
A singular factory is simply a region tion for access to on-chip waveguides.
of the computer assigned by the pro- In part to address some of these ar-
grammer or compiler to the creation of chitectural limitations, the QuDOS ar-
the ancillae discussed here. The size, chitecture was developed.16 QuDOS, if

92 comm unicatio ns o f the acm | octo ber 201 3 | vo l . 5 6 | no. 1 0


review articles

built, would be 100 times faster, largely Science or Nature in which the point is Architecture for a large-scale ion-trap quantum
computer. Nature 417 (2002), 709711.
due to increased parallelism in mea- the results of a quantum computation, 18. Kim, J. and Kim, C. Integrated optical approach
suring the error syndromes. Overall, rather than the machine itself? That is, to trapped ion quantum computation. Quantum
Information and Computation 9, 2 (2009).
coordinated changes of physical tech- when will a quantum computer do sci- 19. Ladd, T., Jelezko, F., Laflamme, R., Nakamura, Y.,
nology, error correction mechanism, ence, rather than be science? Monroe, C. and OBrien, J. Quantum computers.
Nature 464 (Mar. 2010), 4553.
and architecture may gain 34 orders 20. Lanyon, B.P. Universal digital quantum simulation with
of magnitude in performance, demon- Acknowledgments trapped ions. Science 334, 6052 (2011), 5761.
21. Leibrandt, D. et al. Demonstration of a scalable,
strating the impact of the field of quan- This research is supported by the Cabi- multiplexed ion trap for quantum information
tum computer architecture. net Office, Government of Japan, and processing. Quantum Information and Computation 9,
901 (2009).
the Japan Society for the Promotion 22. Levy, J.E. et al. Implications of electronics constraints
for solid-state quantum error correction and quantum
Conclusion of Science (JSPS) through the Funding circuit failure probability. New Journal of Physics 13, 8
A principal lesson learned so far in re- Program for World-Leading Innova- (2011).
23. Lloyd, S. A potentially realizable quantum computer.
search on quantum computer archi- tive R&D on Science and Technology Science 261 (1993), 15691571.
tectures is that systems capable of solv- (FIRST Program). We thank Simon De- 24. Mariantoni, M. et al. Implementing the quantum von
Neumann architecture with superconducting circuits.
ing classically intractable problems vitt, Dave Farber, Joe Touch, Michio Science 334, 6052 (2011), 6165.
will be large, although the search for Honda, Hajime Tazaki, Shigeya Suzu- 25. Maslov, D., Falconer, S. and Mosca, M. Quantum Circuit
Placement. IEEE Transactions on Computer-Aided
the smallest commercially attractive ki, the referees and editors for careful Design of Integrated Circuits and Systems 27, 4
machine continues. Device sizes will reading of the manuscript. (2008), 752763.
26. Metodi, T.S., Thaker, D.D., Cross, A.W., Chong, F.T. and
limit integration levels, affecting archi- Chuang, I.L. A quantum logic array microarchitecture:
tecture, and determining logical clock Supplemental Material Scalable quantum data movement and computation.
Additional information and references for this article are In Proceedings of the 2005 International Symposium
speed requires making many design available under supplemental material in the ACM Digital on Microarchitecture (2005).
decisions but dramatically affects what Library DOI: 10.1145.249568. 27. Monz, T. et al. 14-qubit entanglement: Creation and
coherence. Phys. Rev. Lett 106, 13 (Mar. 2011).
can and cannot be effectively comput- 28. Mosca, M. Quantum algorithms (2008); Arxiv preprint
ed (as shown in Figure 1). Architectural arXiv:0808.0369.
References 29. Oi, D.K.L., Devitt, S.J. and Hollenberg, L.C.L. Scalable
problems cover a broad range, but have 1. Bacon, D. and van Dam, W. Recent progress in quantum error correction in distributed ion trap computers.
algorithms. Commun. ACM 53, 2 (Feb. 2010), 8493. Physical Review A 74, 052313 (2006).
received only modest amounts of at- 2. Beckman, D., Chari, A.N., Devabhaktuni, S. and Preskill, 30. Oskin, M., Chong, F.T., Chuang, I.L., and Kubiatowicz,
tention compared to near-term experi- J. Efficient networks for quantum factoring. Phys. J. Building quantum wires: The long and short of
Rev. A 54 (1996), 10341063; http://arXiv.org/quant- it. In Proceedings of the 30th Annual International
mental hurdles, leaving much room ph/9602016. Symposium on Computer Architecture (June 2003),
3. Brown, K.L., Munro, W.J. and Kendon, V.M. Using
for high-impact research that can help quantum computers for quantum simulation. Entropy
ACM, N.Y.
31. Raussendorf, R., Harrington, J. and Goyal, K.
guide the focus of experimental work. 12, 11 (2010), 22682307. Topological fault-tolerance in cluster state quantum
4. Buluta, I. and Nori, F. Quantum Simulators. Science
To sharpen the community focus on 326, 5949 (2009), 108111.
computation. New Journal of Physics 9, 199 (2007).
32. Schindler, P., Barreiro, J.T., Monz, T., Nebendahl,
building systems, it seems to be time 5. Childress, L. et al. Coherent dynamics of coupled V., Nigg, D., Chwalla, M., Hennrich, M. and Blatt, R.
electron and nuclear spin qubits in diamond. Science
to begin demanding Moores Law-like 314, 5797 (2006), 281285.
Experimental repetitive quantum error correction.
Science 332, 6033 (2011), 10591061.
improvements in system capacity. Re- 6. Clark, C.R., Metodi, T.S., Gasster, S.D. and Brown, K.R. 33. Shor, P.W. Algorithms for quantum computation:
Resource requirements for fault tolerant quantum
viewers of papers and funding propos- simulation: The ground state of the transverse Ising
Discrete logarithms and factoring. In Proceedings
of the 35th Symposium on Foundations of Computer
als should look for realistic estimates model. Phys. Rev. A 79, 6 (June 2009). Science. IEEE Computer Society Press, Los Alamitos,
7. Devitt, S.J., Fowler, A.G., Stephens, A.M., Greentree, CA, 1994, 124134.
of logical Toffoli gate time, incorporat- A.D., Hollenberg, L.C.L., Munro, W.J. and Nemoto, K. 34. Stace, T.M., Barrett, S.D. and Doherty, A.C. Thresholds
ing error correction, for some target Architectural design for a topological cluster state for topological codes in the presence of loss. Physical
quantum computer. New Journal of Physics 11 (2009). Review Letters 102, 20 (2009).
logical fidelity. Even more ambitiously, 8. Devitt, S.J., Fowler, A.G., Tilma, T., Munro, W.J. and 35. Svore, K.M., Aho, A.V., Cross, A.W., Chuang, I. and
we recommend requiring realistic esti- Nemoto, K. Classical processing requirements for a Markov, I.L. A layered software architecture for
topological quantum computing system. International quantum computing design tools. IEEE Computer
mates of application performance. Journal of Quantum Information 8 (2010), 127. (Jan 2006), 7483.
Developing a sense of community is 9. Devitt, S.J., Nemoto, K. and Munro, W.J. Quantum 36. Van Meter, R., Ladd, T.D., Fowler, A.G. and Yamamoto,
error correction for beginners. Reports on Progress in Y. Distributed quantum computation architecture
critical. Creating a shared understand- Physics 76, 8 (Aug. 2013). using semiconductor nanophotonics. International
10. DiVincenzo, D. The physical implementation of
ing including vocabulary, concepts, and quantum computation. Fortschritte der Physik 48, 9-11
Journal of Quantum Information 8 (2010), 295323.
37. Van Meter III, R.D. Architecture of a Quantum
important problems among the physics (2000), 771783. Multicomputer Optimized for Shors Factoring
11. Fowler, A., Mariantoni, M., Martinis, J. and Cleland,
and CS theory, algorithm design, phys- A. A primer on surface codes: Developing a machine
Algorithm. Ph.D. thesis, Keio University, 2006;
arXiv:quant-ph/0607065.
ics experiment, engineering, and ar- language for a quantum computer. Arxiv preprint 38. Vedral, V., Barenco, A. and Ekert, A. Quantum networks
(2012); arXiv:1208.0928.
chitecture communities has proven to 12. Fowler, A.G., Devitt, S.J. and Hollenberg, L.C.
for elementary arithmetic operations. Phys. Rev. A 54
(1996), 147153; http://arXiv.org/quant-ph/9511018.
be difficult, and few journals or confer- Implementation of Shors algorithm on a linear 39. Wootters, W.K. and Zurek, W.H. A single quantum
nearest neighbor qubit array. Quantum Information
ences currently provide good venues for and Computation 4, 4 (2004), 237.
cannot be cloned. Nature 299, 802 (Oct. 1982).
40. Yao, X.-C. et al. Experimental demonstration of
such interdisciplinary endeavors, but 13. Gay, S. Quantum programming languages: Survey and topological error correction. Nature 482 (Feb. 2012),
bibliography. Bulletin of the European Association for 489494.
we expect the number will grow. Theoretical Computer Science (June 2005).
Let us close with a question that pro- 14. Isailovic, N., Whitney, M., Patel, Y. and Kubiatowicz,
J. Running a quantum circuit at the speed of data. Rodney Van Meter (rdv@sfc.wide.ad.jp) is an associate
vokes answers ranging from, Already International Symposium on Computer Architecture. professor in the Faculty of Environment and Information
has, (in reference to direct quantum IEEE (2008), 177188. Studies at Keio Universitys Shonan Fujisawa Campus,
15. Jiang, L., Taylor, J.M., Srensen, A.S. and Lukin, M.D. Kanagawa Prefecture, Japan.
simulation of a specific quantum sys- Distributed quantum computation based on small
quantum registers. Phys. Rev. A 76 (Dec 2007). Clare Horsman (clare.horsman@gmail.com) was a
tem20) to Twenty years, to Never, 16. Jones, N.C., Van Meter, R., Fowler, A.G., McMahon, project assistant professor at Keio Universitys SFC. She is
and all these from people actually P.L., Kim, J., Ladd, T.D. and Yamamoto, Y. Layered currently a research assistant at the University of Oxfords
architecture for quantum computing. Phys. Rev. 2, 3 Department of CS, Oxford, U.K.
working in the field: (July 2012), 031007.
When will the first paper appear in 17. Kielpinski, D., Monroe, C. and Wineland, D.J. Copyright held by Owners/Author(s)

octob er 2 0 1 3 | vo l. 56 | n o. 1 0 | c om m u n icat ion s of t he ac m 93