The Magazine of IEEE-Eta Kappa Nu July 2017 - 21

The IEEE Rebooting Computing Initiative

a wider variety of computing applications, where they
can be much more energy efficient. In the most recent
LPIRC, the winning competitors used commercial
GPUs for image recognition with both high speed and
low power. New neuromorphic processors may be
particularly efficient for pattern recognition applications,
as discussed below.

IV. Superconducting Computing
Superconductors are metals that have the ability to
carry very large currents without resistance, but only at
cryogenic temperatures. While some superconductors
can operate at temperatures up to about 100 K
(=- 173 °C), most practical superconductors for digital
electronic applications operate at temperatures less than
10 K. Niobium (Nb) is the "silicon of the superconducting
world" and dominates electronic applications; Nb circuits
typically operate at about 4 K. While such temperatures are
impractical for mobile computing, they are quite feasible
for large-scale applications to supercomputers and
data centers. Cryogenic refrigerators (sometimes called
cryocoolers) are reliable and commercially available.
The majority of power for operating a superconducting
computer is actually refrigeration power, so that proper
packaging and input/output are critical.
The basic element in a superconducting circuit is
known as a Josephson junction, and it represents two
superconducting layers with an insulator between
them of thickness on the order of a nanometer (nm).
A superconducting loop with a Josephson junction can
support a lossless circulating current, and acts as an
ideal storage element for magnetic flux. But due to the
quantum nature of the superconducting state, the value
of magnetic flux in the loop must be an integer multiple
of a fundamental quantity called the flux quantum Φ0
= h/2e = 2 x 10-15 Wb. A Josephson junction acts as
a switch for magnetic flux, transferring a single flux
quantum at a time, and generating a voltage pulse with

time-integrated value ∫Vdt = Φ0 = 2 mV-ps. As shown
in Fig. 7, this pulse is typically about 1 mV high and 2
ps wide.
The SFQ pulse is the basis for modern superconducting
logic circuits [6], which are both very fast and very low
power. Clock speeds up to 100 GHz should be feasible,
and the switching energy is ~ 10-18 J/bit, orders of
magnitude smaller than that for CMOS transistors. The
higher speed is responsible for the smaller footprint in
Fig. 6, since fewer parallel processors would be necessary
for the same performance. This also contributes to the
sharply reduced power requirements. However, we
should note that the superconducting results in Fig. 5
are projected; no full-scale superconducting processor
has yet been demonstrated.
FIt is important to appreciate that a single-flux-quantum
circuit is NOT the same as a quantum computer, which
has recently been getting much attention in the news.
The superconducting computer of Fig. 6 is based on
classical bits and logic. However, there are also major
research and development projects (and even one
commercial system) on quantum computing based on
Josephson junctions and superconducting circuits similar
to that in Fig. 7. Unlike a classical bit in Fig. 7, which
might correspond to current circulating either clockwise
or counterclockwise, in a quantum bit ("qubit") one
might have a superposition of circulating currents in
both directions at the same time! See ref. [6] for more
information.

V. Neuromorphic Computing

FSince the earliest computers with vacuum tubes in
the 1940s, computer hardware has been based on
the von Neumann architecture of a digital logic unit
linked to memory and a control program, as shown in
Fig. 8 on the left. Modern computers are much more
complicated, but still have the same basic structure.
People have long spoken of
computers as electronic brains,
but they are really quite different.
Brains have dramatically different
structures from von Neumann
computers, as well as dramatically
different capabilities. The righthand diagram in Fig. 8 illustrates
one type of structure known as
a neural network (or neural net),
which represents a simplified
Fig. 7. Superconducting storage loop and single-flux-quantum voltage pulse. From ref. [6].
version of the interconnections in

THE BRIDGE // Issue 2 2017

Table of Contents for the Digital Edition of The Magazine of IEEE-Eta Kappa Nu July 2017