40 | New Scientist | 25 July 2020
digital technology, which needs currents
to be in one state or another – on or off.
Memristors are therefore seen as an
attractive basis for neuromorphic
computing.
Not everyone is convinced. “I see
some marginal benefits of analogue
circuits for neuromorphic implementations,
but these benefits come with very high
costs that are generally not worth paying,”
says Mike Davies, director of Intel’s
Neuromorphic Computing Lab. “Digital
design methods optimise the key figures
of merit that matter – accuracy, power,
speed and chip cost – in ways that analogue
approaches can’t match.”
Few would argue that digital generally
wins hands down in terms of precision.
What seems to still be open to debate is
how much that matters. “These [analogue]
devices and systems emulate real neural-
processing systems,” says Giacomo Indiveri
at the Swiss Federal Institute of Technology
in Zurich. “As such, they are noisy, imprecise
and unreliable.” Indiveri is an advocate
for combining analogue with digital
technologies to get the best from both.
His view is that what biological neural-
processing systems lack in precision, they
make up for with sophisticated feedback
mechanisms for adaptation, learning and
plasticity. Designed correctly, neuromorphic
devices and algorithms could also benefit
from this trade-off.
That could have huge implications for the
power and speed of computing. The tech
giant IBM’s Blue Gene was one of the most
powerful supercomputers in the world in
2011, yet it still struggled to simulate the
billions of neurons and synapses in the brain
of a cat, all while guzzling through enough
energy to run a thousand domestic homes.
Memristor devices could not only be more
compact, but stand a chance of matching the
brain’s energy efficiency, delivering the same
results on a thousandth of the power. As ever
more objects, from smart fridges to children’s
toys, become equipped with computational
power as part of the much-touted internet of
things, Indiveri sees analogue-embracingThe future of the
internet of things
may depend on
analogue techMARK
O^ GEB
ER
/GETTY^ IMAGEShybrid technologies as a way to power these
devices more efficiently.
He isn’t alone. IBM has teams dedicated to
analogue innovation for AI. One of the firm’s
researchers, Hsinyu Tsai, points out that such
analogue devices are already finding use in
everyday technology, such as the specialised
chips that let you rouse your phone with a
spoken command. “Analogue approaches fit
especially well with today’s AI applications,
where models consist of a large number of
computations, yet require only limited
numerical precision,” says Tsai.
It was two thousand years ago that the
Antikythera mechanism sank to the sea floor.
After all this time, perhaps we are only just
coming to understand the true power of its
analogue legacy. ❚Anna Demming is a freelance
writer based in Bristol, UK“ Hybrid digital-
analogue devices
could not only be
more compact, but
stand a chance of
matching the brain’s
energy efficiency”