MICROPROCESSORS ARTIFICIAL INTELLIGENCE
problematic. You can end up creating
a diverse, but confusing market, with
thousands of companies competing
against one another and, to date, few
are at the point of delivering silicon.”
According to Grant companies are
targeting data centres while a growing
number have gone after the IoT.
“The AIoT is a very fragmented
space. In some cases you can run
basic AI tasks on a CPU, or DSP,
or use dedicated neural network
processors,” Grant explains. “It’s a
dynamic and fast changing market and
as a consequence we are seeing neural
networks and traditional computing
coming together – in a sense playing to
the strengths of both.”
Last year, Xmos unveiled xcore.ai,
what it described as an inexpensive
crossover processor that was
designed specifically for the AIoT
market delivering high-performance AI,
DSP, control and IO in a single device.
According to Xmos CEO, Mark
Lippett, “This type of capability
would traditionally have been
deployed either through a powerful
applications processor or a
microcontroller with additional
components to accelerate key
capabilities. With this crossover
processor we have been able to
architect it to deliver real-time
inferencing and decisioning at the
edge, as well as signal processing,
control and communications.”
A major issue for designers is
power consumption, much of which
is consumed as data is moved to
and from memory, which in turn
raises the question as to how much
memory is needed?
Companies looking to reduce the
power required to move data across
the chip are implementing in-memory
architectures that will either move
the compute closer to the memory
or the compute into the array itself
using memory cells to perform
certain computations.
As an alternative, IBM is looking
at processing in the analogue rather
than the digital domain due to the
massive improvement in energy
efficiency it promises to deliver.
With so much innovation,
however, Grant makes the point
that for many companies embracing
new technologies is simply too
challenging, while relying on more
traditional approaches is simply less
risky and certainly less costly.
“Many companies are happy
developing AI chips that use
conventional architectures, especially
when so much has already been
invested in their development. It’s also
a risk for established chipmakers who
may well be wedded to a particular
programming paradigm,” explains
Grant.
For companies in this space
combining traditional computing
architectures with hardware and
software acceleration schemes would
seem to be a sensible solution,
but there does come a point where
starting from scratch can result in a
much better solution.
From established market leaders
like Nvidia, Intel and Qualcomm to
start-ups like BrainChip, Graphcore
and Xmos new devices have started
appearing in short order, all of them
looking to stake a claim in this
emerging market.
Many smaller companies have
focused on specific AI applications and
many unique designs and methods for
AI applications have emerged, from
microprocessor build, to materials and
devices, circuits and architectures.
As chip physics gets smaller and
smaller, so conventional processor
architectures will become less
efficient, and developers will have to
turn to more innovative architectures.
“There is still a role for traditional
CPUs and GPUs,” says Grant.
“Especially in data centres, but where
you need to conduct a particular task
new architectures are going to be
critical in delivering lower costs, lower
power and greater efficiency.
When it comes to machine learning
workloads, for example, they’re
completely new and will require new
structures that will not necessarily be
supported by existing CPUs and GPUs.
“That would suggest that a new kind of
processor architecture will be required,
but I believe there is still a role for
both going forward,” says Grant.
“Today’s CPUs
and GPUs are
not sufficiently
capable of
delivering the
workloads that
are needed, so
there has to be
a better way of
doing this.”
Andrew Grant
“With this
crossover
processor we
have been able
to architect it to
deliver real-time
inferencing and
decisioning at the
edge, as well as
signal processing,
control and
communications.”
Mark Lippett
www.newelectronics.co.uk 9 February 2021 19
/www.newelectronics.co.uk