ARCHITECTURES FOR AI
How are microprocessors architectures evolving to address the needs
of artificial intelligence? By Neil Tyler
D evelopers of AI systems
have been constrained,
to some extent, by the
limitations associated with standard
microprocessors and while companies
have been unveiling a host of new
platforms, questions have been
raised as to whether the CPUs and
GPUs coming to market are capable
of meeting the increased workloads
required. As a result, do we need new
architectures that are engineered
specifically for AI?
AI applications require huge
amounts of compute resources if they
are to produce the desired outputs
and that has resulted in the use of a
large numbers of multi-purpose central
processing units (CPUs) working in
parallel, however, none of these are
actually optimised to address the
processing functionality that AI needs
if it is to perform optimally.
An alternative has been the
use of GPUs which have, for many
companies, become a ‘de facto’
AI co-processor accelerator. Unlike
CPU’s they have simpler cores that
can provide dedicated VRAM memory.
As a consequence, they are able to
better handle statistical computation
and the parallel processing needs
that are required by machine learning
applications.
Like CPUs, GPUs can deliver AI
processing capabilities, but again they
have not been designed specifically
for AI.
So how is the microprocessor
industry looking to address the
requirements of the ‘AI chip’ with
the necessary compute capabilities
required?
“AI has been around since the
1950s but it’s only today that we’re in
the situation where all the elements
are now, finally, coming together,”
that GPUs will lose out to AI-specific
processors and that opportunities
for AI chips will emerge at both data
centres and the edge.
“AI systems developers are
looking for processors that have been
designed and optimised specifically
for AI jobs. That tends to involve
many-cored processors with built-in
parallelism, that can perform analysis
in real-time,” explains Grant.
AI chips tend to focus on lowprecision
arithmetic, novel dataflow
architectures, or in-memory computing
capabilities and have a closely
networked architecture that enables
processors to transfer data between
them keeping energy consumption to
a minimum.
“There’s certainly room for new
players in this space,” according to
Grant.
He believes that the level of
investment and R&D being committed
to AI chip development actually
presents a rare opportunity for
start-ups, who are going up against
established market leaders.
“With new companies we benefit
from more competition but that can be
says Andrew Grant, Senior Director,
Artificial Intelligence, Imagination.
“There’s a huge amount of work going
on in this space, and there has been a
major shift in the compute capabilities
that enable us to deliver AI.”
The industry is certainly seeing
more AI-driven demand for GPUs
but the requirements of AI are now
being taken into account at the very
beginning of the design process.
“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,”
says Grant. “What we are seeing is a
move to new architectures because
it’s not necessarily good to deploy hot
and power hungry CPUs or GPUs, at
scale, and as a consequence we are
seeing a growing number of start-ups
working in this space developing AI
specific architectures.”
A good example of this is
Bristol-based Graphcore, which has
developed the Intelligence Processing
Unit. A new type of microprocessor,
it has been specifically designed
to support artificial intelligence
workloads.
According to Graphcore CEO
and co-founder Nigel Toon,”It’s
a technology that dramatically
outperforms legacy processors such
as GPUs and with a powerful set of
software tools has been tailored to
the needs of AI developers. When it
comes to AI a new kind of processor
architecture is required.’
AI related processors
Research carried out for McKinsey
suggests that by 2025 AI-related
processors could account for almost
20 per cent of all microprocessor
demand, and could be worth in excess
of £50bn. It goes on to suggest
knssr/stock.adobe.com
18 9 February 2021 www.newelectronics.co.uk
/stock.adobe.com
/www.newelectronics.co.uk
/stock.adobe.com