EXPERT OPINION
specific infrastructure decisions can vary
based on industry. Legal or compliance
requirements, such as GDPR, as well as the
type of data and work processes involved,
all factor into AI infrastructure decisions.
The study found that 39% of companies
across industries use major public clouds
– most often these were manufacturers
looking for flexibility and high-speed.
Meanwhile, 29% of respondents prefer
in-house solutions with support from
consultants – most often financial,
energy and healthcare companies that
wish to keep their personally identifiable
information (PII) data under tight security
and greater control.
Elements of successful
AI infrastructure
With so many companies starting from
ground zero, it’s imperative to nail-down
a clear strategy from the start, since
rearchitecting later can cost a lot of time,
money and resources. There are several
boxes companies need to check to
successfully enable AI at scale.
First, businesses need to be able to
ensure they have the right infrastructure
to support the data acquisition and
collection necessary to prepare
datasets used for AI workloads. In
particular, attention must be given to
the effectiveness and cost of collecting
data from Edge or cloud devices where
AI inference runs. Ideally this needs
to happen across multiple worldwide
regions, as well as leveraging highspeed
connectivity and ensuring high
availability. This means businesses need
infrastructure supported by a network
fabric that can offer the following benefits:
• Proximity to AI data: 5G and fixed
line core nodes in enterprise data
centres bring AI data from devices in
the field, offices and manufacturing
facilities into regional interconnected
data centres for processing along a
multi-node architecture.
• Direct cloud access: Provides
high performant access to a cloud
hyperscale environment to support
hybrid deployments of AI training or
inference workloads.
• Geographic scale: By placing
their infrastructure in multiple
data centres located in strategic
geographic regions, businesses
enable cost-effective acquisition of
data and high-performance delivery
of AI workloads worldwide.
As businesses consider training AI/
Deep Learning models, they must
consider a data centre partner that
will in the long-term be able to
accommodate the necessary power and
cooling technologies supporting GPU
accelerated compute and this entails:
• High rack density: To support AI
workloads, enterprises will need to
get more computing power out of
each rack in their data centre. That
means much higher power density. In
fact, most enterprises would need to
scale their maximum density at least
three times to support AI workloads –
and prepare for even higher levels in
the future.
• Size and scale: Key to leveraging
the benefits of AI is doing it at scale.
The ability to run at scale hardware
(GPU) enables the effect of largescale
computation.
A realistic path to AI
Most on-premises enterprise data centres
aren’t capable of handling that level of
scale. Public cloud, meanwhile, offers the
path of least resistance, but it isn’t always
the best environment to train AI models at
scale or deploy them in production due to
either high costs or latency issues.
So, what’s the best way forward for
companies that want to design an
infrastructure to support AI workloads?
Important lessons can be learned by
examining how businesses that are
already gaining value from AI have
chosen to deploy their infrastructure.
Hyperscalers like Google, Amazon,
Facebook and Microsoft successfully
deploy AI at scale with their own core
and Edge infrastructure often deployed
in highly connected, high-quality data
centres. They use colocation heavily
around the globe because they know it
can support the scale, high-density and
connectivity they need.
By leveraging the knowledge and
experience of these AI leaders,
enterprises will be able to chart their own
destiny when it comes to AI. ◊
PUBLIC CLOUD
. . . ISN’T ALWAYS
THE BEST
ENVIRONMENT TO
TRAIN AI MODELS
AT SCALE OR
DEPLOY THEM
IN PRODUCTION
DUE TO EITHER
HIGH COSTS OR
LATENCY ISSUES.
42 Issue 19
www.intelligentdatacentres.com