EDITOR’S QUESTION
KEVIN DEIERLING,
MELLANOX’S VP OF
MARKETING
U
ntil recently, the IT industry could rely on
Moore’s Law to deliver not only more powerful
computers regularly but also falling power
requirements – that as a new chip could have
twice the transistors as the previous generation
but consume the same amount of power.
Now that Moore’s Law is dead, data centres must develop
smarter and more efficient ways to run workloads, including
the use of AI-aided design, domain specific processors, smart
networks, containerisation and Edge Computing. The traditional
compute method was to run all tasks in software and all
software on general-purpose CPUs, such as the Intel and AMD
X86. Software developers would simply add functions and
sophistication to applications (or databases, storage, etc.)
without worrying about optimising efficiency.
Today, software must add new features and become more
efficient simultaneously. Having humans go through source code
line-by-line to optimise it is tedious and slow, as is switching
from high-level languages like Perl to lower-level languages like
C/C++ to increase efficiency. But new Artificial Intelligence (AI)
tools can optimise software code automatically, quickly reducing
data centre power consumption. AI also can improve the
efficiency of data centre power distribution, airflow and cooling,
plus reduce the time needed to move applications to domainspecific
processors such as GPUs, FPGAs, or SmartNICs, which
each can run specific tasks much more efficiently than an X86
CPU. That in turn highlights the benefits of smarter networks
that connect the distributed processing power of multiple CPUs,
GPUs and FPGAs, and also process data in the network itself
using SmartNICs and intelligent switches. For example, many
security, storage, virtualisation and networking tasks run 2X, 3X,
or even 6X faster (and consume proportionally less power) on a
properly designed SmartNIC like the Mellanox ConnectX-5 and
ConnectX-6 than on an X86 CPU.
Another way to make compute greener is to switch from
virtualisation to containerisation, because containers don’t run on
top of a hypervisor and don’t duplicate the entire machine/OS/
kernel stack for each application instance. However, containers
can be less secure than VMs, requiring additional security in the
form of next-generation firewalls and security isolation for bare
metal servers. These additional security requirements would
require more compute power, adding a new challenge to greening
IT operations, but again can be offloaded efficiently to SmartNICs
and I/O Processing Units (IPUs) – such as the Mellanox BlueField
– to keep data centres moving towards green.
Finally, Edge Computing improves efficiency by moving data
collection, processing and analysis tasks closer to devices and
users that both generate and consume the data. It’s greener to
analyse and react to call data in the base station, retail data in
the store and video in the camera than it is to ship all the data to
a central data centre, process it and send the resulting actions or
recommendations back to the edge of the network.
ANOTHER WAY TO MAKE
COMPUTE GREENER
IS TO SWITCH FROM
VIRTUALISATION TO
CONTAINERISATION.
30 Issue 15
www.intelligentdatacentres.com