In a recent survey, International Data Corporation (IDC)
found that accelerated computing is quickly gaining traction in the
enterprise as businesses embrace these technologies to overcome the
limitations of CPUs. To help organizations to better understand where
accelerated computing fits in the computing platforms hierarchy and to
develop a more informed implementation strategy, IDC has published its
first accelerated compute taxonomy.
Accelerated
computing is the ability to accelerate applications and workloads by
offloading a portion of the processing onto adjacent silicon subsystems
such as graphics processing units (GPUs) and field programmable gate
arrays (FPGAs). Accelerated computing is gaining traction in the
enterprise as businesses seek solutions for overcoming the limitations
of central processing units (CPUs) for workloads that require data
processing acceleration. Such apps/workloads rely on (single stream)
pipeline processing where data locality is important for quick handoffs
between various components.
Accelerated
computing is used for unstructured data management workloads, including
cognitive, deep learning, artificial intelligence, machine learning,
and similar types of applications; data analytics workloads, including
visual analytics; technical and scientific workloads; cloud computing,
both internally and in the form of acceleration as a service; and edge
computing. In addition, accelerated computing will affect most workloads
as defined by IDC.
"Compute
will become a lot less homogeneous as today's acceleration
technologies, like GPUs and FPGAs, and yet-to-be-developed accelerators
start transforming server infrastructure to meet the performance demands
of modern workloads, including cognitive and AI," said Peter Rutten, research manager, Servers and Computing Platforms at IDC.
When
asked which specific infrastructure attributes are important for
deploying mission-critical workloads in their organization, nearly three
quarters of the survey respondents identified single or multiple GPUs.
GPUs are especially attractive to businesses as they can be procured off
the shelf and utilize standard libraries that can be incorporated into
applications easily. However, other technologies that offer potentially
higher performance per watt such as FPGAs, many-core processors, and
application specific integrated circuits (ASICs) are starting to gain
traction as well.
"There's
a new world of possibilities with accelerators, which all have unique
technical characteristics and capabilities allowing the right
accelerator to be matched with the right workload," said Gregoire
Robinson, research analyst with IDC's Servers and Computing Platforms
team.
The new taxonomy, IDC's Worldwide Accelerated Compute Taxonomy, 2017 (Doc
#US42878517), provides an overview of key enterprise-focused
accelerated computing hardware definitions. These definitions delineate
the scope of IDC's accelerated compute research, which is tightly
connected to IDC's servers and compute platforms research as well as
IDC's 3rd Platform coverage. IDC sees an increasing emphasis within
enterprises on big data processing, data analytics, and
cognitive/artificial intelligence (AI) applications that are challenging
the capabilities of general-purpose processors in servers deployed in
the datacenter, in the cloud, and at the edge.