White Paper

Data Center
Artificial Intelligence

The AI Readiness Model

Judging an Organization’s Ability to Generate Business Value from
Artificial Intelligence

Table of Contents At Intel, we work with many organizations who are investigating artificial
intelligence (AI) solutions. Image recognition, natural language processing (NLP)
Understanding Where You Are on the and predictive maintenance are emerging as particular hotspots. Some businesses
AI Journey. . . . . . . . . . . . . . . . . . . . . . . . . 1 are exploring these AI use cases for the first time; others are examining how to
Foundational Readiness . . . . . . . . . . . . 2 advance from a successful starting point.
Operational Readiness. . . . . . . . . . . . . . 3 To aid organizations wherever they are on their AI journeys, Intel has created a
Transformational Readiness . . . . . . . . 4 Readiness Model to help decision makers understand where to prioritize efforts.
We have developed this based on our experience working with customers across
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 4
a range of scenarios and industry verticals. Examples include manufacturing
companies wanting to improve quality control, and financial services organizations
looking to use AI in algorithmic trading. This paper provides guidance on how
to judge an organization’s ability and readiness to use AI to generate business
value, and includes a list of questions which you can use to guide your own self-
assessment activities.
To start with, we can group organizations into three categories, according to where
they are in their AI journey – whether they are using AI for the first time, scaling up, or
broadening out use of AI.

Understanding Where You Are on the AI Journey

Some organizations are new to AI
Many organizations are still unfamiliar with deploying AI applications, such as
image recognition, natural language processing (NLP), and predictive maintenance.
We see three common scenarios:
• A
 n organization with existing pools of data can benefit from the use of AI. For
example, a solar energy farm that Intel worked with regularly captured images
of their hardware in an effort to spot damage. In this scenario, the organization
wanted to understand whether image recognition algorithms could be used
to identify flaws automatically. Similar situations can be found in healthcare,
assessing radiology charts or other medical data.
• A
 n organization is running a workload in a traditional environment, and wants
to apply AI or machine learning to explore opportunities for optimization. For
example, these might be enterprises already familiar with analytics deployments,
or research institutions running physics algorithms. Similarly, a manufacturer
using fixed patterns for fault detection in machinery, materials, goods or
processes may recognize the benefit of machine learning.
• S
 ome organizations have been researching the potential of AI – for example, the
use of NLP in customer service scenarios, or retailers using image recognition
to optimize analytics of in-store behaviors. In some cases, the problems that AI
can solve the best may not be obvious – it can, for example, involve a behind-the-
scenes use-case, such as predictive inventory management.
White Paper | The AI Readiness Model 2

For these organizations, it can be a challenge to map out the resources as requirements and resource demands fluctuate
benefits of AI in advance and data is not always available or between each case.
presented in such a way that is suitable for AI.
The Three Types of AI Readiness
Other organizations are ready to scale up AI
While organizations can be at various stages of their AI
These organizations have researched the possibilities journeys, their progress to the next stage or to ongoing
for AI, have an idea of where they want to use it and have success is dependent on having the right elements in place
successfully implemented test models. They may fall into across skills and resources, infrastructure and technology,
a number of scenarios: processes and models.
• An organization may have developed a proof of concept We can also consider readiness in terms of the following areas:
AI solution running on a workstation or a single device.
• F
 oundational – a prerequisite for AI is appropriate
The challenge is to migrate the solution to a data center
infrastructure and interfaces
environment so that it can be moved to production. One
example would be enabling operations to work with • O
 perational – suitable management and governance
predictive maintenance insights. mechanisms are key to the sustainability of AI solutions
• An organization has developed a ‘home-grown’ • T
 ransformational – the ability of an organization to maximize
solution, and is now looking to use industry standard the value it gets from AI
infrastructure and/or software. Whether hardware,
Foundational readiness is the first step, but the success of AI
software or both, the challenge can come from migrating
hinges on operational readiness, and then how receptive the
to an architecture which (in the shorter term) yields less
business is to AI – transformational readiness. This feeds new
optimal results.
and updated requirements, which further drive the foundation
While these organizations are more advanced in their AI for AI deployments.
journey, they can still lack the skills they need to ‘scale up’
their use of AI beyond a smaller number of engineers with Foundational Readiness
a relatively simple hardware and software configuration.
A prerequisite for AI is appropriate infrastructure and
In addition, moving to a multi-node solution will need to
interfaces. While this stands to reason, it is not always evident
address the fact that AI does not scale linearly. For example,
what the needs will be in advance of testing and evaluating
whereas a single-node configuration can process hundreds of
potential scenarios. At the same time, skills and expertise may
images per second, a move to 50 processors will not deliver
be in short supply. With these factors in mind, the following
50 times the performance. Data sources may also bottleneck
should be taken into account.
if required to be used ‘at scale’ versus one-off sampling.
Infrastructure platform
Others are broadly implementing AI
Many organizations want to understand whether existing
A third category of organizations are using machine learning
data center facilities will be suitable for AI workloads. While
or AI to some extent, and are now looking to broaden their
the answer may be yes for a proof of concept, some facilities
adoption across a wider range of use cases. We see the
may not suit the massively scalable processing required for
following scenarios:
machine learning and AI. Depending on the scenario, the data
• An organization may be using AI successfully in a flow required by AI can severely tax your network bandwidth.
line of business, and is now looking to expand. For Some AI solutions can be scheduled at off hours to maximize
example, a company might be using image recognition in throughput. Meanwhile other, more time-critical scenarios may
manufacturing for quality control, and now wants to deploy impose a greater burden on the network, for example if the
NLP in call centers. data was linked into a predictive maintenance capability.
• An organization is successfully using AI to learn from and For further information about platform requirements for AI,
interpret data, and now wants to extend into inference- across processing, storage and networking, see our paper on
based maintenance and updates to models. They may creating a proof of concept.
also want to use outcomes of AI to drive automation,
for example, using inventory data to drive spare-part Cloud resources
inventory management, planning and acquisition. From an Consider cloud-based services as a basis for AI, particularly
infrastructure point of view, the same organization might in areas such as image and natural language processing. The
also be looking to improve power efficiency or performance cloud offers the advantage of low-entry-point, pay-per-use
and reduce total cost of ownership (TCO). services, making it an excellent choice for training and testing.
In these scenarios, challenges come from ensuring that the As organizations look to scale up use of AI, they will need to
platform can fit more than one use case, and managing check if their cloud-based resources are still suitable. Cloud
White Paper | The AI Readiness Model 3

service providers may have graded pricing and service level

agreements, perhaps defined by the underlying compute
infrastructure running the workload. Decision makers will need AI software optimizations for Intel®
to strike the optimum balance between performance and cost Xeon® processors
– possibly on a workload by workload basis.
To allow data scientists and developers to work with
Data sources their preferred framework, Intel has optimized a
number of deep learning libraries for many of the
Critical data sources need to be available and accessible.
most popular AI frameworks including, TensorFlow*,
Decision makers and engineers will require an assurance that
Theano*, and MXNet*.
data is available in the right quantity and quality for deep
learning algorithms, whether it is internally or externally Operating underneath these frameworks, the Intel®
sourced: in the latter case, this may need contractual Math Kernel Library for Deep Neural Networks (Intel®
arrangements to be agreed with third parties. MKL-DNN) is a new accelerator with specific math for
deep learning, optimized on top of x86 with Intel®
Software packages Advanced Vector Extensions 2 (Intel® AVX-2) and
A wide variety of inference, machine learning and AI software Intel® Advanced Vector Extensions 512 (Intel® AVX-
packages are available today, each with benefits and 512) instructions. As an open source project, it will
constraints. Open source packages such as, TensorFlow* continue to track new and emerging trends in all major
and BigDL sit alongside commercial solutions such as the frameworks.
Intel® Saffron™ software and cloud-based services. Each of Alternatively, BigDL is an open-source distributed
these needs to integrate with tools for data management, deep learning library for Spark* that can run directly
visualization and so on: each use case will require a software on top of existing Spark or Apache Hadoop* clusters.
architecture that selects the best tools for the job in hand, Developed by Intel, it allows for loading of pre-trained
potentially taking into account downstream systems, Torch* models into the Spark framework, and can
customization, optimization and other modifications. efficiently scale out to perform data analytics at big
data scale.
Operational Readiness
Learn more about both of these solutions in our
Suitable management and governance mechanisms are key infographic The Anatomy of an AI Proof of Concept.
to the sustainability of AI solutions. IT decision makers can
examine several areas to ensure their AI deployments will be
operationally ready. We recommend a review of the following.

Agile delivery architects who can tune proof-of-concept solutions; however,

Best practice models such as DevOps and other agile it is beneficial to build up in-house skills across both IT
methodologies can significantly help organizations at any and lines of business. Proceeding beyond proof of concept
stage of their AI journey – in earlier stages, continuous without in-house skills adds unnecessary risk to an AI project,
development and delivery responds to quickly changing particularly in the review and evaluation stages.
requirements, unclear outcomes and the need for repeated
review, evaluation and testing. And for more advanced use, Cybersecurity
continuous improvement can drive changes in both data
Given how the relationship between AI and automation
sources and models.
can minimize human intervention and therefore oversight,
cybersecurity of data, infrastructure and algorithms should
Operational management
be seen as a high priority. Potential security risks come from
The ‘Ops’ side of DevOps also needs to incorporate criteria the corruption of input data into the AI, tampering with the
around effective service delivery (across the platform selected models, or unauthorized access to the resulting insights.
as a foundation for AI), as well as effective management of all
data sources, both internal and external. Additionally, criteria Governance, compliance and risk
should exist around measuring the business effectiveness of
The relationship between AI and governance is complex and
AI – for example, are the generated insights or automation
many-faceted. In earlier stages of an organization’s journey,
delivering the expected business value?
governance questions will not be any different to other
data-centric IT projects – can the project deliver, is customer
Skills and expertise
privacy protected and so on. As its use is extended, AI brings
Lack of skills is a frequent challenge for organizations in additional ramifications: for example, in predictive planning
the earlier stages of the AI journey. At the outset, you may and maintenance, how much human involvement is required
need to bring in skills from outside, in particular solution in spares purchasing (if any)?
White Paper | The AI Readiness Model 4

Business opportunity
From data to insight to innovation In the same vein, AI’s chances of success can be improved
to revenue: Intel® Xeon® Scalable if it unlocks new opportunities for business growth, new
processors ways to engage with customers or new types of operational
process. If this is the case, the organization should be
Intel® Xeon® processors are the heart of the data looking to structure itself in order to leverage the AI-driven
center, running the majority of the most critical and business opportunities that may exist. It is important to be
most innovative workloads. While existing Intel® clear on how existing and desired operating models can
architecture in your data center will allow you to get integrate the results and benefits of AI, through automation
started with machine learning and deep learning, or augmentation.
the new Intel® Xeon® Scalable processors offer the
most agile Intel platform for AI, capable of taking Clarity of business case
your AI to the next level. In terms of the compute A precursor to any business change is to have a clear picture
capability required for deep learning training of the benefits that change will bring. In Intel’s experience,
and inference, the Intel Xeon Scalable processor organizations earlier on in their AI journey have more focus
delivers up to 2.2x performance compared to the on TCO: whether AI delivers the expected results (and saves
previous generation for deep learning training money through automation) at an acceptable cost. More
and inference performance. And, benefitting from developed projects are looking to increase AI performance,
additional software optimizations – for example, and the most advanced are looking to see ROI in business
TensorFlow*, Theano* and Torch* – can achieve up terms, for example the amount of time that is released to
to 113x performance compared to non-optimized perform other tasks. The business case should present clear,
three-year-old servers for deep learning, providing costed criteria for what constitutes success.
a solid foundational architecture for AI workloads1.
Six memory channels compared to previous Business acceptance
generations’ four makes for significantly increased The solution should be adapted to business needs right
memory bandwidth and capacity for memory- through to the daily activities of front-line staff, and the
intensive workloads. people impacted. Achieving acceptance may not always be
straightforward, particularly if job roles and responsibilities
change as a result of implementing AI.

Conclusion: Start Planning Now for AI

Success
Transformational Readiness Whether your organization is at the beginning of its AI
The third (and possibly most important) area of readiness is journey, or has already made some progress, you can ask
whether an organization can maximize the value it gets from yourself a number of questions to assess where you are
AI. Whether AI is being used for pattern recognition, insight and increase your readiness as a result. Some suggested
generation or process automation, it will generally have one questions that may assist your planning are below:
of the following impacts:
• Support better decision making, at senior or line Using AI for the first time
management level • I s the scenario, use case or problem to be solved with AI
clearly defined?
• Automate a part of a business process, or drive an
automated response • A
 re priorities set around where AI will deliver the most
business value?
In either case, AI can have a major impact on the day-to-day
running of the business – which means that the business • I s the planned infrastructure architecture clear and
needs to be able to embrace resulting changes. appropriate?
• A
 re all necessary data sources clearly understood and
Strategic leadership accessible?
It is not essential that the organization’s board sees digital • C
 an your chosen software packages deliver the AI solution
technologies in general, and AI in particular, as a driver for end-to-end?
business growth. However, it certainly helps if the right
mindsets exist, to drive delivery from the very top of the • A
 re sufficient skills and resources available (either in-house
or externally)?
organization. If AI is seen as a source of strategic advantage,
priorities and budgets will be set accordingly. • H
 ave expectations been set around training and learning
times?
White Paper | The AI Readiness Model 5

• A
 re measures in place to monitor business effectiveness
Making the most of your existing Intel® of AI solutions?
Xeon® processor-based infrastructure • I s the architecture for AI provided as a platform, rather than
Leveraging your existing data center infrastructure is as one-off solutions?
an ideal opportunity to prove the value of AI to your • A
 re lines of business fully engaged in how AI will affect their
business from a flexible, general purpose foundation. processes?
Intel® Xeon® processors use a consistent • A
 re the governance needs of the AI solution clearly
infrastructure and programming model for existing understood?
analytics pipelines and support the large memory
requirements of AI models. The previously mentioned • I s AI seen as a central pillar of an IT-enabled business
software optimizations mean Intel Xeon processor- strategy?
based infrastructure can continue to support an By addressing these questions, you will increase the
organization’s AI journey from experimentation to probability of success – and will increase the acceptability of
proof-of-concept to production, depending on AI-based solutions in your organization. Getting everything
the workload. right from the start is not a realistic expectation: rather, you
Learn more about both of these solutions in our should be looking to build skills and expertise as you discover
infographic The Anatomy of an AI Proof of Concept. the benefits of AI.
Most important is to have a clear grasp of the problem you
are looking to solve. Establish a problem statement, then
work to make the solution a reality. Whether it is automating
• Is the TCO of the end-to-end solution clear and signed off? processes or delivering insight, the ultimate goal of AI is to
give yourself, and your organization, freedom to innovate
Scaling up use of AI and grow.
• Can the planned solution scale beyond initial testing and
evaluation?
• Is a clearly defined business case confirmed with a business
unit? Learn More
• Is sufficient direct resourcing available, with time allocated To read more about the Intel AI portfolio and how it
and reserved? can support your journey to AI, visit: ai.intel.com

• Is network bandwidth sufficient to ensure timely data Intel’s performance-optimized machine and deep
delivery at scale? learning libraries and frameworks are available here:
https://software.intel.com/en-us/ai-academy
• Are operational management processes in place which
cover AI delivery?
• Does the architecture align with industry standards and
best practices?
• Has a cybersecurity risk assessment been undertaken and
acted upon?
• Have realistic deployment plans been set and
communicated?

Broadening out use of AI

• Is a team in place to oversee AI-based continuous
improvement?
• Are the broader AI opportunities for the organization
researched and clear?
• Are AI solutions developed and deployed following agile
best practice?
White Paper | The AI Readiness Model 6

1
INFERENCE using FP32 Batch Size Caffe GoogleNet v1 256 AlexNet 256.
Performance estimates were obtained prior to implementation of recent software patches and firmware updates intended to address
exploits referred to as “Spectre” and “Meltdown.” Implementation of these updates may make these results inapplicable to your device
or system. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.
Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations
and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance
tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other
products. For more complete information visit http://www.intel.com/performance Source: Intel measured as of June 2017 Optimization
Notice: Intel’s compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to
Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee
the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent
optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture
are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the
specific instruction sets covered by this notice.
Configurations for Inference throughput
Processor :2 socket Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz / 28 cores HT ON , Turbo ON Total Memory 376.46GB (12slots
/ 32 GB / 2666 MHz).CentOS Linux-7.3.1611-Core , SSD sda RS3WC080 HDD 744.1GB,sdb RS3WC080 HDD 1.5TB,sdc RS3WC080
HDD 5.5TB , Deep Learning Framework caffe version: f6d01efbe93f70726ea3796a4b89c612365a6341 Topology :googlenet_v1 BI
OS:SE5C620.86B.00.01.0004.071220170215 MKLDNN: version: ae00102be506ed0fe2099c6557df2aa88ad57ec1 NoDataLayer.
Measured: 1190 imgs/sec vs Platform: 2S Intel® Xeon® CPU E5-2699 v3 @ 2.30GHz (18 cores), HT enabled, turbo disabled, scaling
governor set to “performance” via intel_pstate driver, 256GB DDR4-2133 ECC RAM. CentOS Linux release 7.3.1611 (Core), Linux
kernel 3.10.0-514.el7.x86_64. OS drive: Seagate* Enterprise ST2000NX0253 2 TB 2.5” Internal Hard Drive. Performance measured
with: Environment variables: KMP_AFFINITY=’granularity=fine, compact,1,0‘, OMP_NUM_THREADS=36, CPU Freq set with cpupower
frequency-set -d 2.3G -u 2.3G -g performance. Deep Learning Frameworks: Intel Caffe: (http://github.com/intel/caffe/), revision
b0ef3236528a2c7d2988f249d347d5fdae831236. Inference measured with “caffe time --forward_only” command, training measured
with “caffe time” command. For “ConvNet” topologies, dummy dataset was used. For other topologies, data was stored on local storage
and cached in memory before training. Topology specs from https://github.com/intel/caffe/tree/master/models/intel_optimized_models
(GoogLeNet, AlexNet, and ResNet-50), https://github.com/intel/caffe/tree/master/models/default_vgg_19 (VGG-19), and https://github.
com/soumith/convnet-benchmarks/tree/master/caffe/imagenet_winners (ConvNet benchmarks; files were updated to use newer
Caffe prototxt format but are functionally equivalent). GCC 4.8.5, MKLML version 2017.0.2.20170110. BVLC-Caffe: https://github.com/
BVLC/caffe, Inference & Training measured with “caffe time” command. For “ConvNet” topologies, dummy dataset was used. For other
topologies, data was st ored on local storage and cached in memory before training BVLC Caffe (http://github.com/BVLC/caffe), revision
91b09280f5233cafc62954c98ce8bc4c204e7475 (commit date 5/14/2017). BLAS: atlas ver. 3.10.1.
Configuration for training throughput:
Processor :2 socket Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz / 28 cores HT ON , Turbo ON Total Memory 376.28GB (12slots
/ 32 GB / 2666 MHz).CentOS Linux-7.3.1611-Core , SSD sda RS3WC080 HDD 744.1GB,sdb RS3WC080 HDD 1.5TB,sdc RS3WC080
HDD 5.5TB , Deep Learning Framework caffe version: f6d01efbe93f70726ea3796a4b89c612365a6341 Topology :alexnet BIO
S:SE5C620.86B.00.01.0009.101920170742 MKLDNN: version: ae00102be506ed0fe2099c6557df2aa88ad57ec1 NoDataLayer.
Measured: 1023 imgs/sec vs Platform: 2S Intel® Xeon® CPU E5-2699 v3 @ 2.30GHz (18 cores), HT enabled, turbo disabled, scaling
governor set to “performance” via intel_pstate driver, 256GB DDR4-2133 ECC RAM. CentOS Linux release 7.3.1611 (Core), Linux
kernel 3.10.0-514.el7.x86_64. OS drive: Seagate* Enterprise ST2000NX0253 2 TB 2.5” Internal Hard Drive. Performance measured
with: Environment variables: KMP_AFFINITY=’granularity=fine, compact,1,0‘, OMP_NUM_THREADS=36, CPU Freq set with cpupower
frequency-set -d 2.3G -u 2.3G -g performance. Deep Learning Frameworks: Intel Caffe: (http://github.com/intel/caffe/), revision
b0ef3236528a2c7d2988f249d347d5fdae831236. Inference measured with “caffe time --forward_only” command, training measured
with “caffe time” command. For “ConvNet” topologies, dummy dataset was used. For other topologies, data was stored on local storage
and cached in memory before training. Topology specs from https://github.com/intel/caffe/tree/master/models/intel_optimized_models
(GoogLeNet, AlexNet, and ResNet-50), https://github.com/intel/caffe/tree/master/models/default_vgg_19 (VGG-19), and https://github.
com/soumith/convnet-benchmarks/tree/master/caffe/imagenet_winners (ConvNet benchmarks; files were updated to use newer
Caffe prototxt format but are functionally equivalent). GCC 4.8.5, MKLML version 2017.0.2.20170110. BVLC-Caffe: https://github.com/
BVLC/caffe, Inference & Training measured with “caffe time” command. For “ConvNet” topologies, dummy dataset was used. For other
topologies, data was st ored on local storage and cached in memory before training BVLC Caffe (http://github.com/BVLC/caffe), revision
91b09280f5233cafc62954c98ce8bc4c204e7475 (commit date 5/14/2017). BLAS: atlas ver. 3.10.1.
Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service
activation. Performance varies depending on system configuration. No computer system can be absolutely secure. Check with your system
manufacturer or retailer or learn more at intel.com
Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance
tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions.
Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you
in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more
complete information visit www.intel.com/benchmarks
Benchmark results were obtained prior to implementation of recent software patches and firmware updates intended to address exploits
referred to as “Spectre” and “Meltdown”. Implementation of these updates may make these results inapplicable to your device or system.
All information provided here is subject to change without notice. Contact your Intel representative to obtain the latest Intel product
specifications and roadmaps.
Intel, Xeon, Saffron, and the Intel logo, are trademarks of Intel Corporation in the U.S. and/or other countries.
* Other names and brands may be claimed as the property of others.
© Intel Corporation 0318/CAT/RD/PDF 337350-001EN