Overview
AI is one of the most promising technologies in the world, but the field is plagued by high failure rates and necessarily long development cycles to move from pilot to production. This article examines AI in the enterprise space: challenges faced in AI implementation, benchmarked solutions as generated from the Intel and Aible “30-day to value” program, and what this means for business owners looking to invest in AI technology. We will also detail how Intel AI software and hardware technologies are central to Aible's approach in delivering impact.
Challenges with AI
According to Gartner*, 85% of AI and machine learning projects fail to deliver, and only 53% of projects make it from prototypes to production. The deeper issue, however, as uncovered by the Massachusetts Institute of Technology, Boston Consulting Group (MIT-BCG), is that "a mere 10% of organizations achieve significant financial benefits with AI." Even when AI projects are found to be successful, the technology has the potential to be unsuccessful in unexpected ways and produce unexpected results. Perhaps more significantly, the same corporations that claim to use AI are also using human biases to shape the outcomes of their algorithms; it is creating the illusion that AI makes decisions without bias when in reality, humans are still driving the decisions.
Three challenges faced by data scientists and AI machine learning developers as per Aible customer case studies are:
- No one has perfectly clean data.
- The business user role is integral.
- AI needs to be tied to business objectives and needs.
AI in 30 Days
The Aible Impact
Reducing the risk and time-to-value of AI projects is imperative for the success of this ecosystem, and Aible is one of such promising startups pledged to create value from AI within 30 days. The company is working with enterprise AI companies to transform AI's "art of possibilities" into practical applications for business—it's an example of how AI is already creating value for companies and it will only become more important in the years ahead. These are ambitious goals, but they are made possible when organizations partner with companies that can deliver on such promises.
Aible is part of the Intel® Disruptor Initiative, a global program that identifies and works with the most exciting companies building on AI. The program brings together the world’s best AI talent with the most promising start-ups to identify and support the most promising companies building on AI today. The goal of this initiative is to shift the conversation from, "Will we implement AI in our organization?" to "How fast can we get value from AI?"
Aible answers the latter by having a portfolio of products that targets the end-to-end automated machine learning (AutoML) lifecycle. From gathering requirements for specific business use cases and optimizing models via data enhancement and tuning to making available actionable insights, Aible focuses on delivering rapid business impact by automating this lifecycle. Working with Intel, key aspects of this multistep process have been identified and optimized and have empirically impacted how models are optimized for performance and adopted by the industry.
The Solution
Gains of Serverless Approaches
Aible takes advantage of a serverless approach to improve workload performance and further optimize its applications. This avoids the challenges with server-oriented architectures where upwards of 70% of time and costs are tied up in infrastructure overhead which isn't improved by the performance of the processor. This includes overheads for cluster scale-out, VM launch, establishing network connections, provisioning control planes, copying data, and other latencies associated with managing the operation and costs of server infrastructure.
With Aible’s serverless infrastructure, these unrelated activities and costs are mostly eliminated, and infrastructure spend goes directly toward useful compute.
Figure 1. Server and serverless infrastructure comparison.
Aible and Intel benchmarked the overall impact on three key areas: cost per job, total cost per ownership (TCO), and elapsed training time.
All are benchmarked on two architecture environments: a serverless architecture based on AWS* Lambda and AWS server-oriented architecture based on Kubernetes*.
The study demonstrated a better experience on serverless computing compared to traditional server architectures with comparable older-generation Intel processors due to limitations on the AWS Lambda side.
When deployed to serverless functions the application was:
- 2–3x more cost-effective
- 3–4x lower TCO
- 2–3x faster than on server architecture
Figure 2. A server versus serverless comparison on cost, time, and TCO.
Given the lack of a newer AWS Lambda that is based on an Intel® Xeon® Scalable processor, to understand the benefits of the newer Intel® architecture, we studied the performance on the server-oriented architecture, specifically m6i versus m4 instances on AWS.
We found model training to be more than twice as fast on the latest 3rd generation Intel Xeon Scalable processors over Intel® Xeon® E5 processors v4.
Dataset used: 250k and 500k row samples from Kaggle Bank loan default. Includes time to feature encodem model train/test on TensorFlow* 2.7, and to generate model metrics and model driver analysis.
Figure 3. Processor comparison: Intel Xeon E5 processor v4 versus 3rd gen Intel Xeon Scalable processor
An important thing to note about Aible's benchmarks is that they are not simply based on running algorithms on a set number of data points. They also factor in things like hardware requirements, networking infrastructure, and other environmental conditions that can affect performance. This makes their results more reliable than those from other sources, which may only test specific scenarios.
Analysis & Benefits of Intel® Xeon® Processors & Developer Toolkits
Aible was built ground up with a serverless-first mindset and from standard libraries and open source frameworks. For example, Aible was using stock machine learning packages like TensorFlow 1.15 with graph running: Not all Intel optimizations were part of this framework.
Working with Intel, Aible team migrated the solution to 3rd generation Intel Xeon Scalable processors and Intel-optimized machine learning packages with oneAPI support.
The Aible solution uses Intel® toolkits and benefits from faster performance and lower TCO using the 3rd generation Intel Xeon Scalable processors and Intel-optimized versions of TensorFlow, NumPy, and SciPy.
Highlighted Improvements
TensorFlow 2.7 and higher is Intel optimized: Aible benefitted from Intel® Advanced Vector Extensions 512 and showed the following performance gains:
- 2x for classification or regression models
- Up to 20x for business use cases that use transfer learning incorporating language (transformer) or convolutional neural network (CNN) image models
Intel-optimized versions of NumPy and SciPy: Aible showed the following performance gains:
- 2x across-the-board speedup without any changes to Aible code.
- 15–20x speedup for transcendental functions, for example, np.exp, np.log.
- 2x and more speedup of scipy.special.logsumexp and scipy.stats.norm.
The overall performance gains showed that Aible, when run on 3rd generation Intel Xeon Scalable processors, delivers results in half the time over older generation Intel Xeon E5 processors v4. Since Aible is serverless (if serverless configuration were to use newer generation processors), companies like Aible would benefit from higher performance, which in turn would translate to their customers benefitting from gaining faster insights in less time.
The Result
Organizations are enabled to make quicker and better strategic and tactical decisions that deliver business value quickly.
Customer Case Studies
The Aible MLOps solution on Intel architecture has offered 25 organizations access to AI, enabling them to solve their toughest business problems ranging from organizing the company holiday party to optimizing the company's IT strategy. Each of these organizations has worked closely with a dedicated team of Aible and Intel experts to see the solution in action. In many cases, these engagements have been completed in fewer than 30 days.
The Enterprise AI Solution of Choice
Today’s enterprise IT infrastructure leaders face significant challenges in building a foundation that is designed to help teams drive value from AI initiatives in the data center. Aible has been helping business teams across key industries deliver measurable business impact from AI within days by using Intel Xeon Scalable processors with Intel-optimized AI software. The reduced risk and time-to-value delivered by the Aible enterprise AI solutions, powered by Intel, are central to the vision of an AI Everywhere future.
Learn More
About Aible
Aible is an end-to-end and automated AI solution that eliminates the need for organizations to manage the AI infrastructure and data. With Aible, organizations can focus on their business challenges and maximize the impact of AI. This is validated by long list of case studies published here, that details the challenges and timeline to success. Aible provides enterprise AI solutions, transforming how companies make strategic decisions, act optimally, react to changes, and align across the organization using AI as an enabler for collaboration at scale. The solution offers tools for conducting scenario analysis and assumption testing, advises on the predictive model and resourcing, creates datasets for model retraining, enables end users to provide direct feedback, monitors business outcomes, and more.
Related Articles
- Optimize Distributed AI Training Using Developer Toolkits
- Accelerate AI Deep Learning with Intel® oneAPI Deep Neural Network Library (oneDNN)
- Deliver Fast Python* Data Science and AI Analytics on CPUs
- A Scale-Out Training Solution for Deep Learning Recommender Systems