Tech Insights

7 spectrums of edge AI to consider before deploying

To truly realize the possibilities of edge AI, modern organizations should consider seven workload characteristics and develop solutions that deliver the capabilities needed for desired outcomes. Here, HPE's Arti Garg and Lin Nease discuss what you should consider before deploying edge AI.

HPE-Edge AI-GTC21-blog.jpg

Artificial Intelligence (AI) has generated tremendous buzz due to its potential to transform enterprise operations with real-time decision-making and automated event response. It does this by processing and analyzing data locally—at the edge where it’s generated—thereby avoiding bandwidth and connectivity challenges associated with streaming everything to a centralized data center.

Despite the promise of this intelligent tomorrow, excitement around edge AI is tempered by complexity, uncertainty around edge security, and inexperience among IT teams. But these concerns can be mitigated when you look at edge AI as a spectrum—a range of use cases and patterns—instead of as a single application type. To truly realize the possibilities of edge AI, modern organizations should consider these seven workload characteristics and develop the solutions that deliver the capabilities they need for their desired outcomes.

1. Ruggedized vs. controlled-environment edge

Central to making decisions about the infrastructure for your edge AI solution is understanding the location where it will be deployed. Manufacturing quality assurance applications, for instance, may require placing compute servers on the factory floor in order to quickly deliver computer vision inference results as products move through the assembly line. Likewise, electric utility transformer yards process video surveillance streams onsite to avoid the high costs of sending terabytes of data to a data center.

These deployments represent harsh environments, where you need ruggedized equipment that can withstand metal dust in the air, explosive materials, or the direct heat of the sun.

But not all edge deployments require such robust protection. Office buildings, retail stores, and warehouses are examples of edge environments that might be suitable for the same kind of equipment you can deploy in a data center. Safe within a wiring closet—or even the air-conditioned car of a train—these controlled-environment edge deployments may include infrastructure that looks very similar to what you might see in a more standard IT environment.

2. Streaming vs. batch inference

In a perfect world, all edge AI tasks would be handled by streaming inference, or real-time analysis of collected data. A perfect world is expensive. That’s why you need to make important decisions around when you need to continuously apply your AI models to data as it’s created, and when you can do it “offline” at regular intervals.

Consider quality assurance and fraud detection workloads. Companies need this information right away to check on server functionality or to discover instances of fraud before they need to be remediated. By contrast, a workload like supply chain monitoring doesn’t require immediate turn-around, and batch inference conducted on data transferred to a centralized data center may be easier to implement and more cost-effective than real-time analysis.. In some cases, you may find a middle ground, running batch inference, at either the edge or in a centralized facility, at regular 20-minute intervals to give you “near real-time” results when you need them.

Depending on your network infrastructure, data sources, and edge workloads, you’ll most likely want to take a hybrid approach to batch and streaming inference to optimize costs while informing better decision-making.

3. Always connected vs. asynchronous communications

It’s impossible to maintain constant connection to all of your edge devices. An oil tanker in the middle of the ocean can’t communicate with your data centers every moment of its journey; the ease and cost effectiveness of data transfers fluctuates as it crosses geographies and environments. In fact, rarely can edge devices be “always” connected.

And yet, we often need connectivity the most during the one percent of the time that we lose it. Imagine that your connection to a critical battery is severed. Did the battery fail? Is the computer down? Or did you simply lose contact?

This is where most edge deployments go wrong. You can’t walk in with the same assumptions of connectivity to edge assets that you would expect in a datacenter. Rather, it’s essential to determine your tolerance for missed connections and design your algorithms to expect communication lapses. That could mean building in delays and multiple touch points to the decision-making process. Some companies, like oil and gas drillers, take this to the extreme, helicoptering servers with petabytes of data from offshore drilling platforms to centralized datacenters and running batch analytics to draw conclusions from their edge data.

4. Continuous vs. periodic model retraining

AI is only as good as your model. For many workloads, AI models are updated regularly based on system feedback and enriched data. In most scenarios, organizations mimic continuous model retraining by batching information uploads at scheduled intervals. An autonomous vehicle, for example, sends updates to a centralized data center when it’s most efficient to do so.

But what if automakers and other enterprises could create a closed-loop system at the edge?

New AI applications make it easier to make continuous adjustments to edge AI models. They can review which input led to a particular conclusion, and then predict when that event may happen again. The applications of these technologies are exciting and diverse, including facial recognition workloads and more.

5. Uniform vs. mixed edge sources

How many edge sources will your solution require? It’s a great question to ask early on in your journey. Too often, enterprises deploy without thinking about the logistical challenge of myriad edge devices, including costly part replacement and maintenance.

While fewer (uniform) edge sources are usually preferred, some workloads require many input types to arrive at a desired conclusion. A battery factory may be running QA on their product based on images of product output on the assembly line—but they also require an additional sensor to detect humidity levels during electrode production, which can affect how they interpret the photographs.

Edge AI tools can help reduce the complexity of mixed edge sources, speeding your time to insight. Using sensory fusion, AI applications can synthesize data from multiple sources to create a multivariable decision tree.

6. Hierarchical vs flat edge capabilities

Should your edge AI solution consist of only devices that have advanced AI capabilities? Or should you introduce a hierarchy, where less capable devices transfer data to more intelligent edge-compute servers for aggregation and inference? The answer is usually in between. You need to design a solution that does the job without unnecessarily wasting time or adding cost.

Video surveillance is a great example. Smart cameras have enough onboard computing capability to process and analyze the data they’re collecting by themselves. But it also makes sense to build a hierarchical solution that includes standard cameras and an edge server that can collect and process the data on site for significantly lower cost.

Before you set up your environment, be sure you’re not throwing money at a problem that the right edge configuration could solve better and more cost-effectively.

7. Automated response vs. decision support

To truly reap the benefits of edge AI, it’s essential to think about how it will influence your decision-making processes. In some cases, edge AI might allow you to completely transform your operations. For example, if you can predict that a manufacturing robot is going to fail before it does, you might want to configure a more responsive maintenance program that enables higher uptime on your assembly line.

Be careful about automating your responses based on edge AI, though, as it can create significant consequences when it doesn’t work right. Take COVID-19 protocols: Companies are looking at fever detection doors that facilitate touchless entry into a facility. While this may be a great example of how edge AI can help keep your workforce safer, you need to remember to include a failsafe if something goes wrong. Otherwise, you could lock out your entire workforce due to a malfunctioning infrared camera!

Living on the edge

Like most things in life, edge AI isn’t black and white. It encompasses a spectrum of characteristics, capabilities, and use cases. To successfully implement edge AI, organizations need an IT team that knows how to make optimal decisions about the best solution for their environment and application. Among recently surveyed executives, 84 percent fear missing their growth objectives if they don’t scale their artificial intelligence efforts. By thinking through some of these edge AI application characteristics ahead of time, you can scale your deployments while avoiding unforeseen costs, operational challenges, and inefficiencies in the future.

Edge experts at HPE can help you make the right decisions and plan your overall edge strategy, while providing the right combination of hardware to meet your goals. HPE ProLiant DL380 servers and HPE Apollo 6500 Gen 10 systems are NVIDIA-certified, meaning the systems have been comprehensively tested by HPE engineers to help ensure a faster, more reliable and predictable deployment. These GPU-accelerated applications can help organizations deliver unique technology solutions, develop new business models and increase operational performance. In addition, HPE Edgeline Systems provide enterprise-class IT in a single, rugged system suited for harsh environments to enable innovative new abilities at the edge. This comprehensive AI solution brings datacenter-level compute and management technology to the edge, helping accelerate time-to-value for enterprises.

Related articles:

Want to know the future of technologySign up for weekly insights and resources

Meet our Tech Insights bloggers

Arti Garg-HPE.pngArti Garg is Head of Advanced AI Solutions & Technologies in the AI Strategy & Solutions team at HPE. Previously she held data science leadership roles in a number of sectors including renewable energy, industrial products, and data center operations. In the past, Dr. Garg worked for the White House Budget Office where she oversaw R&D investments at the Department of Energy. She holds a PhD in Physics from Harvard University and an MS in Aeronautical & Astronautical Engineering from Stanford University.


Lin Nease-HPE.pngLin Nease is CTO for HPE’s IoT Advisory Practice. In this role, he sets strategy and builds technology roadmaps both for key customers and for HPE in the emerging areas of edge infrastructure and IoT. In his 30+ years at HPE, Lin has been CTO for HPE’s Networking business, CTO for Business Critical Servers, and co-founded multiple HPE businesses that have enjoyed widespread growth, including edge compute, the server blades business, and the Superdome franchise of products prior to that. He holds 9 patents, a BS Computer Science (Ariz State U), and an MBA (Cal State Sacramento).

Insights Experts
Hewlett Packard Enterprise

0 Kudos
About the Author


Our team of HPE and other technology experts shares insights about relevant topics related to artificial intelligence, data analytics, IoT, and telco.

HPE Webinars
Find out about the latest live broadcasts and on-demand webinars
Read more
Online Expert Days
Visit this forum and get the schedules for online Expert Days where you can talk to HPE product experts, R&D and support team members and get answers...
Read more
View all