Edge Computing: Outlook of Future Compute 2022

A spotlight is now shining brightly on edge computing architecture as it appears to be taking over jobs now confined to established cloud computing methods.

Proponents hope that edge computing will reduce the amount of data sent to the cloud, enable real-time responses, and perhaps eliminate some of the mysterious line items that show up on a company’s cloud computing bills.

Moving some of the runtime AI processing from the cloud to the edge is a commonly cited goal. However, using graphics processing units (GPUs) for AI processing at the edge also incurs costs.

Edge is still a frontier with much to explore, as seen at a recent session on Edge Intelligence Implementations at Future Compute 2022, sponsored by MIT Technology Review.

How much does AI cost?

At Target Corp. Edge methods gained acceptance as the COVID-19 pandemic disrupted normal operations, according to Nancy King, senior vice president of product development at the mass-market retailer.

Local IoT sensor data has been harnessed in new ways to help inventory management, she told Future Compute attendees.

“We send raw data back to our data center toward the public cloud, but often we try to process it at the edge,” she said. Data is available faster there.

Two years ago, as COVID-19 lockdowns mounted, target managers began processing some sensor data from freezers to guide central planners regarding overstocks or stock shortages, King said.

“Edge gives us the answer we need. It also gives us the ability to respond faster without clogging the network,” she said.

However, she raised concerns about the cost of running GPU-intensive AI models in stores. So the problem of AI processor costs does not seem to be limited to the cloud.

With edge AI implementations, King says, “computational costs are not coming down fast enough.” Also, she said, “some problems don’t require deep AI.”

Edge orchestration

Orchestrating workflows at the edge requires the coordination of various components. That’s another reason the move to Edge will be gradual, according to session attendee Robert Blumofe, executive vice president and CTO at content delivery giant Akamai.

Edge computing approaches, which are closely tied to the increasing use of software container technologies, will continue to evolve, Blumofe told VentureBeat.

“I don’t think you would see a recording without a container,” he said. He marked this as part of another general trend in distributed computing: bringing the computation to the data, not the other way around.

In Blumofe’s view, Edge is not a binary edge/cloud equation. On-site and mid-level processing will also be part of the mix.

“Ultimately, much of the computation required can be done locally, but not all of a sudden. What’s going to happen is that data is going to leave the premises and go to the edge and the middle and the cloud,” he said. “All of these layers must work together to support modern applications securely and efficiently.”

The move to support developers working at the edge plays no small role in Akamai’s recent $900 million acquisition of cloud service provider Linode.

Akamai’s Linode operation recently released new support for distributed databases. This is important because the database space will have to change with new edge architectures. Architects will weigh edge and cloud database options against each other.

Balance and rebalance

Of course, early work with edge computing leans more toward prototyping than actual implementation. Implementers today must expect a learning curve in which to balance and rebalance types of processing across sites, said session attendee George Small, CTO at Moog, a maker of precision controls for aerospace and Industry 4.0.

Small cited oil rigs as an example of a place where fast-flowing timescale data needs to be processed, but not all data needs to be sent to the data center.

“You could end up working very intensively on site,” he said, “and then just push up the important information [to the cloud].” Architects need to be aware that different processes take place on different time scales.

In IoT or industrial IoT applications, this means that edge implementers need to think in event systems that combine stringent embedded edge requirements with looser cloud analytics and recording systems.

“Bringing those two worlds together is one of the architectural challenges,” Small said. While learning continues on the fringes, “it doesn’t feel too remote,” he added.

AI can explain

Much of the learning process involves edge AI or edge intelligence that injects machine learning into a variety of real-world devices.

But there are also people on this edge. According to Sheldon Fernandez, CEO of Darwin AI and moderator of the MIT Edge session, many of these devices will ultimately be managed by professionals, and their trust in the devices’ AI decisions is critical.

“What we’re learning is that with more and more powerful devices, you can do a lot more things on the edge,” he told VentureBeat.

But these cannot be “black box” systems. They must provide workers with explanations “that complement that activity with their own human understanding,” said Fernandez, whose company is pursuing alternative approaches that support “XAI” for “explainable artificial intelligence.”

On a side note, people doing jobs need to understand why the system classifies something as problematic. “Then,” he said, “you may or may not agree to that.”

Meanwhile, he hinted, AI processing users can now choose from a range of hardware, from regular CPUs to powerful GPUs and edge-specific AI ICs. And it’s a good general rule to do operations close to where the data is. As always, it depends.

“If you’re doing basic video analysis without hardcore timing, a CPU might be good. What we’re learning, like everything in life, is that there are few hard and fast rules,” Fernandez said. “It really depends on your application.”