Artificial Intelligence has been effectively used across industries and platforms, including the industrial edge, for some time. What’s important to focus on now is how to expand those opportunities. For example, it’s currently possible to use AI to regulate the temperature of, say, a tank holding certain chemicals. What organizations can’t do in that situation is integrate additional input from external meteorological forecasts to conserve the power required to heat and cool the tank — yet.
It won’t be long before we get to the “yet.” Actually, it won’t be long before that level of integration seems behind the curve because AI will have leapfrogged in application, capability, and, importantly, interoperability. Organizations that are prepared to harness this whirlwind of AI change at the edge will be competitive. The key is meeting the technology where it is now and getting ready for where it will go. But, first, it’s important to think about where it has been.
For years, AI has been used within edge environments in retail, healthcare, manufacturing, and other sectors. The opportunities to expand these opportunities, however, have been impacted by purpose-built, siloed systems that lock in data, fragmented and low performance devices, gaps in AI expertise and staff, capabilities of the available AI technology, and most importantly, vision for how AI at the edge could evolve.
Today, several factors have come together to transform decades-old systems — among them, more powerful edge devices, enterprise-grade open source edge platforms and tools, shrinking large language models (LLMs), a growing understanding of the need to break down IT and OT silos to optimize data, and — of course — modern AI applications that can process large amounts of data locally, as it’s collected.
Biggest Factors for Accelerating Edge AI Evolution
As we see edge and AI converge, it’s crucial to think about the ability to perform training at the edge. And one of the biggest enablers will be taking AI models, including foundational models and LLMs, and shrinking them down to work at the edge.
While the biggest foundational models or LLMs may contain billions if not trillions of parameters, there is a push to make these models “compute optimal” and to leverage models that are more organization-, industry- or use-case- specific. Shrinkage of large models and specific optimized models will eventually enable them to be trained at the edge based on data generated at the edge. With this in place, data and algorithms could be adapted locally, and the time and cost it takes to move data back and forth from the edge to the cloud would be significantly reduced, if not eliminated.
The other big factor in moving the needle for AI at the edge is the amount of compute power that is rapidly increasing in edge devices that are decreasing in size and weight just as quickly. For example, right now you can get a Mac notebook with a 16-core CPU, 40-core GPU, 48GB of memory, and 1TB of SSD storage — for about $4,000. Devices with that much performance and storage weren’t conceivable even a couple years ago, but today they can enable organizations to address workloads up to and including AI training and inferencing — even in space-constrained locations with little or no onsite technical support.
Taking Notes from the Container Playbook
Augmented by automation and self-healing capabilities, the use of container-based technology will also be key to not just deploying but also optimizing AI workloads at the edge. In fact, the container playbook is very similar to the edge AI playbook: With containers, you go from purpose-driven machines to more general-purpose platforms, and from a monolithic application architecture to a microservices architecture. With edge AI, we are moving from siloed data that is very difficult to share to an underlying infrastructure that is more flexible and agile — one that can support multiple different use cases without the need to reinvent the wheel each time.
And, just as the use of containerization technology exploded, lightweight Kubernetes layers are poised to provide many of the benefits of container orchestration where workloads are run and data is collected — in a lightweight footprint that uses the same tools that developers and administrators are already familiar with.
Indeed, it’s important to note how important open standards, open source code and open collaboration will be to ensure organizations can make the most of AI at the edge. Siloed thinking is just as limiting — if not more so — than siloed data.
AI at the edge has the potential to literally change — and even save — people’s lives. The pieces are falling into place for evolving what’s possible at the edge; now is the time to take advantage.
