Edge vs. Cloud: Which AI Infrastructure to Opt For?

Edge vs. cloud, the ever-growing debate in this fast-paced tech evolution

Edge computing is all the buzz these days. Touted as the most exciting technology shift in recent years, discussions about its transformative powers are aplenty! With increasingly powerful AI/ML algorithms redefining "intelligence" and the availability of cheaper more powerful "edge" devices, the hype is turning out to be largely real. But if one were to consider the history of edge computing, it goes further back than what the recent interest would have us believe. In fact, computing and intelligence first started at the edge at a time when high bandwidth network connections were virtually non-existent for most applications. Even in the late 1990s, critical measurement devices deployed remotely in a plant or a field often had dedicated computational capability for processing incoming sensor data. The algorithms in these devices were however only rudimentary in their "intelligence": mostly signal processing or data transformations. With improvements in network capability and increased connectivity, cloud-based computing started to gain traction in the late 2000s. In parallel, powerful AI algorithms rose to prominence as a means to unlock meaningful information from swathes of structured and unstructured data. In just about a decade, cloud AI had become the go-to choice for AI applications. But the shift to the cloud brought with it several concerns as well: data upload and download costs, network reliability, and data security to name a few. At the same time, the trade-off for edge computing between processing capability and cost or footprint was diminishing with the rise of affordable yet powerful edge devices. It seems that we have now come full circle back to considering "edge computing" as a viable and attractive option for building intelligent applications.

As the debate rages on about which option is better – edge AI or cloud AI, anyone who is familiar with these two frameworks would likely respond that "it depends!". The reason is that edge and cloud infrastructures are not competing but complementary frameworks. Both have seen tremendous evolution and sophistication in the last few years, particularly as a base for AI development and deployment. As with any technology selection, the choice really boils down to the specific application: the objectives, value drivers, and economics as well as any constraints on power, footprint, and connectivity. Thus, it is imperative to understand the pros and cons of both cloud and edge AI before attempting to build the right infrastructure.

Cloud-based AI is an attractive choice when seeking flexibility, scalability, and ease of deployment. Most cloud service providers today provide robust frameworks for training and deployment of AI models along with pay-as-you-go packages with little to no upfront commitment or investment. The cloud offers computational and storage options with few limitations, making it particularly suited for large AI models. But it can become an unwieldy option for real-time applications that require continuous assessment of sensor or image data as they incur significant costs by having to stream data back and forth. This data transfer also makes the cloud largely unsuited for low latency applications requiring closed-loop control or immediate actions.

Edge AI on the other hand is the logical choice for real-time data analysis for automated alarms or closed-loop control. While edge infrastructure does require an upfront investment in edge hardware, the operational costs are significantly lower compared to that for the cloud. Today, a wide variety of edge AI hardware options are available including NPUs (neural processing units), TPUs (tensor processing units) as well as SoCs (system on chip) and SoMs (system on module) with dedicated AI accelerators. Low cost and low power hardware for AI is an area of active research and is likely to provide superior options going forward. On the flip side, AI-based consumer applications have to deal with rather diverse edge devices (mobiles, tablets, PCs, etc), making edge deployment a potentially daunting prospect. Hence, edge infrastructure may not be conducive to rapid prototyping and doesn't scale as easily either. While federated learning, the concept of distributed training for AI models allows for both training and deployment on the edge, the cloud remains the logical choice for training large models requiring adequate computational power.

But the solution doesn't necessarily have to be an either-or choice. As applications transition to a more microservices-based architecture, they can be broken down into smaller functionalities or microservices with their own specific deployment framework. So instead of having to choose between cloud and edge, the focus can be on using both optimally for a specific application. For example, an application might start off with a quick prototype on the cloud. As it evolves, functionalities that require low latency and real-time decisions can be transitioned to the edge, while those that require scale and flexibility can be retained in the cloud. Model training or re-training can be centrally managed in the cloud while some federated learning on the edge can increase accuracy locally. Similarly, sensitive data can be processed on the edge and more generic data relegated to the cloud.

Organizations, developers, and practitioners would do well to think of the cloud and the edge not as distinct alternatives, but really as a continuum from edge to cloud with many diverse infrastructure options in between. That includes different types of edges – operational edge, network edge, mobile endpoint, etc, and different types of distributed processing on the network – private cloud, public cloud, cloudlet, fog computing, and so on. While the complexity can be a challenge, finding the right mix of technologies is starting to present a unique opportunity for organizations to maximize the value of AI while simultaneously minimizing the cost and risk.

Author:

Anusha Rammohan, Member of The IET and Senior Technology Leader, Myelin Foundry