gesponsertBuilding Systems for the Long Term Industrial AI at the Edge

Cloud AI is powerful, but it is not the right answer in every situation. For industrial, medical, and robotics applications, Advantech is focusing on local AI, customized hardware and systems that give customers greater control over data, latency, and system performance.

For a surgical robot, the latency of a cloud connection is not an abstract IT metric. If an image has to be transmitted to a remote location for processing and then sent back, latency and dependence on a stable connection can become real operational concerns. Tony Chen, Vice President and Head of Advantech’s Applied Computing Group (ACG), which provides Design and Manufacturing Services (DMS), sees this as a clear example of why the future of industrial AI will extend beyond the data center. Chen emphasizes that, in a hospital environment, latency and connectivity can become critical considerations in time-sensitive medical applications. That does not mean the cloud is going away. Large models and compute-intensive workloads will continue to have their place there. But another class of systems is growing alongside them: machines, medical devices, vehicles and robots that can run AI locally and respond to data without having to ask a data center for every decision.

Chen sees particularly strong potential for local AI adoption in Europe. Not because European developers are fundamentally opposed to the cloud, but because data protection, latency, security, and the desire to keep proprietary know-how under their own control tend to play a particularly important role in Europe. But AI adoption is not only a hardware challenge. Integrating AI into existing software environments — particularly legacy systems — can be equally complex.

Not every form of intelligence needs to live in the cloud

Advantech comes from a world in which embedded and industrial computers remain inside machines for years or even decades. The first wave of digitalization connected devices to networks, and later more and more data moved into cloud systems. With Edge AI, part of the processing is now moving back toward the device. For Chen, this is not an either-or decision. Cloud and Edge AI serve different purposes. Large models can run in the data center, while smaller or more specialized models handle tasks directly at the machine.

In a production system, that could mean making a quality decision. In a medical device, it could mean processing an image. In a robot, it could involve assessing the surrounding environment. The important point is that not every piece of information has to be transmitted before the system can react. This reduces latency while also allowing sensitive data to remain within the company’s infrastructure.

Chen observes the latter particularly often among European customers. They do not necessarily want to transfer their proprietary knowledge, algorithms, or process data to an external service. For a medical technology company, software expertise developed over many years can be a major component of the product itself. Chen puts it in a somewhat unusual way: European companies want to keep their own “personality” within their AI systems. By this, he means that their own software logic, algorithms, and know-how should remain within their system. A company that has developed its own diagnostic functions, machine control systems, or image-processing technologies may not want the associated data and algorithms to leave its own infrastructure. For Edge AI, this represents a practical way to keep proprietary intelligence and data within the company’s infrastructure.

A new processor can mean two years of software work

In theory, building an AI system sounds straightforward: choose a suitable computing platform, install the required software, and get started. In industrial applications, however, it is rarely that simple. Advantech offers a wide range of standard platforms. In parallel, Chen’s Applied Computing Group works with customers whose requirements do not fit an existing configuration. That may involve computing performance, form factor, or environmental conditions — from high or low temperatures and limited space to special interfaces or specific AI acceleration requirements.

But the more difficult adaptation is typically not the hardware. “European customers place a strong emphasis on their own software,” Chen says. Moving from one processor platform to another can create considerable additional work. If an application has been developed and validated for a particular x86 platform over many years, switching to another processor platform may sound easier than it actually is. Some customers estimate that such a migration could require as much as two years of software work. In medical technology in particular, a platform change can also trigger additional validation and testing.

A new chip is then not simply a new chip. This is also why Advantech does not see customization as merely changing a PCB. The work can just as easily involve migrating existing software to new hardware while minimizing disruption to the customer’s existing development environment.

WEDA bridges Edge AI deployment across industrial verticals

To address this challenge, Advantech provides WEDA, the WISE-Edge Developer Architecture, as a common software layer for AI and IoT applications. Advantech describes WEDA as a modular development framework intended to provide common software functions across different edge systems. Chen describes the idea behind it as a kind of “glue layer” – a layer that holds the different elements together.

The goal is to reduce the need for developers to work with completely different interfaces and software conventions across Intel- and AMD-based systems, NVIDIA platforms, or Arm-based architectures. WEDA provides development tools such as containers, APIs and SDKs, while Advantech tries to absorb some of the hardware differences underneath. Those differences do not disappear entirely, of course. An Arm-based system is still technically different from an x86 platform, and an API alone cannot automatically make every software application portable across architectures. But the migration effort can become smaller.

According to Advantech’s own evaluation, WEDA can reduce development time by around 30 percent in suitable projects, Chen says. That is a vendor figure and will depend on the individual project. For Advantech, however, there is also a simple business rationale behind it: hardware has little value if the customer cannot finish the product. “If our customer cannot deliver a product, we are in the same boat,” Chen says.

Standard hardware remains important – but not for every application

The shift toward customized systems does not, however, make standard products obsolete. For many applications, they are still the fastest route to a finished product. Advantech offers both a broad portfolio of standardized products and its Applied Computing Group (ACG), which provides Design and Manufacturing Services (DMS) for customers whose requirements go beyond standard configurations. Sometimes the necessary modification is minor. In other projects, the result is a substantially different platform.

The starting point is not necessarily a specific processor or architecture. Chen names Intel, AMD, NVIDIA and several Arm-based platforms as possible foundations. The key question is which platform best fits the application. Focusing on a single architecture can make development more efficient. At the same time, it increases dependence on that architecture remaining successful and available over the long term.

Advantech accepts the trade-off: supporting several platforms in parallel requires additional investment and engineering resources. Chen nevertheless considers this necessary to give customers greater flexibility in choosing the platform that best fits their applications.

Betting on only one ecosystem can go wrong

The distinction between x86 and Arm is becoming less clear. At the same time, AI accelerators, NPUs, and GPUs are reshaping traditional embedded architecture. No one can reliably predict which platform will dominate a particular application five or ten years from now. Advantech deliberately keeps several technology paths open. Chen puts it simply: do not put all of your eggs in one basket.

That strategy is not free. Maintaining a multi-platform strategy requires continued engineering investment, but Advantech sees this as essential to giving customers long-term flexibility and reducing dependency on any single technology architecture. This approach also allows Advantech to remain prepared as market and geopolitical conditions change.

Embedded systems add another complication: product lifecycles are much longer than the typical innovation cycle of a processor. Chen speaks of projects that run for five or ten years. Some products are still being shipped 15 years after the first delivery. Advantech aligns technology roadmaps not only with chip suppliers, but also with the long-term needs of its customers.

Just-in-Time is becoming Just-in-Case

Recent supply constraints show that this dependency is not just a hypothetical possibility. Memory prices and availability can come under pressure again as market conditions change. According to Chen, Advantech tries to identify such developments early, negotiate contracts with suppliers and qualify alternative components for customers.

The logic has changed compared with previous years. “Supply chain used to mean Just-in-Time,” Chen says. “Today, it is more like Just-in-Case.” More inventory does not solve every problem, however. An alternative memory device or processor still has to be tested and qualified in the target industrial system. Advantech qualifies multiple memory variants and configurations for selected platforms.

For an individual machine builder, the same second-source qualification work requires engineering resources that could otherwise be spent on its own product. The service may be less spectacular than an AI demonstration at a trade show. For a product that is designed to remain available for ten years, it can be more important.

Europe needs more than hardware nearby

Chen’s view of Europe is notably pragmatic. He considers the market attractive for customized systems, but it does not function as one homogeneous economic region. Germany is not France, France is not Italy, and Italy is not Scandinavia. Advantech maintains local technical and sales teams across different European regions (Advantech's Design & Manufacturing Services. Made in Europe.). For Chen, it is not enough to ship hardware from Taiwan and then try to solve every problem by email or video call.

He describes a project in Italy where a mechanical issue could not be resolved clearly despite several discussions. Only when an engineer visited the customer did it become apparent that both sides had simply been looking at the problem from different angles. “It is about people and communication,” Chen says.

That does not entirely fit the image of an increasingly autonomous AI world, and it does not have to. The more tightly hardware, software, and application are connected, the more likely problems are to emerge at the interfaces. Sometimes a local engineer can solve more than the next digital support system.

A machine does not have to be allowed to do everything it can do

In the end, the Edge AI discussion comes back to a fairly traditional embedded question: what does the system actually need to do? A medical device requires different computing performance from an autonomous vehicle. A vision-processing robot needs different accelerators from a battery-powered system where every watt matters. And not every machine needs the largest AI processor available.

Chen describes Advantech’s role as more like that of a master chef than a semiconductor manufacturer. The ingredients come from different platform suppliers. The customer defines the desired outcome. Advantech then tries to build a system that fits the application.

Edge AI does not become suitable for industrial use simply by putting as much computing power as possible into a machine. It has to work with existing software, operate under real-world environmental conditions, remain available for years and continue to be supportable even if a chip or memory component disappears from the market. For European developers, another requirement often comes on top: they want to know where their data remains and who controls the system.

The cloud can be part of the solution. It just does not automatically have to be the place where every decision is made. (sb)

(ID:50949074)