The biggest topic facing embedded system designers today is undoubtedly AI. However, countless edge devices around us—refrigerators, door locks, industrial sensors, and more—still struggle to make decisions on their own without the help of the massive infrastructure of the cloud. This is because, despite the power of cloud AI, there are three clear limitations: latency, communication costs, and privacy.

Against this backdrop, microcontrollers (MCUs) are rapidly evolving beyond mere processors into "edge AI platforms" that directly interpret and judge data in the field. Infineon, a global semiconductor leader, is marking a milestone in this transformation with its next-generation MCU architecture, "PSoC™ Edge." In an interview with Jinyong Lee, Manager of Edge AI Solutions at Infineon, we explored the technical nature of edge AI and the practical challenges faced by field engineers.

1. Edge AI: Why Now?: The Limitations of the Cloud and "On-the-Face Knowledge"

In the past, IoT devices were merely "relays" that simply collected data and transmitted it to the cloud. However, with the advancement of autonomous driving, robotics, and smart home appliances, the importance of "instant decision-making" has grown. As sensor data continues to surge, transmitting it all over the network has become a practical challenge, both in terms of cost and security.

Manager Lee Jin-yong explained the changes on the ground as follows.

Cloud processing was common in the past, but network latency and privacy issues began to pose practical limitations. Infineon responded by evolving its MCUs beyond simple control into edge AI platforms that make decisions directly on the spot.

At the heart of this evolution is the PSoC™ Edge, a strategic outcome of Infineon's complete expansion of its design direction to enable high-performance AI inference even in low-power environments.

2. A revolution in hardware architecture: SRAM-centric design and 'flashless'

Technically, edge AI MCUs differ from general-purpose MCUs in their design. The most notable change is their memory structure. While conventional MCUs have relatively small SRAM and large internal flash, MCUs designed for AI computing have an overwhelmingly high proportion of SRAM.

AI models perform neural network operations by traversing numerous layers. The data layers (activation data) generated at each layer in this process instantly require enormous memory space. Manager Lee Jin-yong emphasized, "AI models absolutely require large SRAM capacity. In-built capacity in megabytes (MB) is required, a requirement that is on a whole different level from the hundreds of kilobytes (KB) of conventional MCUs."

Additionally, PSoC™ Edge boldly adopted the 'Flashless' concept, which eliminates internal flash. Because AI model sizes vary widely depending on user needs, the design allows for flexible external flash connections rather than insisting on a specific capacity of internal flash. This approach, similar to that of a typical application processor (AP), maximizes the flexibility of edge AI.

Infineon's PSOC Edge MCU image

3. A Power Efficiency God's Move: Multi-Domain and 'NN Lite'

The most severe constraint in edge environments is power. High-performance AI computation inevitably consumes significant power. Infineon addressed this issue by separating "low-power" and "high-power" domains.

At the heart of the system are the Cortex-M55 CPU and Ethos-U55 NPU, which handle high-performance computing. However, if these were always running, the battery would drain quickly. This is where Infineon's expertise shines. In situations where always-on is required, the low-power domain Cortex-M33 and Infineon's proprietary technology, 'NN Lite', operate.

During the interview, Manager Lee Jin-yong shared some impressive insights into the design philosophy of Edge AI.

"The key to edge AI design isn't about boasting maximum performance figures, but rather achieving a 'balanced design' that delivers predictable performance and power efficiency within limited resources. PSoC™ Edge delivers its greatest value when implementing systems that are always awake yet power-efficient."

In a voice recognition device, for example, NN Lite constantly detects wakeup words like "Hi Bixby" using very little power, waking up the high-power Ethos-U55 to perform precise inference only when a complex command is input. This dramatically reduces average power consumption while maintaining user-perceived performance.

4. Software that solves engineers' problems: DEEPCRAFT

No matter how excellent the hardware, it's useless if development is difficult. Most MCU engineers are skilled in control and communication design, but they often feel unfamiliar with Python-based AI modeling and frameworks like TensorFlow.

For this purpose, Infineon offers a powerful tool called DEEPCRAFT. This tool automates the entire process, from data collection to labeling, model optimization, and C code generation for MCUs. Manager Lee Jin-yong, as a practitioner who has witnessed countless trials and errors faced by engineers in the field, offered the following advice:

"MCU engineers don't need to be AI experts. Tools like DEEPCRAFT help engineers seamlessly integrate only the AI features they need while maintaining their existing firmware architecture. Simply defining the problem to be solved with AI is the starting point for a successful edge AI design."

He further emphasized the importance of "labeling," the most arduous task in AI model development, and emphasized that high-quality data determines model quality. He believes that embedded engineers should first consider where to deploy AI within the system architecture, rather than becoming bogged down in Python syntax.

DeepCraft UI image

5. The Future of Edge AI: Physical AI and Robotics

The interview concluded with a discussion on the future direction of edge AI. Manager Lee Jin-yong pointed out that in places like Korea, where the internet environment is perfect, it's easy to overlook the need for edge AI. However, in marine plants with unstable communications, industrial sites with severe communication shielding, and robotics where real-time response is crucial, edge AI is the only solution.

"Physical AI," which has recently been attracting attention, is the most attractive battleground for edge AI MCUs. This is because robots must independently process visual and auditory data and provide immediate physical feedback without external assistance. Furthermore, in the home appliance sector, efforts are underway to exponentially increase recognition rates by upgrading the "fall detection" function, which previously relied on simple radar sensor probabilities, to an AI model.

Manager Lee Jin-yong expressed his pride, saying, "This market is just beginning. Invisible limitations will continue to be broken down as technology advances, and Infineon will be at the forefront, providing optimal solutions."

Welcome to a New Era for MCU Engineers

Infineon's PSoC™ Edge is more than just a single chip; it's reshaping the grammar of embedded design. The perfect harmony of CPU and NPU, a high-performance SRAM-centric memory architecture, and ultra-low-power NN Lite technology are the triumvirate driving the popularization of edge AI.

Now the ball is in the engineers' court. There's no need to be intimidated by complex AI formulas. As Manager Lee Jin-yong said, by defining AI's place within the system architecture and leveraging the powerful toolbox provided by Infineon, anyone can become a creator of intelligent edge devices. The evolution of MCUs, from the era of control to the era of intelligence, is ongoing.