Lecture
Where Should the AI Run? Making the MCU, NPU, or Cloud Decision
- at -
- Hall C5C5.100
- Language: English
- Type: Lecture
Lecture description
"When adding AI to an embedded product, the first architecture question is often the hardest: should the workload run on a standard MCU, use dedicated AI acceleration, or move partly to the cloud?
This session presents a practical framework for making that decision based on five factors: model complexity, latency, memory, power, and connectivity/privacy requirements. Using a small open-source AI model running on a Cortex-M-class device as a real-world case study, we’ll show how to translate product requirements into measurable engineering constraints, then use those measurements to determine when an MCU is sufficient and when the architecture needs to change. The session will give attendees a reusable way to evaluate edge AI workloads across platforms.
Attendees will leave with a simple decision framework they can apply early in development to avoid both over-engineering and choosing hardware that cannot meet the product’s needs."
This session presents a practical framework for making that decision based on five factors: model complexity, latency, memory, power, and connectivity/privacy requirements. Using a small open-source AI model running on a Cortex-M-class device as a real-world case study, we’ll show how to translate product requirements into measurable engineering constraints, then use those measurements to determine when an MCU is sufficient and when the architecture needs to change. The session will give attendees a reusable way to evaluate edge AI workloads across platforms.
Attendees will leave with a simple decision framework they can apply early in development to avoid both over-engineering and choosing hardware that cannot meet the product’s needs."
