From Cloud to Edge AI: Embedded Hardware Guide

Artificial Intelligence has shifted from the data centre to the edge. The AI of Things (AIoT) is no longer a buzzword but a core engineering reality. From smart homes and industrial automation to ADAS in EVs and autonomous robotics, the demand for localised intelligence is driving innovation in embedded hardware.

Design engineers are on the frontlines of this transformation. Their role is to bring intelligence directly into end products by tackling the challenges of running AI workloads on constrained devices, while selecting the right SoCs, SOMs, and development kits that make edge intelligence both practical and deployable.

Why Running AI on Edge Devices is Challenging

Cloud-based AI models thrive on massive compute and storage. Edge devices must instead meet:

• Low latency – milliseconds matter for robotics, ADAS, or production line defect detection.
• Bandwidth efficiency – sending raw video/audio data to the cloud is expensive and impractical.
• Privacy and security – facial images, health signals, and voice commands often cannot leave the device.
• Power and cost budgets – embedded AI must fit into fanless boxes, battery-powered nodes, and cost-sensitive consumer devices.

The solution is localised AI models accelerated by NPUs, DSPs, FPGAs, or ultra-low-power Neural Network engines. These accelerators allow real-time inference without draining power or requiring cloud connectivity.

Design engineers play a critical role in selecting the right development kits and reference platforms, bridging high-level AI concepts with deployable hardware solutions.

AI-Capable Dev Kits and Boards

The transition from cloud-centric AI to edge intelligence would not be possible without a new generation of AI-ready hardware platforms. These development kits and boards serve as the practical foundation for design engineers, offering a balance of compute performance, energy efficiency, and integration flexibility. More than just processors, they bundle together optimised silicon, software frameworks, pre-trained models, and peripheral interfaces, effectively reducing the complexity of deploying AI workloads on constrained devices. Whether the goal is ultra-low-power keyword detection in a wearable or high-performance machine vision in industrial automation, these kits provide the building blocks that turn AI concepts into deployable products. Let's explore the most advanced AI-capable devkits and SBCs to stay ahead of the game.

AMD Kria™ K24/K26 SOMs & Starter Kits

• Kria Range K24/26: Built on Zynq™ UltraScale+™ MPSoC: Arm Cortex-A53 + R5F cores with programmable FPGA fabric.
• The FPGA enables custom accelerators such as Xilinx’s DPU, ideal for CNN inference and video pre-processing.
• KV260 Vision AI Starter Kit: delivers up to 3× higher AI throughput per watt than many SOMs.
• Primary use cases: industrial machine vision, robotics, ADAS prototyping, and deterministic real-time analytics.

Arduino Portenta & Nicla Sensors

• Arduino Portenta H7 / X8: Dual-core MCUs/MPUs supporting TensorFlow Lite and MicroPython, ideal for connected IoT + AI workloads.
• Arduino Nicla Sense ME: Tiny sensor-rich module with IMU, gas, humidity, and barometer—optimized for ML-based condition monitoring and environmental sensing.
• Portenta + Nicla ecosystem bridges prototyping and production, letting developers scale from proof-of-concept to deployable solutions.
• Community-driven software and accessible APIs lower the barrier to entry while maintaining professional-grade hardware reliability.