Beginner

Edge AI Hardware

From microcontrollers with kilobytes of RAM to GPU-powered single-board computers - the hardware landscape for running AI at the edge.

Hardware Comparison

DeviceCPU/GPURAMAI PerformancePrice
Arduino Nano 33 BLEARM Cortex-M4256 KB~1 MOPS~$25
ESP32-S3Dual-core Xtensa512 KB + PSRAM~5 MOPS~$5
Raspberry Pi 5ARM Cortex-A764-8 GB~10 TOPS (with HAT)~$60-80
Google Coral Dev BoardARM + Edge TPU1-4 GB4 TOPS~$130
NVIDIA Jetson NanoARM + 128-core GPU4 GB472 GFLOPS~$150
NVIDIA Jetson Orin NanoARM + Ampere GPU8 GB40 TOPS~$250

Microcontrollers for TinyML

Microcontrollers (MCUs) are the most constrained edge devices. They run TinyML models for always-on sensing:

  • Arduino Nano 33 BLE Sense: Built-in microphone, accelerometer, and gyroscope. Perfect for keyword spotting and gesture recognition. Uses TensorFlow Lite for Microcontrollers.
  • ESP32: Extremely cheap ($5), built-in Wi-Fi and Bluetooth. Good for IoT applications that need connectivity with basic ML inference.
  • STM32: STMicroelectronics' family offers a range of ARM Cortex-M MCUs with STM32Cube.AI tooling for model deployment.

Raspberry Pi

The Raspberry Pi is the most popular edge AI platform for prototyping. It runs a full Linux OS and supports Python, making it easy to deploy ML models:

  • Pi 5: Quad-core ARM Cortex-A76, 4-8 GB RAM. Fast enough for real-time image classification and lightweight object detection.
  • AI HAT+: Hailo-8L accelerator add-on providing 13 TOPS of AI performance for the Pi 5.
  • Camera Module: Official camera modules for vision AI projects. Supports 12MP stills and 4K video.

NVIDIA Jetson

The Jetson family provides GPU-accelerated edge computing for compute-intensive AI workloads:

  • Jetson Orin Nano: Entry-level with 40 TOPS. Runs YOLOv8 at 30+ FPS for real-time object detection.
  • Jetson Orin NX: Mid-range with 100 TOPS. Handles multiple camera streams and complex vision pipelines.
  • Jetson AGX Orin: Top-end with 275 TOPS. For autonomous robots, drones, and advanced multi-model workloads.
  • Software: JetPack SDK includes CUDA, cuDNN, TensorRT, and DeepStream for optimized inference.

AI Accelerators

AcceleratorForm FactorPerformanceBest For
Google Coral Edge TPUUSB stick / PCIe / SoM4 TOPSEfficient INT8 inference
Intel Neural Compute StickUSB stick~1 TOPSOpenVINO inference
Hailo-8M.2 / PCIe26 TOPSHigh-performance edge AI
Apple Neural EngineBuilt into Apple Silicon15.8 TOPSiOS/macOS on-device AI
Qualcomm Hexagon DSPBuilt into Snapdragon15+ TOPSAndroid on-device AI

Choosing the Right Hardware

  • Always-on sensing (keyword detection, motion): Microcontroller (Arduino, ESP32) - microwatts of power, pennies per unit.
  • Image classification, simple detection: Raspberry Pi + AI HAT or Google Coral - good balance of capability and cost.
  • Real-time video, multi-model pipelines: NVIDIA Jetson - GPU-accelerated, handles complex workloads.
  • Mobile apps: Use the device's built-in NPU (Apple Neural Engine, Qualcomm Hexagon).
Key takeaway: Choose hardware based on your model's compute requirements, power budget, and deployment scale. Start with Raspberry Pi for prototyping, then move to purpose-built hardware (Jetson, Coral, or MCUs) for production. The right hardware choice depends on the tradeoff between performance, power, and cost.

Ready to Go Deeper?

Live instructor-led courses from our partners. Affiliate disclosure.