Price: $35-75 | Best For: General edge computing, learning

Most accessible edge computing platform. Full Linux computer in credit-card size with GPIO pins for sensors, camera support, and enough power for edge ML inference.

Specifications (Pi 4):

• Quad-core ARM CPU (1.5-1.8GHz)
• 2GB/4GB/8GB RAM options
• WiFi, Bluetooth, Ethernet
• 40-pin GPIO for sensors
• Camera interface (CSI)
• USB 3.0 ports

Perfect For:

• Edge ML inference with TensorFlow Lite
• IoT gateway and sensor hub
• Computer vision projects (with Pi Camera)
• Home automation and smart devices
• Learning edge computing concepts

ML Performance:

• Runs TensorFlow Lite models
• YOLO object detection: ~1-3 FPS (depending on model size)
• MobileNet inference: ~5-10 FPS
• Best for keyword spotting, simple vision tasks

Get Started →

Price: $149-499 | Best For: Computer vision, robotics, edge AI

Powerful edge AI platform with GPU acceleration. Runs full-size deep learning models with real-time performance for vision and robotics applications.

Key Features:

• GPU: 128-core Maxwell GPU (Nano) or up to 1024-core Ampere (Orin)
• 4GB-8GB RAM
• CUDA and cuDNN support
• Multiple camera inputs (CSI)
• GPIO pins for sensors
• Pre-installed ML frameworks

ML Performance:

• YOLOv5/v8 at 20-30 FPS (Nano) or 60+ FPS (Orin)
• Real-time pose estimation
• Multiple object tracking
• Runs full PyTorch and TensorFlow models
• Excellent for autonomous robots and drones

Use Cases:

• Autonomous vehicles and robotics
• Real-time video analytics
• Industrial vision inspection
• Smart retail and surveillance

Learn More →

Price: $59-149 | Best For: Fast ML inference on resource-constrained devices

Google's Edge TPU provides hardware acceleration for TensorFlow Lite models. USB Accelerator adds ML to any device; Dev Board is standalone computer.

Edge TPU Capabilities:

• 4 TOPS (trillion operations/second)
• Optimized for MobileNet, SSD, YOLO
• Runs quantized TensorFlow Lite models
• Ultra-low power consumption (2W)
• MobileNet inference: 100+ FPS

USB Accelerator ($59):

• Plug into Raspberry Pi, PC, or any USB host
• Instant ML acceleration without changing hardware
• Perfect for adding AI to existing edge devices
• Low power, no cooling needed

Dev Board ($149):

• Standalone computer with integrated Edge TPU
• NXP i.MX 8M CPU, 1GB RAM
• WiFi, Bluetooth, USB-C
• 40-pin GPIO header

Explore Coral →

Price: $5-50 | Best For: Ultra-low-power edge, TinyML, sensors

Microcontrollers for battery-powered edge devices. Limited compute but extremely power-efficient for sensor data and simple ML tasks.

Popular Options:

• Arduino Nano 33 BLE Sense ($35): Built-in IMU, mic, temp/humidity, gesture sensor
• ESP32 ($5-15): WiFi/Bluetooth, dual-core, affordable
• Arduino Portenta H7 ($99): Dual-core ARM M7+M4, vision shield available
• Seeed XIAO ESP32S3 Sense ($14): Camera, mic, TinyML capable

TinyML Capabilities:

• Keyword spotting ("Hey device")
• Gesture recognition from IMU
• Vibration anomaly detection
• Simple image classification (low-res)
• Power consumption: milliwatts (battery lasts months)

Perfect For:

• Always-on sensor nodes
• Wearable devices
• Industrial monitoring (vibration, sound)
• Smart home sensors

Browse Arduino →

Type: Full ML Development Platform | Price: Free for developers, paid for commercial

Most accessible end-to-end platform for edge ML. Collect data, train models, deploy to any edge device—all through web interface. No PhD required.

What Makes It Special:

• Complete Workflow: Data collection → Training → Deployment in one platform
• No-Code Interface: Build ML models without writing code
• Device Support: Arduino, Raspberry Pi, ESP32, mobile, more
• Pre-Built Blocks: Audio (keyword spotting), vision (object detection), sensor data
• AutoML: Automatically finds best model architecture
• EON Tuner: Optimizes models for your specific hardware

Supported Use Cases:

• Vision: Object detection (FOMO, YOLOv5), image classification
• Audio: Keyword spotting, audio classification, anomaly detection
• Motion: Gesture recognition, activity classification from IMU
• Sensor Fusion: Combine multiple sensor types
• Anomaly Detection: Industrial predictive maintenance

Learning Path:

• Start with pre-built public projects
• Follow step-by-step tutorials for your hardware
• Collect your own data using phone or device
• Train model with one click
• Deploy to device as Arduino library or binary

Start Building →

Type: ML Framework | Price: Free & Open Source

Google's framework for running TensorFlow models on mobile and edge devices. Industry standard with extensive device support.

Key Features:

• Convert TensorFlow models to optimized .tflite format
• Quantization for smaller models (INT8, float16)
• Hardware acceleration (GPU, Edge TPU, NNAPI)
• Support for Raspberry Pi, Android, iOS, microcontrollers
• Pre-trained models available (image classification, object detection)

TensorFlow Lite Micro:

• Runs on microcontrollers (Arduino, ESP32)
• Footprint as small as 16KB
• No operating system required
• Perfect for TinyML applications

Getting Started:

• Use pre-trained models from TensorFlow Hub
• Convert your own TensorFlow models
• Optimize with post-training quantization
• Deploy with Python or C++ APIs

Get Started →

Type: Computer Vision Platform | Price: Free tier, paid plans

End-to-end platform for computer vision. Upload images, annotate, train models, deploy to edge. Specializes in making YOLO and other vision models accessible.

What It Does:

• Dataset Management: Upload, organize, version control images
• Annotation: Web-based labeling tools for object detection, segmentation
• Augmentation: Automatically generate training variations
• Training: One-click training of YOLOv5, YOLOv8, and other models
• Deployment: Export for edge devices or use hosted API

Supported Models:

• YOLOv5, YOLOv8, YOLOv9 (object detection)
• Faster R-CNN, RetinaNet
• Semantic segmentation models
• Instance segmentation
• Export to TensorFlow Lite, ONNX, CoreML

Perfect For:

• Custom object detection for edge devices
• Quality inspection systems
• Retail analytics
• Security and surveillance

Try Roboflow →

Type: Transfer Learning Platform | Price: Free

NVIDIA's toolkit for training AI models without large datasets or ML expertise. Use transfer learning to adapt pre-trained models to your use case.

Key Capabilities:

• Pre-trained models for common vision tasks
• Transfer learning with small datasets (100s of images)
• AutoML for model architecture search
• Pruning and quantization for edge deployment
• Optimized for Jetson deployment

Supported Applications:

• Object detection (DetectNet, YOLO, Faster R-CNN)
• Image classification
• Semantic segmentation
• Action recognition
• Re-identification

Get TAO →

Latest Version: YOLOv9 / YOLOv10 (2024) | Type: Real-time Object Detection

The gold standard for real-time object detection. Fast enough for edge devices while maintaining high accuracy. Multiple versions optimized for different hardware.

YOLO Evolution:

• YOLOv5: Most popular, excellent balance of speed/accuracy
• YOLOv7: Improved accuracy, still fast
• YOLOv8 (Ultralytics): Easiest to use, great documentation
• YOLOv9: Cutting-edge architecture improvements
• YOLO-NAS: Neural Architecture Search optimized

Edge-Optimized Versions:

• YOLOv5n (nano): Smallest, fastest (1.9MB model)
• YOLOv5s (small): Good balance for Raspberry Pi
• YOLOv8n: Even more efficient nano model
• All can be quantized to INT8 for 4x speedup

Performance on Edge Hardware:

• Raspberry Pi 4: YOLOv5n at 5-10 FPS
• Jetson Nano: YOLOv5s at 20-30 FPS
• Jetson Orin: YOLOv8 at 60+ FPS
• Coral TPU: YOLOv5 TFLite at 15-25 FPS

Easy Deployment:

• Ultralytics library: pip install ultralytics
• Export to TensorFlow Lite, ONNX, CoreML
• Pre-trained on COCO dataset (80 common objects)
• Easy to fine-tune on custom datasets

Get YOLO →

Type: Lightweight Image Classification | Optimized For: Mobile & Edge

Google's efficient neural networks designed specifically for resource-constrained devices. Excellent accuracy with minimal compute.

MobileNet Family:

• MobileNetV1: Depthwise separable convolutions
• MobileNetV2: Inverted residuals, linear bottlenecks
• MobileNetV3: Neural architecture search optimized
• Models from 1MB to 16MB
• Runs smoothly on any edge hardware

EfficientNet:

• Compound scaling of depth, width, resolution
• Better accuracy than MobileNet at similar size
• EfficientNet-Lite optimized for edge
• Used in many mobile apps

Use Cases:

• Image classification (1000 ImageNet classes)
• Transfer learning base for custom tasks
• Feature extraction for other models
• Real-time classification on video streams

Explore Models →

Type: Pre-Built ML Pipelines | Price: Free & Open Source

Google's collection of ready-to-use ML solutions for common tasks. Optimized pipelines that work out-of-the-box on edge devices.

Available Solutions:

• Pose Detection: Track 33 body landmarks in real-time
• Hand Tracking: 21 hand landmarks for gesture recognition
• Face Detection/Mesh: 468 facial landmarks
• Object Detection: Optimized for mobile
• Selfie Segmentation: Real-time background removal
• Holistic: Face, pose, and hands together

Why MediaPipe:

• No training needed—works immediately
• Runs on CPU (no GPU required for many tasks)
• Cross-platform (Python, C++, Android, iOS, Web)
• Real-time performance on edge devices
• Great for prototyping interactive systems

Example Use Cases:

• Fitness tracking and form correction
• Sign language recognition
• AR filters and effects
• Gesture-controlled interfaces

Try MediaPipe →

Type: Cutting-Edge Architectures | Status: Research → Production

Latest models pushing the boundaries of edge AI performance with transformer-based architectures and novel optimizations.

Notable Recent Models:

• FastViT: Vision Transformer optimized for edge (2023)
• EfficientViT: Memory-efficient vision transformers
• MobileViT: Combines convolutions with transformers
• PP-YOLOE: Paddle Paddle's anchor-free YOLO variant
• YOLO-NAS: Neural Architecture Search for optimal structure
• RT-DETR: Real-time transformer-based detection

Trends in Edge Vision:

• Transformer Efficiency: Making attention mechanisms work on edge
• Anchor-Free Detection: Simpler, faster object detection
• Neural Architecture Search: Auto-designed optimal models
• Quantization-Aware Training: Models designed for INT8 from start
• Dynamic Networks: Adjust compute based on input complexity

Where to Find:

• Papers with Code - Latest research
• GitHub Edge AI - Open implementations
• Hugging Face Model Hub - Pre-trained models

Explore Latest →

Type: Edge Runtime & Cloud Integration | Price: Free tier, usage-based

Amazon's edge computing platform. Run Lambda functions, Docker containers, and ML models locally on edge devices while syncing with AWS cloud.

Key Capabilities:

• Deploy AWS Lambda functions to edge devices
• Run ML models (SageMaker Neo optimized)
• Local message processing with MQTT
• Device management and OTA updates
• Works offline, syncs when connected

Learn More →

Type: Edge Computing Platform | Price: Free tier, usage-based

Microsoft's edge platform. Deploy Azure services and custom code to edge devices. Strong support for industrial IoT.

Features:

• Deploy containerized workloads to edge
• Azure Cognitive Services at edge
• Stream Analytics for real-time processing
• Integration with Azure ML
• Enterprise security and compliance

Get Started →

Type: Fleet Management Platform | Price: Free for up to 10 devices

Developer-friendly platform for building, deploying, and managing IoT fleets. Excellent for Raspberry Pi and edge device management.

What It Does:

• Deploy apps to thousands of devices
• OTA updates with rollback
• Remote SSH access to devices
• Docker-based applications
• Environment variables and secrets management

Try Balena →

Type: Visual IoT Programming | Price: Free & Open Source

Flow-based programming tool for wiring together IoT devices, APIs, and services. Perfect for rapid prototyping without coding.

Features:

• Drag-and-drop visual programming
• Huge library of pre-built nodes
• MQTT, HTTP, WebSocket support
• Dashboard creation for IoT visualization
• Runs on Raspberry Pi, servers, cloud

Install Node-RED →