AI Compute & NPU

A dedicated TPU/NPU (Edge TPU, Ethos-U55) runs models far faster than the CPU — but pure-MCU boards are fine for small models.

RAM & PSRAM

Memory, not clock speed, usually limits what model fits. Vision needs PSRAM for the frame buffer — 8 MB is the sweet spot.

Camera & Mic

Vision and audio are the big edge-AI use cases. A built-in camera and microphone save a lot of wiring and debugging.

Software Support

Edge Impulse, TensorFlow Lite Micro, PyTorch or a vendor SDK — good tooling is what turns hardware into a working model.

Power & Form Factor

Battery projects need a low-power MCU board; a Jetson needs a wall supply. Size matters for wearables and sensors.

The Coral USB Accelerator isn’t a board — it’s a USB stick that bolts Google’s Edge TPU (4 TOPS) onto any host, most famously a Raspberry Pi. Plug it into USB 3.0 and a Pi that struggled with one frame per second can suddenly run real-time object detection. It uses TensorFlow Lite models compiled for the Edge TPU and draws very little power. It’s the easiest way to add serious vision throughput to an existing SBC build.

✅ Pros: 4 TOPS in a USB stick; turns a Pi into a vision machine; very low power; mature TF Lite tooling.

❌ Cons: not standalone (needs a host); models must be Edge-TPU-compiled (int8); the software stack has aged; ~$60–80.

The Arduino Nicla Vision packs a standalone smart camera into a 22.86 × 22.86 mm square. It pairs a powerful STM32H747 dual-core (Cortex-M7 at 480 MHz + M4 at 240 MHz) with a 2 MP color camera, a 6-axis IMU, a time-of-flight distance sensor, a microphone and Wi-Fi/BLE. It runs OpenMV and MicroPython and integrates cleanly with Edge Impulse, making it a favorite for asset tracking, object recognition and predictive-maintenance prototypes that need to ship.

✅ Pros: standalone smart camera; OpenMV + Edge Impulse; ToF + IMU + mic onboard; industrial-grade build; tiny.

❌ Cons: ~$80; only 2 MP and 1 MB RAM limits big models; pricier than a XIAO.

The Arduino Tiny Machine Learning Kit bundles the sensor-packed Nano 33 BLE Sense (nRF52840 Cortex-M4F), an OV7675 camera, a custom ML shield and a cable — and pairs with the well-known Harvard/Google TinyML courses on edX. The board senses motion, sound, gestures, color, light and pressure, so you can work through real TensorFlow Lite Micro examples with hardware that matches the curriculum. It’s less about raw power and more about a structured, guided way to actually learn TinyML.

✅ Pros: complete learning bundle; matches the edX TinyML courses; tons of onboard sensors; excellent documentation.

❌ Cons: older 64 MHz M4F is slow; low-res OV7675; costs more than buying parts separately; stock can be spotty.

The Grove Vision AI Module V2 is remarkable for its price: it carries a Himax WiseEye2 with an Arm Cortex-M55 plus an Ethos-U55 micro-NPU — a genuine neural accelerator for around $20. It supports TensorFlow and PyTorch models and Seeed’s no-code SenseCraft AI deployment, so you can flash a person-detection or object-counting model and visualize results without writing inference code. It connects to a XIAO, Arduino, Raspberry Pi or ESP board and a small camera.

✅ Pros: real Ethos-U55 NPU for ~$20; no-code SenseCraft deployment; ultra-low power; pairs perfectly with a XIAO.

❌ Cons: needs a camera + host board; smaller community than Arduino/Pi; docs can be sparse in places.

Espressif’s official ESP32-S3-EYE is a compact AI camera board built around the ESP32-S3 with a 2 MP camera, a small LCD, a microphone and 8 MB PSRAM / 8 MB flash. It ships with ESP-WHO and ESP-DL, so out of the box it does on-device face detection/recognition and voice wake-word commands — all processed on-chip, no cloud. The LCD makes it a satisfying demo board, and it’s a natural next step if you already build with ESP32.

✅ Pros: out-of-box face + voice recognition; LCD for live preview; official Espressif support; fully on-device.

❌ Cons: ESP-WHO/ESP-DL has a learning curve; ~$48; tied to Espressif’s toolchain.

Sony’s Spresense is the oddball that shines for audio. Its CXD5602 chip runs six Arm Cortex-M4F cores at 156 MHz on an ultra-efficient FD-SOI process, with integrated GPS, a 192 kHz/24-bit hi-res audio codec and support for up to 8 microphone inputs. That combination makes it superb for acoustic event detection, sound classification and low-power, GPS-aware sensor nodes. It programs from the Arduino IDE or the NuttX SDK.

✅ Pros: 6 cores + GPS + hi-res multi-mic audio; extremely power-efficient; excellent for audio ML.

❌ Cons: niche ecosystem and smaller community; no onboard Wi-Fi; listed as end-of-life at several retailers — confirm stock before buying.

Just Starting / Best All-Rounder?

Get the Seeed XIAO ESP32-S3 Sense (~$15). Camera, mic, 8 MB PSRAM and Edge Impulse support — the complete TinyML starter.

Need Real Horsepower?

The NVIDIA Jetson Orin Nano Super (~$249) runs LLMs, VLMs and vision transformers with CUDA. The only true generative-AI pick here.

Already Have a Raspberry Pi?

Add the Coral USB Accelerator (~$60–80) for 4 TOPS of real-time vision, or the Grove Vision AI V2 (~$20) for a tiny NPU.

Building a Smart Camera?

The Arduino Nicla Vision (~$80) ships ready to deploy; the ESP32-S3-EYE (~$48) is the best plug-and-play demo.

Want to Learn the Theory?

The Arduino Tiny ML Kit (~$80–100) follows the Harvard/Google edX TinyML courses step by step.

Audio or Battery Sensor Node?

The Sony Spresense (~$65) is unbeatable for low-power, multi-mic audio ML with onboard GPS.

What’s the difference between TinyML and edge AI?

“Edge AI” is the broad idea of running models on the device instead of in the cloud — that includes everything from a Jetson running an LLM to a sensor classifying vibration. “TinyML” is the subset that runs on microcontrollers with kilobytes of RAM and milliwatts of power, like keyword spotting on the XIAO or Nano 33 BLE Sense. Every TinyML project is edge AI, but not every edge-AI board is “tiny.”

Do I need a board with a dedicated NPU or TPU?

Not for most starter projects. Keyword spotting, gesture and simple sensor classification run fine on a plain microcontroller like the XIAO ESP32-S3. You want a dedicated accelerator (Coral’s Edge TPU, the Grove module’s Ethos-U55, or the Jetson’s GPU) once you need real-time camera inference at higher frame rates or larger models — that’s where the CPU becomes the bottleneck.

Can these kits run real LLMs or generative AI?

Only the Jetson Orin Nano Super realistically runs small language models, vision-language models and vision transformers — that’s exactly what its 67 TOPS and 8 GB of LPDDR5 are for. The microcontroller boards (XIAO, Nicla, ESP32-S3-EYE, Spresense) are for classic TinyML: classification, detection and keyword tasks, not text generation. Don’t expect a chatbot from a $15 board.

What software do I use to train and deploy models?

Edge Impulse is the most beginner-friendly path and supports the XIAO, Nicla Vision and Nano 33 BLE Sense end to end. TensorFlow Lite Micro is the underlying runtime for most microcontroller boards. Espressif boards use ESP-DL/ESP-WHO, the Grove module uses Seeed’s no-code SenseCraft AI, and the Jetson runs the full JetPack/CUDA stack with PyTorch and TensorFlow.

Which kit is best for a battery-powered sensor node?

For low-power always-on sensing, a microcontroller board wins — the XIAO ESP32-S3 for general use, the Sony Spresense for audio with GPS, or a XIAO paired with the Grove Vision AI V2 when you need a vision NPU without a power-hungry host. Avoid the Jetson and Coral for battery projects; they’re designed for mains or a sizeable power bank.